In 2023, we ran a small pilot with a primary care practice in Houston. The job sounded simple: help the practice complete Optum's patient assessment forms faster.

We had a lot of questions. What is this form? Why does it exist? Why does nobody want to do it? And the one that surprised us most: why do so many of them get rejected?

The last question is where we first learned about MEAT. Once we understood it, the rest made sense.

1. What the form is

A patient assessment form is a payer's checklist for one patient. The health plan lists the conditions it believes the patient has, based on past claims, pharmacy data, and labs. Some forms also list open quality gaps. The doctor reviews each condition during a visit, usually the annual wellness visit, confirms or rules it out, and sends the form back with the supporting notes.

2. Why it exists

Under risk adjustment, a patient's chronic conditions must be documented in a face-to-face visit every calendar year. If they are not, they do not count toward the patient's risk score, and the plan is paid as if the patient were healthier than they are. The timing also looks back: conditions documented this year set next year's payment.

The form is the plan's way of making sure those conditions get looked at, confirmed, and documented every year. Plans often pay practices a fee for each completed form, because the information is worth a lot to them.

3. Why nobody wants to do it

Most doctors we met saw the form as one more piece of paperwork, and it is easy to see why.

• It repeats work they already did. The diagnosis is already in the note. The form asks them to find it again and copy it somewhere else.
• It takes real time. In the practices we worked with, it took anywhere from 15 to 45 minutes per patient to search the chart, find the evidence, and fill in the form.
• It comes on top of a full day. The forms pile up, so the work spills into evenings and weekends.
• It does not feel like medicine. Nothing on the form changes what the doctor does for the patient that day.

Some doctors we met paid their nurses $50 a form to fill them in over the weekend. That tells you how much the work was costing them, and how little anyone wanted to do it.

4. Why so many got rejected

Here is the part nobody explained up front. Each condition on the form is accepted only if the patient's note supports it. Checking a box is not enough. The plan's reviewers read the note and ask: is there evidence, in this visit, that the doctor actually dealt with this condition?

When the answer was no, the condition was rejected. And the standard the reviewers used had a name: MEAT.

Acceptance by the plan is not the end of it, either. Years later, CMS can pull the same chart in a RADV audit and ask the same question. The note has to hold up both times.

Ruling a condition out is also a valid answer. If the patient no longer has it, saying so on the form, and in the note, is the right call. The form is not asking for more diagnoses. It is asking for accurate ones.

5. What MEAT means

MEAT is a simple test for whether a diagnosis is supported in a note. For each condition, the note should show that the doctor did at least one of these in the visit:

• Monitor: signs, symptoms, disease progression or regression, test results. For example, "eGFR 38, down from 44 in March."
• Evaluate: exam findings, test results, response to treatment. For example, "No edema today, lungs clear."
• Assess: thinking about the condition, ordering tests, discussing it, reviewing records. For example, "CKD stage 3b, likely diabetic, progressing."
• Treat: medications, therapies, referrals, other plans of care. For example, "Continue lisinopril, refer to nephrology."

The letters overlap, and you do not need to label them. A reviewer is not checking whether you wrote "Monitor." They are checking whether the note shows that you dealt with the condition.

Most coders look for at least one MEAT element for each condition. More is safer, especially for conditions that are often questioned in audits.

You may also hear TAMPER (Treatment, Assessment, Monitor or Medicate, Plan, Evaluate, Referral). It is the same idea with different letters.

6. Where MEAT comes from

MEAT is not written into CMS rules. It is the industry's way of applying the official ICD-10-CM coding guidelines for outpatient visits. Those guidelines say to code all documented conditions that coexist at the time of the visit and require or affect patient care, treatment, or management. They also say not to code conditions that were treated in the past and no longer exist.

In other words, a condition counts when it mattered in the visit. A diagnosis that sits on the problem list but never comes up in the note did not affect care that day, so it does not count.

7. Weak vs. strong documentation

The strong versions are not longer for the sake of it. Each one shows, in a sentence or two, what the doctor saw and what they did about it.

8. The most common reasons a condition fails

Most rejections we saw came from a short list of problems. CMS's own RADV guidance flags several of the same ones.

• Problem list only. The condition appears in the problem list or past medical history, but the visit note never addresses it.
• "History of" coded as active. A cancer treated years ago, with no current treatment, is history, not an active diagnosis.
• Complications not linked. "Diabetes" and "CKD" listed separately is weaker than "CKD due to diabetes." The link has to be in the note.
• No plan. The condition is named, but there is no sign of monitoring, evaluation, or treatment.
• Missing signature or credentials. An unsigned note, or one without the provider's credentials, may not be accepted at all.
• Wrong date or wrong year. The visit date is missing, or the visit happened in a different year than the one being reviewed.
• The form and the note disagree. The form says one thing, and the note says something else.
• Copy-forward with nothing new. The same assessment pasted visit after visit, word for word, can make a reviewer doubt the condition was actually addressed that day.

9. What helps

• Address each chronic condition in the assessment and plan, not just the problem list. One line per condition is enough if it shows MEAT.
• Write the numbers and the plan: the lab value, the exam finding, the medication change, the referral.
• Link complications in words: "CKD due to type 2 diabetes," "neuropathy due to diabetes."
• Be honest about history. If a condition is resolved or treated in the past, document it as history. That protects you in an audit.
• Sign every note, with your credentials.
• Fill the form from the note, not the other way around. The note is what gets audited.

10. The takeaway

MEAT is not a coding trick. It is a way of showing that a condition mattered in the visit. When the note shows that you monitored, evaluated, assessed, or treated a condition, the diagnosis holds up. When it does not, the form gets rejected, the payment may be taken back, and the patient looks healthier on paper than they really are.

The forms were never the real problem. The real problem was that the evidence was already in the doctors' notes, and nobody had a fast way to find it and show it.

How we think about it at ParaDocs

That pilot is why ParaDocs exists. We read the note, find the MEAT evidence for each condition, and show it next to the code, so doctors and coders can see right away which diagnoses are supported, which need better documentation, and which should come off.

TL;DR

• What it is: LEAD (Long-term Enhanced ACO Design) is CMS's new 10-year ACO model for Traditional Medicare. It runs from January 1, 2027, to December 31, 2036, and replaces ACO REACH.
• Who it is for: Existing ACOs, plus practices that have never joined one, including small, rural, and independent practices and groups caring for high-needs patients.
• The big change: Risk adjustment. Scores lean more on age and sex, growth is capped, and CMS plans to phase in AI-inferred risk scores starting in 2029.
• Why it matters: Coding more will pay less. Clear documentation and real care will matter more.

For the past few years, ACO REACH has been Medicare's main test of what happens when provider groups take on real financial risk for Traditional Medicare patients. REACH ends on December 31, 2026. On January 1, 2027, CMS starts its replacement: the LEAD model.

LEAD is the longest model the CMS Innovation Center has ever tested. Ten years is long enough for a practice to build a business around it, which is exactly the point. Here is what doctors and practice leaders should know.

1. What LEAD is

LEAD is a voluntary ACO model. A group of doctors, practices, or health systems agrees to be accountable for the total cost and quality of care for a group of Traditional Medicare patients. If they keep spending below a benchmark while meeting quality goals, they share the savings. If spending runs over, they share the losses.

Total cost means everything Medicare Part A and Part B pay for those patients:

• Part A: hospital stays, skilled nursing facility stays, hospice, and some home health.
• Part B: doctor visits, outpatient and ER care, labs, imaging, preventive services, and medical equipment.

Prescription drugs (Part D) are not included.

CMS wants a wider range of practices in the room this time:

• Current ACO REACH and Medicare Shared Savings Program ACOs.
• Practices and providers that have never been part of an ACO.
• Small, independent, and rural practices, community health centers, and rural health clinics.
• Groups that care for high-needs patients, including people who are dually eligible, homebound, or frail.

CMS opened applications in March 2026 for the first cohort, which starts on January 1, 2027.

2. Two ways to take risk

• Professional Risk: The ACO shares 50% of savings and losses. It must take monthly primary care capitation, and it stays in this option for at least four years before it can move up.
• Global Risk: The ACO keeps 100% of savings and owes 100% of losses. "Global" here means full risk for the total cost of care of the ACO's own patients. It does not mean nationwide. In return for the bigger upside, CMS takes a discount of 1.75% to 3% off the benchmark up front. Global ACOs can also choose total care capitation, a monthly payment that covers the Part A and B services their own providers and partner providers deliver.

Very large savings or losses are shared at lower rates, a safety limit called risk corridors, so no group carries unlimited exposure.

Here is a simplified example. The real formulas have more steps, but the logic is the same. Say an ACO has 10,000 patients and a benchmark of $12,000 per patient per year. That is a $120 million benchmark.

Illustration only. Leaves out the quality withhold, risk corridors, and other CMS adjustments.

For scale, the average Shared Savings Program ACO had about 23,500 patients in 2025. The $12,000 per patient is a round number for illustration. Real benchmarks vary by region and by how sick the patients are.

In plain terms, Professional is the safer entry point: half the upside, half the downside. Global is for groups ready to manage the full cost of care. The discount means a Global ACO has to beat the benchmark by more before it saves anything. In this example, Global only pays more than Professional once spending falls below about $112.8 million, roughly 6% under the benchmark.

3. How the money works

Today, Medicare pays most doctors through the fee schedule: a set price for each billed service. A mid-level office visit (99214) pays one rate. A 20-minute chronic care management call (99490) pays another. The more services billed, the more the practice is paid.

LEAD replaces that, for primary care, with capitation: a set monthly payment per patient. For example, an ACO paid $40 per patient per month for 10,000 patients receives $400,000 every month, whether a patient comes in twice or ten times. Claims are still sent, so CMS can see the care, but the payment for those services comes through the capitation instead. That gives practices predictable cash flow and the freedom to spend time where patients need it, such as care coordination or home visits, instead of where the fee schedule pays.

CMS sets the base capitation from the ACO's own past primary care spending. ACOs can also take an enhanced amount up front to invest in staff and workflows, but that extra has to be paid back at the end of the year.

Where the benchmark comes from. The benchmark is built mostly from the ACO's own patients' past spending, not a national average. For the first year, CMS looks at what Medicare spent on those patients in 2024, 2025, and 2026. Older years are adjusted to today's prices, and for new ACOs the most recent year counts most: 10% for 2024, 30% for 2025, and 60% for 2026.

For example, if those adjusted costs were $11,600, $11,900, and $12,000 per patient, the baseline is 0.1 × $11,600 + 0.3 × $11,900 + 0.6 × $12,000 = $11,930. CMS then trends that number forward to 2027, using a blend of national and regional spending growth, and adjusts it for how sick the patients are. The baseline stays fixed for years instead of resetting, which fixes a long-standing complaint from ACOs that got penalized for their own past success.

Region still matters in two ways. ACOs that already spend less than their region can get a boost toward the regional level, and CMS uses regional spending to decide who counts as a higher-spending ACO. CMS plans to move toward regional rates, but no sooner than 2031.

The add-on payment. Higher-spending ACOs get an extra 1.5% of their benchmark, paid monthly, to help build the staff and systems an ACO needs. On a $120 million benchmark, that is $1.8 million a year, about $150,000 a month. It is never reconciled, so it does not have to be paid back, even if the ACO loses money.

4. The big change: risk adjustment

This is the part I would read twice.

Risk adjustment decides how much money follows each patient, based on how sick they are. If you are new to this, our article on risk vs. quality explains the basics. LEAD changes it in four ways.

• More weight on age and sex. For most patients, LEAD uses a modified version of the CMS-HCC V28 model that gives more weight to demographics and less to diagnosis codes. CMS says this should reduce the payoff from aggressive coding.
• Capped growth. A group's average risk score can only grow so much against a fixed reference year: 3% for most patients, and 4% for high-needs patients. If a group's coding pushes scores higher than that, the extra does not count.
• Concurrent scores for high-needs patients. For the sickest patients, the risk score uses diagnoses from the same year as the care, not the year before.
• AI-inferred risk scores. CMS plans to test an AI-based risk score in 2028, then blend it in: one-third of the score in 2029, two-thirds in 2030, and fully AI-inferred from 2031. CMS says the timeline depends on testing and could change.

Here is what concurrent scoring means in practice. Say an 84-year-old patient with heart failure is diagnosed with metastatic lung cancer in May 2027 and starts treatment. Under the usual model, her 2027 risk score is based on her 2026 diagnoses, so it reflects heart failure only. The money that follows her in 2027 is far less than what her care will actually cost. Under the concurrent model, the cancer documented in 2027 counts toward 2027. Her score rises in the same year her costs do. That fits high-needs patients, whose health can change quickly.

What AI-inferred means. Today, a risk score is mostly addition. Each coded condition has a weight, and the weights are added to an age and sex factor. Two patients with the same codes get the same score.

An AI-inferred score is a prediction instead. A model trained on large amounts of Medicare data looks at the patient's whole profile and estimates what their care is likely to cost. Take two patients, both coded with diabetes with complications. One has rising A1c values, new insulin fills, and two ER visits this year. The other has the code and little else. Today they score the same. An AI model may not score them the same.

CMS has not published how the model works or exactly which data it will use. It has said diagnosis codes will be one input alongside other data sources. In 2028, CMS will send ACOs an AI-inferred score for each patient, next to the usual score, before it counts for payment. That year is the chance to compare the two and see where they disagree.

What this means for doctors: the era of finding more codes to lift a risk score is ending. Diagnoses still matter, but mostly when they are accurate, supported, and tied to real care. A clear note that shows MEAT for each condition is still the foundation. What changes is that inflating codes buys much less.

5. Quality: what LEAD measures

LEAD holds back 3% of the benchmark as a quality withhold. On a $120 million benchmark, that is $3.6 million. It is not taken out of the monthly payments. Instead, CMS sets it aside when it calculates savings or losses at the end of the year, then gives back some or all of it based on the ACO's quality scores. An ACO with strong quality earns back the full $3.6 million. An ACO that earns back only half leaves $1.8 million on the table.

The measures include:

• Hospital readmissions within 30 days.
• Unplanned admissions for patients with multiple chronic conditions.
• Days at home for patients with complex chronic conditions.
• Timely follow-up after acute exacerbations of chronic conditions.
• Patient experience surveys (CAHPS).

Most of these come from Medicare claims, so CMS calculates them without the ACO sending anything extra.

LEAD also adds two electronic clinical quality measures from the Shared Savings Program: controlling high blood pressure, and diabetes with an A1c above 9% (poor control). They are optional in the first two years, pay-for-reporting in years three and four, and pay-for-performance from year five.

These two are different. Medicare claims usually do not carry the blood pressure reading or the A1c number, so the ACO has to pull those values from its EHR and report them electronically to CMS. There is no health plan in between. LEAD is Traditional Medicare, so CMS is the payer. If those numbers are typed only into a note or sit in an outside lab report, the EHR report may miss them. Each ACO also has to write a Prevention and Quality Plan.

6. High-needs patients and post-acute care

LEAD has a special path for groups that care for the sickest patients. The criteria below apply to patients. An ACO qualifies as a High Needs ACO if at least 40% of its aligned patients each meet at least one of them:

• Mobility impairment or frailty.
• Serious chronic illness.
• Two or more unplanned hospital admissions in the past year.
• 45 or more days in a skilled nursing facility in the past year.

CMS decides which patients qualify, using Medicare claims.

The ACO, in turn, has to meet its own requirements. High Needs ACOs must offer 24/7 access to a provider, train staff in advance care planning, and be able to deliver care at home. In return, they need far fewer aligned patients to start: 800 in the first year, compared with 5,000 for a standard ACO.

Several flexibilities carry over from REACH and matter for SNF and long-term care. These are rule waivers the ACO's providers can use for their aligned patients:

• Waiver of the 3-day hospital stay rule before a skilled nursing stay.
• Home visits after discharge from a hospital, skilled nursing, or long-term care stay.
• Concurrent hospice care, and nurse practitioners and physician assistants allowed to certify hospice.
• Home health without the homebound requirement.

For groups that already care for nursing home residents, this is the most important part of LEAD. The patients they already serve are exactly the ones the model is built around.

7. CARA: working with specialists

LEAD also introduces CARA (CMS-Administered Risk Arrangements). It lets Global Risk ACOs set up episode-based payment arrangements with specialists and post-acute providers, which CMS calls Preferred Providers.

An episode is a defined stretch of care for one condition, with a target price. For example, an ACO and a cardiology group agree on a heart failure episode for the ACO's patients, with a target of $18,000 per patient for the year. If the cardiology group's patients average $16,000, the $2,000 saved per patient is split between the ACO and the cardiologists, as they agreed. If they average $20,000, the cardiologists share the loss.

CMS does the heavy lifting. It builds the episodes from Medicare claims, sends both sides data and target price reports, tracks actual spending, and reconciles the result at the end of the year. The ACO and the specialist do not have to build their own claims analytics or chase each other for payment. CMS says target price reports come in the second quarter of 2027, the first episodes count in 2028, and the first reconciliation follows in mid-2029. One of the first is a falls prevention episode, with chronic conditions such as diabetes, heart failure, CKD, COPD, and depression phased in later.

8. What to do now

Whether or not your group joins LEAD, the direction is clear. A few things are worth doing now:

• Document once, clearly. Address each chronic condition in the assessment and plan with MEAT, and stop relying on problem lists.
• Make sure your numbers reach the right fields. Blood pressure and A1c values need to land where quality reports read them: the vitals section and lab results in your EHR, and CPT II codes on claims for health plan measures. A value typed only into the note, a home reading, or an outside lab report may never be counted.
• Know your high-needs patients. Frailty, SNF days, and repeat admissions decide which track fits, and how your patients are risk adjusted. Use the claims data CMS sends ACOs, hospital admission and discharge alerts, SNF records, and frailty or function assessments in your EHR to see them before CMS does.
• Do not build a strategy on coding intensity. Capped scores and AI-inferred risk will reward accuracy and real care, not volume.
• Watch the AI risk adjustment closely. It is the biggest change to risk adjustment in years. Use the 2028 shadow scores to see where your patients land.

9. The takeaway

LEAD is CMS's clearest signal yet about where value-based care is heading: longer commitments, more predictable payments, more room for small practices and high-needs care, and much less reward for coding games. For doctors, the job gets simpler in one way. Take good care of patients, and write it down clearly. The model is being built to reward exactly that.

How we think about it at ParaDocs

As risk adjustment moves away from code counting, the value shifts to evidence. We read the clinical note and outside records, find the support for each condition and the values quality measures need, and link them to the exact sentence in the chart, so the evidence can reach the places that count.

TL;DR

• The benchmark: It is a scorecard, not a payment. Medicare keeps paying hospitals, specialists, and nursing facilities directly, all year.
• What the ACO gets: A share of the savings at the end of the year, if it spent less than the benchmark and met its quality goals. Newer models also pay some money up front.
• Medicare Advantage is different: There, CMS pays the health plan the full amount every month, and the plan keeps what it does not spend.
• Reaching patients: Mostly, nobody can tell whether the money does. What can be tracked is results: admissions, readmissions, blood pressure, A1c, and patient experience.

My intern asked me this almost every week. So did some of my investors, and doctors who have only ever worked in fee-for-service. The question was always some version of: "If an ACO is responsible for $120 million in patient care, does it get $120 million? And then keep whatever is left?"

It is a fair question. The words make it sound that way. ACOs talk about "taking risk" and "managing a benchmark," and it is easy to picture a pot of money handed over at the start of the year.

That is not how it works. Here is the plain version.

1. The short answer

In Traditional Medicare, the ACO does not receive the benchmark. The benchmark is a target. Medicare keeps paying for care the same way it always has, claim by claim, and at the end of the year it compares what it spent with the target.

If spending came in under the target and quality was good, the ACO gets a share of the difference. If spending came in over, the ACO may owe part of the difference back. The ACO's real money is the difference, not the whole amount.

2. A year in an ACO, step by step

Here is how it works in the Medicare Shared Savings Program, the largest ACO program, with 476 ACOs and about 11 million patients in 2025.

• January to December: Patients see their doctors, go to the hospital, and stay in nursing facilities as usual. Every provider bills Medicare and gets paid by Medicare. That money never passes through the ACO.
• Throughout the year: CMS adds up everything Medicare paid for the ACO's assigned patients, from every provider, inside or outside the ACO.
• After the year ends: CMS compares that total with the ACO's benchmark and checks the ACO's quality scores.
• Settlement: If the ACO saved enough and met its quality standard, CMS sends it a check for its share of the savings. If it is in a two-sided track and spent more than the benchmark, it pays part of the loss back.

A simple example. Say an ACO has 10,000 patients and a benchmark of $12,000 per patient per year, so $120 million. The ACO's patients end up costing Medicare $114 million. That is $6 million in savings. In the Shared Savings Program, ACOs keep a share of the savings, up to 75% depending on their track and quality score. At a 50% share, the ACO gets a check for $3 million. Medicare keeps the other $3 million.

So "responsible for $120 million" really means "graded on $120 million, paid on the $3 million difference."

Real numbers look like this. In 2024, 75% of Shared Savings Program ACOs earned shared savings, and CMS paid them $4.1 billion in total. Sixteen ACOs owed money back, $20 million in total. Spread across all 476 ACOs, that is about $8.6 million per ACO on average, before it is split among the practices inside it.

3. Newer models put some money in earlier

Waiting until the year is over is hard on a practice's cash flow. So newer models, like ACO REACH and its replacement, LEAD, pay some money up front.

• Primary care capitation: Instead of billing each primary care visit, the ACO receives a set monthly amount per patient for primary care. In our example, $40 per patient per month for 10,000 patients is $400,000 a month.
• Total care capitation: ACOs taking full risk can choose a monthly payment that covers the services their own providers and partner providers deliver. The ACO then pays those providers itself.
• Add-on payments: LEAD pays higher-spending ACOs an extra 1.5% of their benchmark, paid monthly, to build the staff and systems an ACO needs.

Even here, the ACO does not get the whole benchmark. A hospital outside the ACO is still paid by Medicare directly, and that cost still counts against the ACO at the end of the year. Our article on the LEAD model walks through this in more detail.

4. Medicare Advantage is the version people picture

The picture of "getting the whole pot and keeping the rest" is much closer to Medicare Advantage.

In Medicare Advantage, CMS pays a private health plan a set amount every month for each member, adjusted by that member's risk score. The plan then pays the hospitals, doctors, and nursing facilities. Whatever it does not spend, it keeps, within limits. Plans must spend at least 85% of their revenue on medical care and quality improvement.

Some plans pass that model down to provider groups. A group may receive a percentage of what the plan gets for each patient, often called a percent-of-premium deal, and keep whatever it does not spend on care. That is where "you get the money up front and keep what is left" is actually true.

It is also why risk adjustment matters so much more in Medicare Advantage. The risk score sets the money that arrives every month. In a Shared Savings Program ACO, the risk score mostly adjusts the target.

Simplified. Details vary by track, model, and contract.

5. So where does the savings check go?

Once the check arrives, the ACO decides how to use it. CMS does not tell it how to split the money. It does require two things:

• A written plan. Every ACO must tell CMS how it will distribute savings among its practices and providers.
• A public report. Every Shared Savings Program ACO must post on its website how much of its savings went to infrastructure, how much went to redesigning care, and how much was paid out to its practices and providers.

How the money is split varies a lot. Some ACOs pay practices based on how much each one contributed to the savings. Others weigh quality, such as annual wellness visits, diabetes control, and blood pressure control, or how engaged a practice was in meetings and data sharing. Many keep part of the money to run the ACO: care managers, analytics, software, and staff.

For a doctor inside an ACO, the result often shows up as a once-a-year bonus check, months after the year ended, based on a formula they may never have seen.

6. Can anyone tell if the money reached patients?

Mostly, no.

• The public reports are percentages in broad categories. "Distributed to ACO participants" can mean physician bonuses. "Infrastructure" can mean software or care managers.
• In Medicare Advantage, payments from a plan to a provider group count as medical spending toward the 85% rule, even if the group keeps part of it as profit.
• Nobody tracks whether a higher risk score for one patient turned into more care for that patient.

There was one place CMS tied savings directly to patients. ACOs with a history of savings could take part of their expected savings early, and at least half of that money had to go to services for patients that Medicare does not normally cover. Few ACOs used it, and CMS has proposed ending the option after 2027.

So CMS has largely stopped trying to follow the money. Instead, it measures what happened to the patients: hospital admissions, readmissions, days at home, blood pressure, A1c, and patient experience surveys. That is why newer models like LEAD and ACCESS lean so heavily on those measures. The dollars are hard to trace. The results are not.

7. What this means for doctors

If you work inside an ACO, or are thinking about joining one:

• Ask for the distribution plan. You are entitled to know how savings are split, and what you are measured on.
• Know what drives savings. Most savings come from avoiding hospital stays, readmissions, and unnecessary post-acute days, not from coding. In 2024, ACOs made up mostly of primary care clinicians saved Medicare $403 per patient, compared with $224 for other ACOs.
• Know which model you are in. In a Shared Savings Program ACO, the risk score adjusts the target. In Medicare Advantage, it sets the payment. The documentation rules are the same, but the stakes are different.
• Look at the public report. Every Shared Savings Program ACO posts one. It will not tell you everything, but it is a start.

8. The takeaway

An ACO is not handed a pot of money. It is handed a target and a scorecard, and paid a share of what it saves. Medicare Advantage is the closer fit to "get the money, keep the rest," and that is exactly why risk scores matter most there.

Whether the money reaches the patient is the harder question, and the honest answer is that nobody can fully trace it. What we can trace is whether patients stayed out of the hospital, got their blood pressure and A1c under control, and felt cared for. That is where the attention is moving.

How we think about it at ParaDocs

We cannot follow the dollars either. What we can do is show the evidence on both sides: that a patient is as sick as their risk score says, and whether they got the care their conditions need. A risk score without the second half is just a number.

TL;DR

• Recapture: A condition that was documented last year and has to be documented again this year. The patient is known to have it.
• Suspect: A condition that has never been documented, or not recently, but other data in the chart suggests the patient may have it. Nobody has confirmed it yet.
• Why it matters: They carry different risks. Recapture fails when the note is thin. Suspects fail when there is no real diagnosis and no real care behind them.
• What we found: In our data, suspect gaps backed by a lab result were approved by clinicians 88.5% of the time. Gaps backed only by a medication were approved 31% of the time.

The first payer forms I worked on had one list of conditions for each patient. It took me a while to see that the list was really two lists mixed together.

Some conditions were already in the patient's history. The doctor had diagnosed them, coded them, and treated them the year before. Others had never been diagnosed at all. They were there because a lab value, a medication, or an old record hinted at them.

The form treated them the same way: a box to check. But they are not the same, and treating them the same is where a lot of risk adjustment goes wrong. The first kind is called recapture. The second is called a suspect gap.

1. Why gaps exist at all

In Medicare Advantage and most risk-based contracts, chronic conditions reset every year. A condition only counts toward a patient's risk score if it is documented in a face-to-face visit during the calendar year. If heart failure is documented in 2025 but not in 2026, it does not count for 2026, even if the patient still has it.

That reset creates two kinds of gaps:

• Recapture gaps: Conditions documented before that have not been documented yet this year.
• Suspect gaps: Conditions that have never been documented, or not for a long time, but that the data suggests may be there.

2. Recapture: known conditions that need to be documented again

A recapture gap starts from something real: a past diagnosis by a clinician. The job is to see the patient, evaluate the condition again, and document it with MEAT, showing it was monitored, evaluated, assessed, or treated in the visit.

Recapture is usually the safer of the two, but it fails in predictable ways:

• The condition was resolved. A cancer treated years ago, a past stroke with no lasting effects, or a depression in remission should not be carried forward as active.
• The note is thin. "CHF, stable" with nothing else may not hold up in an audit.
• Copy-forward. The same assessment pasted visit after visit can make a reviewer doubt the condition was addressed that day.

A good recapture review asks one question: is this still true, and does this year's note show it?

3. Suspect gaps: conditions the data hints at

A suspect gap is a hypothesis, not a diagnosis. It comes from evidence somewhere in the record:

• A lab result. An eGFR below 60 on two tests more than three months apart, with no CKD diagnosis.
• A medication. Insulin, an anticoagulant, or an antipsychotic with no matching diagnosis.
• A diagnosis from somewhere else. A specialist letter, a hospital discharge summary, or an old claim that never made it into the primary care note.
• An exam finding or report. An echo showing a low ejection fraction, or an amputation noted in the exam.

Some of these are strong. Others are weak, and the difference is the whole point of this article.

4. Not all evidence is equal

A lab result is usually tied to how the condition is defined. Chronic kidney disease is defined by kidney function below a threshold for at least three months. Diabetes can be diagnosed with an A1c of 6.5% or higher. When the lab meets the definition, the suspect is close to a diagnosis already. The doctor still has to evaluate it and decide.

A medication alone is weaker, because most drugs have more than one use:

• Furosemide is used for heart failure, but also for edema from other causes and for kidney disease.
• Metformin is used for diabetes, but also for prediabetes and polycystic ovary syndrome.
• An anticoagulant may mean atrial fibrillation, or a past blood clot, or a mechanical heart valve.
• Gabapentin may mean diabetic neuropathy, or shingles pain, or anxiety, or sleep.

A suspect built on a medication alone is a question for the doctor, not an answer.

5. What our data showed

We studied 1,989 suspect gaps ParaDocs surfaced in a multi-state SNF and long-term care population of about 8,000 patients. Every gap was reviewed by a clinician, who approved or rejected it. Overall, 76.3% were approved. But the approval rate depended heavily on the kind of evidence behind the gap.

1,989 suspect gaps, multi-state SNF/LTC population, CMS-HCC V28, 2024 to 2025.

It also varied by condition. Emphysema had the highest approval rate, at 93.1%, usually backed by strong diagnostic evidence. Unstable angina had the lowest, at 11.6%. A past cardiac event or a nitrate prescription is not enough to say a patient has unstable angina today.

The lesson is simple. The number of gaps tells you very little. The evidence behind each gap tells you whether it is likely to be real.

6. Why this matters beyond payment

Suspect gaps have a reputation problem, and it is earned.

The HHS Office of Inspector General has looked at diagnoses that show up only in health risk assessments and chart reviews, with no other record of care for that condition. In 2016, those diagnoses drove an estimated $9.2 billion in Medicare Advantage payments. In 2023, diagnoses reported only on health risk assessments, and the chart reviews linked to them, drove an estimated $7.5 billion. OIG's concern was simple: either the diagnosis is not accurate, or the patient has a serious condition and is not getting care for it. Neither is good.

That is the difference between a suspect gap done well and one done badly. Done badly, a suspect becomes a code with nothing behind it. Done well, a suspect becomes a doctor looking at an eGFR of 38, confirming CKD stage 3b, adjusting medications, and referring to nephrology. The second patient is better off. The first is just a number.

7. What helps

• Separate the lists. Recapture and suspects need different reviews. Do not send them to doctors as one undifferentiated list.
• Rank suspects by evidence. Lab-backed and specialist-backed suspects first. Medication-only suspects last, and framed as questions.
• Show the evidence. A suspect with the lab value, the date, and the source is something a doctor can act on in seconds. A suspect with no evidence is just a request to add a code.
• Let doctors rule things out. Rejecting a suspect is a valid clinical answer, and it protects the practice. A high rejection rate on weak evidence is a sign the process is working, not failing.
• Close the loop with care. A confirmed condition should come with a plan: a medication change, a referral, a follow-up. That is what MEAT looks for, and it is also what the patient needs.

8. The takeaway

Recapture asks: is this known condition still true, and did we document it this year? Suspect gaps ask: is there something here we have not diagnosed yet? Both matter. But they are different questions, with different risks, and they need different evidence.

More gaps is not the goal. Better evidence is. When the evidence is strong, doctors confirm the condition, the code holds up, and the patient gets care for something that was being missed.

How we think about it at ParaDocs

We built ParaDocs to show the evidence behind every gap, not just the gap. Each suspect comes with the lab, note, or report that triggered it, and its date and source, so clinicians can confirm what is real, rule out what is not, and spend their time on the ones most likely to matter.

TL;DR

• The short answer: HEDIS is the biggest quality measure set, but it is mainly for health plans. Doctors, ACOs, nursing facilities, hospitals, and home health agencies each have their own programs.
• The good news: Many programs use the same measures. Blood pressure control, A1c, and cancer screening show up almost everywhere.
• The catch: Each program has its own rules for who counts, what time window counts, and how the data has to be reported.
• What helps: Capture the value once, clearly, in a place every program can find it.

For a long time, I used "quality" and "HEDIS" as if they were the same word. So do most people in value-based care. When a payer sends a quality gap list, it is usually HEDIS. When someone talks about Star Ratings, HEDIS is a big part of it.

But when I started working with skilled nursing and long-term care groups, ACOs, and doctors in fee-for-service practices, the word "quality" kept meaning something different. The nursing facility cared about its Five-Star rating. The ACO talked about eCQMs. The independent practice worried about MIPS. They were all talking about quality, and none of them were talking only about HEDIS.

Here is a plain map of who measures what.

1. HEDIS: quality for health plans

HEDIS (Healthcare Effectiveness Data and Information Set) is run by NCQA. It has more than 90 measures across six areas, such as prevention, chronic disease, behavioral health, and access to care. Most health plans in the US report it, across commercial, Medicaid, Medicare Advantage, and ACA marketplace plans.

HEDIS measures the plan, not the doctor. But plans need the data from doctors, which is why practices get gap lists asking for an A1c, a blood pressure, or proof of a colonoscopy. Our article on risk vs. quality explains how those gap lists work.

2. Programs built on top of HEDIS

Several rating programs use HEDIS as one ingredient and add their own:

• Medicare Advantage Star Ratings: About 40 measures from HEDIS, patient surveys, Part D pharmacy data, and plan operations. Plans rated four stars or higher earn a quality bonus.
• ACA Marketplace Quality Rating System: CMS rates marketplace health plans from one to five stars, using clinical measures and a member survey.
• Medicaid Core Sets: Federal sets of measures for adults and children. States must now report the Child Core Set and the behavioral health measures in the Adult Core Set.

3. Programs for doctors and ACOs

In Traditional Medicare, there is no health plan in the middle, so CMS measures the clinicians directly.

• MIPS (Merit-based Incentive Payment System): Many clinicians who bill Medicare Part B report quality measures each year. Their payments can go up or down by as much as 9% based on their score.
• Shared Savings Program ACOs: Report a set of measures called APP Plus. Meeting the quality standard is required to share in savings.
• LEAD and other CMS models: Use claims-based measures like readmissions and days at home, patient surveys, and electronic measures for blood pressure and A1c.

Many of these are reported as eCQMs, electronic clinical quality measures. An eCQM is a measure written as a precise specification that an EHR can run against its own data. The ACO or practice pulls the values from the EHR and reports them to CMS.

4. Programs for nursing facilities, hospitals, and home health

Facilities have their own programs, and most are built from their own assessment data and claims:

• Nursing facilities: Care Compare's Five-Star rating combines health inspections, staffing, and quality measures. Many of those quality measures come from the MDS, the standard assessment completed for every resident.
• SNF Quality Reporting Program: Skilled nursing facilities that do not report required data lose two percentage points of their annual Medicare payment update.
• SNF Value-Based Purchasing: CMS withholds 2% of a facility's Medicare Part A payments and pays back about 60% of the pool as incentives, based on performance, starting with hospital readmissions.
• Hospitals: Programs like the Hospital Readmissions Reduction Program can cut payments by up to 3%.
• Home health: Agencies are measured through the OASIS assessment and a value-based purchasing program that now runs nationwide.

For groups that care for nursing facility residents, this matters. The facility, the physician group, the ACO, and the health plan may all be measured on the same patient, in different programs, at the same time.

5. The same measure, everywhere

Here is the part that surprises people. Many of these programs measure the same few things.

Measure names and IDs vary by program and year. Shared Savings ACOs report several of these eCQMs too.

Controlling blood pressure is a HEDIS measure, a Star Ratings measure, a Medicaid Core Set measure, a MIPS measure, an eCQM for ACOs, and part of LEAD. The clinical goal is the same: is the patient's blood pressure under control? The difference is in the details.

6. Where the details differ

• Who counts: Each program has its own rules for which patients are included and excluded. A patient in hospice, or with advanced illness and frailty, may be excluded in one program and included in another.
• What time window counts: Most measures use the calendar year. Screenings look back further. A colonoscopy counts for 10 years, for example, but only if the date is recorded.
• Where the data comes from: HEDIS can use claims, CPT II codes, lab feeds, and chart review. eCQMs come from the EHR's structured fields. Nursing facility measures come from the MDS.
• How it is reported: Health plans report HEDIS to NCQA and CMS. Practices and ACOs report to CMS. Facilities submit assessments.

This is why the same blood pressure reading can count in one program and be missed in another. If it was typed only into a note, HEDIS chart review may find it, but an eCQM pulling from the vitals section will not.

7. What helps

• Capture values in structured fields. Blood pressure in the vitals section, A1c as a lab result, screenings with a date. That is where eCQMs and most automated reports look.
• Write the date and the source for outside results. "Colonoscopy 3/2023 at Memorial, normal." A line like that can close a gap across several programs.
• Use CPT II codes when payers ask for them. They carry results like an A1c below 7% straight to the health plan on the claim.
• Know which programs you are in. A physician group serving nursing facility residents in a Medicare Advantage plan and an ACO may be measured in three or four programs at once.
• Do the work once. One clear note and one structured value should serve every program. If your team is entering the same number in three places, something is wrong.

8. The takeaway

HEDIS is the biggest quality system, but it is not the only one. Health plans, clinicians, ACOs, nursing facilities, hospitals, and home health agencies are each measured by their own programs, and they all overlap on the same few things: blood pressure, A1c, screenings, readmissions, and patient experience.

For doctors, the job is the same in every program. Measure it, write it clearly, and put the number where systems can find it. The programs will keep changing. The patient's blood pressure is still the patient's blood pressure.

How we think about it at ParaDocs

We read the note and the outside records to find the values quality programs need, like the blood pressure reading, the A1c, and the screening date, and link each one to its source. That way one piece of evidence can count everywhere it should, instead of being re-entered program by program.

When I first discovered value-based care, I spent time shadowing a few primary care clinics. I wanted to understand how they actually practiced it: what slowed them down, what they vented about, and what was taking the joy out of practicing medicine. I talked with everyone, from the front desk and admin staff to billing, coders, nurses, and the doctors, and watched what each of them did all day.

So many "whys" came out of those visits. Two words kept coming up: risk and quality. People used them in the same sentence, sometimes as if they were the same thing. Payers sent lists for both. Meetings covered both. It felt like a labyrinth, and every piece of the puzzle led back to the same place: the clinical note. Nobody could quite explain why the note mattered so much to both.

In 2023, I even completed the Optum Education Series and got certified, just to understand it better.

If risk and quality still feel confusing to you, you are not alone. Most doctors were never taught this in training. Yet both shape how much a practice gets paid, how its performance is judged, and how much time it spends on paperwork.

Here is the plain-language version I wish someone had given me.

1. The short answer

Risk adjustment asks: how sick are your patients?

Quality asks: did your patients get the right care?

Both are measured from the same place: what happens in the visit and what gets written in the note. But they look for different things, follow different rules, and pay out in different ways.

2. What risk adjustment measures

In Medicare Advantage, ACO REACH, and many other value-based contracts, payment is adjusted for how sick a patient is. CMS does this with the CMS-HCC model, which turns a patient's diagnoses and demographics into a risk score, often called the RAF score.

• A RAF of 1.0 is roughly the average Medicare beneficiary.
• A RAF of 2.0 means the patient is expected to cost about twice as much.
• The sicker the patient, the more money follows them to cover their care.

What counts as evidence for risk adjustment:

• Diagnoses, coded in ICD-10, that map to an HCC.
• Documented in a face-to-face visit with an acceptable provider.
• Documented again every calendar year. Chronic conditions do not carry over. The model also looks back: diagnoses documented in 2026 set the risk score that pays for the patient's care in 2027. If heart failure is not documented in 2026, it does not count toward 2027, even if the patient still has it.
• Supported in the note. Coders and auditors look for MEAT: was the condition Monitored, Evaluated, Assessed, or Treated in that visit? MEAT is not written into CMS rules, but it is the standard most coders and auditors use.

In Medicare Advantage, the audit behind it is RADV, short for Risk Adjustment Data Validation. Here is how the money flows:

• CMS pays the Medicare Advantage plan a monthly amount for each member, adjusted by that member's risk score.
• The plan pays the doctor or provider group. In a risk contract, that payment is often a share of what the plan receives.
• In a RADV audit, CMS reviews a sample of charts to check whether each submitted diagnosis is supported by the medical record.
• If a diagnosis is not supported, CMS takes that money back from the plan.

Whether the plan then takes money back from the doctor depends on the contract. In upside-only arrangements, the plan usually keeps that risk. In full-risk or capitated contracts, where the provider group controls its own coding, the contract may pass part of the recovery down. It is worth knowing what yours says.

CMS also tried to extrapolate audit findings across a plan's whole membership. A federal court vacated that rule in September 2025 and HHS has appealed, so for now recoveries apply only to the charts that were audited. Audits continue either way.

3. What quality measures

Quality is about whether the right care happened, on time, with the right result. The most common set of measures is HEDIS, run by NCQA. Medicare Advantage plans are also rated with Star Ratings, which use HEDIS along with patient surveys and other data.

Examples you will recognize:

• Glycemic status for patients with diabetes: was the most recent A1c of the year below 8.0% (good control), or above 9.0% (poor control)?
• Controlling high blood pressure: was the most recent BP reading below 140/90?
• Kidney health evaluation for patients with diabetes: did they get both an eGFR and a urine albumin-creatinine ratio?
• Eye exams, colorectal cancer screening, and breast cancer screening: were they done, and when?

Who sets these targets? NCQA does. It develops HEDIS measures with expert panels and public comment, based on clinical guidelines and evidence, and updates them every year.

The targets are set for whole populations, so they are often looser than what you would aim for with one patient. Many guidelines suggest an A1c below 7% for most adults, but the HEDIS measure uses 8% and 9%, because the same target applies to everyone, including older or frail patients where tighter control can cause harm. A patient with no A1c on record for the year counts as poor control.

What counts as evidence for quality:

• Services and results, not diagnoses.
• Specific values and dates: the A1c number, the BP reading, the date of the screening.
• Often reported through CPT II codes on the claim, such as 3044F for an A1c below 7.0%, or through clinical data sent to the plan.
• Proof that the service happened inside the measurement window. An order is not enough. The result or the record of the service is what counts.

The measurement window is usually the measurement year, January 1 to December 31. Some screenings look back further, because they are not needed every year. A colonoscopy counts if it was done this year or in the 9 years before. A stool DNA test such as Cologuard counts for this year and the 2 years before. A FIT test only counts in the current year.

Not every measure needs a result. For colorectal cancer screening, the measure only asks whether the screening was done, and when. A line in the history such as "Colonoscopy 6/2021" counts, even without the result. The result still matters clinically, and a history of colorectal cancer takes the patient out of the measure entirely. For A1c and blood pressure, the value is the whole point.

The money works differently too. Medicare Advantage plans rated four stars or higher earn quality bonus payments, and many ACO contracts only share savings if quality targets are met. So quality usually earns a bonus or unlocks one, while risk adjustment sets the base payment.

4. Where they overlap

The same patient, the same visit, and the same note often feed both.

Take a patient with type 2 diabetes and kidney disease:

• For risk, what matters is the diagnosis: type 2 diabetes with diabetic chronic kidney disease, the CKD stage, and evidence in the note that you monitored and managed both.
• For quality, what matters is the numbers: the latest A1c value, the latest blood pressure, and whether the eGFR and urine albumin tests were done this year.

One clear assessment can serve both. For example: "Type 2 diabetes with CKD 3b. A1c 7.4 today. eGFR 38, down from 44 in March. BP 132/78. Continue lisinopril, refer to nephrology." That supports the diagnoses for risk, and the A1c and blood pressure values for quality. A vague line like "DM, CKD, stable" can fail both.

5. Why people keep them separate

If both come from the same note, why do so many organizations treat them as two different worlds? A few reasons.

• Different teams. Risk adjustment usually sits with coding, revenue, or finance. Quality usually sits with clinical leadership or a quality department.
• Different vendors and lists. Practices often get a suspect list for risk and a care gap list for quality, sometimes from different payers, in different formats, with different deadlines.
• Different rules. Risk cares about diagnoses and documentation support. Quality cares about services, results, and dates. The logic, codes, and timelines do not line up neatly.
• Different audits. RADV asks whether a diagnosis is supported. A HEDIS audit asks whether a service is proven. Each team learns to defend against its own audit.
• Reputation. Risk adjustment has been under heavy scrutiny for coding intensity, so some organizations keep it at arm's length from clinical work. Quality feels like "good medicine," so it is easier to talk about with clinicians.

There are good reasons for some of this separation, especially for compliance. But for the doctor in the room, it creates a problem.

6. Why it is harder for providers

Payers see risk and quality as two departments. Providers see one patient.

In a single visit, a doctor may be asked to:

• Re-document every chronic condition for the year, with enough detail to hold up in an audit.
• Close quality gaps: record the A1c, recheck the BP, document the eye exam that was done elsewhere.
• Do all of this while actually caring for the patient, in a visit that is already too short.

Across the whole panel, the lists pile up. Each health plan sends its own lists for its own members. A doctor with patients in five plans may get five risk lists and five quality lists, each in a different format and on a different schedule. The plans' lists don't connect to each other, and they rarely show up inside the EHR where the doctor actually works. So before every visit, someone has to match the patient to the right plan's list.

Some of the hardest parts are not clinical at all:

• The evidence lives in different places. The diagnosis is in your note, the A1c is in the lab system, the eye exam is in a letter from the ophthalmologist, and the colonoscopy was done three years ago at another hospital.
• Timing is unforgiving. A condition documented on December 31 counts toward this year's data. The same condition documented on January 2 counts toward next year's. A quality result from last year does not close this year's gap.
• Feedback comes late. Doctors often learn months later that a diagnosis was unsupported or a gap stayed open, long after anything could be done in the visit.

7. What actually helps

Getting risk and quality right is less about coding more and more about documenting clearly once.

• Address every active chronic condition at least once a year, with its status and a plan. That is MEAT in practice.
• Write the numbers and be specific. "Type 2 diabetes with hyperglycemia, A1c 8.2 today" helps both risk and quality. "Diabetes, uncontrolled" helps neither: ICD-10 has no code for "uncontrolled," and there is no value for the quality measure.
• Record outside results with dates, such as "Colonoscopy 3/2023 at Memorial, normal."
• Use CPT II codes for results that quality measures need, so they reach the plan through the claim.
• Do one pre-visit review that covers both: which conditions need documenting this year, and which gaps are still open for this patient.
• Ask for both lists in one place. If your organization sends risk and quality separately, ask whether they can be combined for the care team.

8. The takeaway

Risk adjustment and quality are two scorecards built from the same visit. Risk asks how sick your patients are. Quality asks whether they got the right care. Both depend on the same thing: a clear, specific note that shows what you saw, what you thought, and what you did.

Separating them makes sense for audits and accounting. But for doctors, the work is the same visit and the same note. The more organizations treat them as one picture of the patient, the less time clinicians spend chasing lists, and the more time they spend on care.

How we think about it at ParaDocs

We built ParaDocs to read the note the way a clinician wrote it, and to show both sides at once: which diagnoses are supported, which need better documentation, and which quality gaps are still open, with the evidence linked to the exact sentence in the chart.

TL;DR

• What it is: A 10-year voluntary CMS Innovation Center model that pays for chronic care based on whether patients actually get better.
• Who it covers: Traditional Medicare patients with cardio-kidney-metabolic conditions, chronic musculoskeletal pain, depression, or anxiety.
• How it pays: A recurring payment per patient, with half of the Medicare share held back until outcomes are measured.
• The catch: Full payment requires at least half of your patients to hit their outcome targets. Data and documentation have to be good from day one.

For years, Medicare has mostly paid for chronic care one service at a time. A visit, a care management call, 20 minutes of time logged. Whether the patient's blood pressure or A1c actually improved was mostly someone else's problem.

The ACCESS Model (Advancing Chronic Care with Effective, Scalable Solutions) changes that question. CMS announced it in December 2025, and the first care periods started in July 2026. The idea is simple: pay a set amount to manage a patient's condition, and tie the full payment to measurable improvement against that patient's own baseline.

1. Who can join, and who can't

ACCESS is open to organizations enrolled in Medicare Part B as providers or suppliers:

• Durable medical equipment suppliers and laboratory suppliers are excluded.
• Each organization needs a Medicare-enrolled clinical director to oversee care quality and compliance.
• Medicare Advantage plans can't participate directly. The model is for Traditional Medicare patients.
• Patients enroll voluntarily or are referred by their clinician, and ACCESS runs alongside their usual care.

Care can be in person, virtual, asynchronous, or delivered through technology, including FDA-authorized devices and software. That flexibility is the point. CMS wants to see whether tech-enabled teams can manage chronic disease well at scale.

2. The four tracks

ACCESS starts with four tracks, covering conditions that affect about two-thirds of people with Medicare:

• Early cardio-kidney-metabolic (eCKM): hypertension, dyslipidemia, obesity or overweight with central adiposity, and prediabetes.
• Cardio-kidney-metabolic (CKM): diabetes, chronic kidney disease, and atherosclerotic cardiovascular disease.
• Musculoskeletal (MSK): chronic musculoskeletal pain.
• Behavioral health (BH): depression and anxiety.

3. How the money works

For 2026 and 2027, CMS published annual payment amounts per patient:

• eCKM: $360 for the initial period, $180 for follow-on periods.
• CKM: $420 for the initial period, $210 for follow-on periods.
• MSK: $180, with no follow-on period.
• BH: $180 for the initial period, $90 for follow-on periods.

Payments come monthly, at one-twelfth of the annual amount. The allowed amount includes the patient's 20% coinsurance. CMS holds back 50% of the Medicare portion and settles up after the 12-month care period.

To earn the full amount, at least 50% of an organization's aligned patients have to meet their outcome targets. A patient counts if every required measure is reported on time and the patient either reaches a control threshold or shows clinically meaningful improvement from baseline.

The measures are what you would expect: blood pressure, weight and BMI, A1c, LDL cholesterol, validated pain and function scores, and standardized depression and anxiety scores. CMS will also publish risk-adjusted outcomes in a public directory. There is also a substitute spend adjustment, which reduces payment when patients get similar services from other Medicare providers.

4. What actually changes for clinics

Under fee-for-service, the question was "did we deliver the service?" Under ACCESS, it is "did the patient improve, and can we prove it?" That shifts the hard work in three places.

Finding the right patients:

• The patients who qualify are already in your panel, but often not on your problem list.
• Prediabetes, early CKD, and depression are some of the most under-documented conditions in primary care and long-term care.
• The evidence usually sits in labs, PHQ-9 scores, and free-text notes, not in diagnosis codes.

Setting an honest baseline:

• Outcomes are measured against each patient's starting point.
• If the baseline is wrong, incomplete, or missing, you can't show improvement, and you can't defend the result.

Reporting every measure, on time:

• A patient who improves but has one missing measure does not count.
• That makes data completeness a payment issue, not just a quality issue.

5. The hard parts

ACCESS is promising, but it is not easy money.

The rates are modest. A few hundred dollars per patient per year only works with efficient, tech-enabled teams and a meaningful number of enrolled patients.

The 50% threshold is real. Organizations will need to pick tracks where they can actually move outcomes, and avoid enrolling patients they can't follow.

Patients have to opt in. Some organizations will need to do outreach and enrollment work they have never done before.

And the data plumbing matters more than ever. Labs, device readings, and symptom scores have to flow cleanly from the chart to CMS, every month, for every patient.

Payers are paying attention too. More than a dozen health plans, including UnitedHealthcare, Humana, CVS Health, Cigna, and Centene, signed an ACCESS-aligned payer pledge to move toward similar outcome-based payment. If this model works, its logic will not stay in Traditional Medicare.

6. What to do now

If you are considering ACCESS, or just want to be ready for where chronic care payment is going:

• Run your panel against the four tracks. Count how many patients qualify based on labs, vitals, and notes, not just codes.
• Check your baselines. For each eligible patient, confirm you have a recent, documented starting value for every required measure.
• Pick the track you can win. Choose the one where your team already has the clinical model and the data.
• Fix the documentation gaps first. Outcome payment depends on records that show the condition, the baseline, and the follow-up clearly.

How we think about it at ParaDocs

ACCESS rewards the same thing value-based care always has: knowing what is really going on with each patient, and being able to show it. We build clinical intelligence that reads the free text in clinical notes, finds patients whose conditions are documented but not coded, and links every finding to the evidence behind it. Under an outcomes model, that is how you find the right patients and set baselines you can defend.

TL;DR

• Old Standard: Healthcare AI was measured by how much it surfaced. More output looked like progress.
• What Changed: RADV made weak or unclear evidence a real audit and financial risk.
• Hidden Cost: When evidence isn’t clear, work doesn’t disappear—it moves to manual review, chart chasing, and reviewer fatigue.
• Key Insight: Accuracy alone isn’t enough if teams can’t quickly trace recommendations to evidence.
• The Real Shift: This isn’t a model problem—it’s a workflow design problem.
• Bottom Line: The goal is no longer more output, but accurate, defensible, audit‑ready decisions.

For years, many healthcare AI and risk adjustment tools were judged on one thing above all else: could they surface more opportunity?

If a system produced more suspects, more diagnoses, or more output, that was often enough to get attention. Volume was treated as progress.

That standard is getting harder to defend.

Today, the more important question is no longer whether a system can identify a possible condition.

It is whether an organization can clearly understand why it surfaced, what evidence supports it, and whether that evidence is visible enough to verify without redoing the work manually.

That is the shift toward healthcare AI traceability.

And in the RADV era, it is no longer optional.

RADV Changes the Burden of Proof

CMS’s RADV program exists to recover improper risk adjustment payments when diagnoses are not supported by the medical record.

When extrapolation applies, a small number of unsupported diagnoses can create population-level financial risk, not just isolated errors. A federal court vacated CMS’s 2023 extrapolation rule in September 2025; HHS has appealed, and audits continue.

CMS has also signaled a more aggressive audit posture, increasing review volume and audit coverage.

You do not need to be a compliance expert to understand what this means operationally.

When audit pressure rises, the cost of weak evidence rises with it.

A recommendation that once felt “probably fine” now has to meet a higher bar:

• Can it be defended clearly, quickly, and consistently under scrutiny?

That is what RADV changes.

It raises the burden of proof.

The Black-Box Problem Is Not Theoretical

In healthcare AI, black-box systems are often framed as a philosophical issue. Transparency. Explainability. Trust.

But in a real risk adjustment workflow, the problem is not abstract.

It is operational.

If a system surfaces a diagnosis without clearly showing the supporting note, sentence, lab, or documentation, the burden of validation does not go away.

It moves.

The work does not disappear. It moves downstream to the human reviewer.

And the organization still pays for it:

• Chart chasing
• Manual verification
• Reviewer fatigue
• Inconsistent decisions
• Slower follow-through
• Weaker confidence in submissions

A workflow can look intelligent on the surface while preserving the hardest part of the work underneath.

This is why the black-box problem in healthcare AI is not just about models.

It is a workflow design problem.

If the output cannot be traced, it cannot be trusted in a high-scrutiny environment.

Why Traceability Matters More Under V28

The V28 transition has raised the bar on evidence quality and specificity.

Fewer payment-mapped ICD-10 codes and tighter model logic mean broad suspecting is far less valuable when poorly supported.

Organizations are now operating under two simultaneous pressures:

• A stricter model environment (V28)
• A stricter audit environment (RADV)

Together, they redefine what a high-quality risk adjustment workflow looks like.

Under a looser environment, teams could tolerate high-volume suspect lists.

Now, that same approach creates noise before it creates value.

If reviewers still have to search for evidence, reconstruct context, and decide what is defensible, the system is not meaningfully reducing burden.

Healthcare AI traceability is what turns a possibility into an actionable recommendation.

Without it, the workflow still depends on manual rescue.

Accuracy Without Traceability Is Not Enough

This is a point the market still underestimates.

A model can appear highly capable. It can generate plausible output. It can look impressive in a demo.

But if an organization cannot move quickly from recommendation to evidence,

the practical value of that intelligence drops sharply in real workflows.

This is where complexity shows up, not in the model, but in the people who rely on it.

Clinicians need medical validity. Coding and compliance need supportability. Operations need efficiency. Executives need defensibility.

When reasoning is not visible, confidence breaks down.

And when confidence breaks down, adoption follows.

This is why the principle is simple:

• Accuracy without traceability is not enough.
• Because in healthcare, output is not the end of the job. The system must also support verification, adoption, and audit readiness.

What Healthcare AI Traceability Should Look Like

Traceability does not require deep model transparency or technical explanations.

In practice, it is simpler than that.

Can a user answer a few critical questions immediately?

Why was this surfaced?

What evidence supports it?

Where exactly is that evidence?

Is it current and specific?

Is it worth acting on now?

If those answers are not clear, the workflow is incomplete.

This is what strong healthcare AI traceability enables:

• Less time hunting
• More time deciding
• Less ambiguity
• More consistency
• Stronger trust across teams

That is far more valuable than high output volume that leaves proof work to the user.

The Real Issue Is Workflow Design

What RADV and V28 expose is not just a documentation problem.

They expose a workflow design problem.

Too many systems were built around generating more opportunities.

But today, value comes from helping teams act on the right opportunities with clear evidence and less waste.

That shift changes how tools should be evaluated:

• Is evidence immediately visible?
• Does the system reduce validation work or shift it?
• Does it prioritize defensible opportunities?
• Can stakeholders trust it for their own needs?
• Will it hold up under RADV scrutiny?

These are the questions that define modern risk adjustment workflows.

A More Useful Standard for Healthcare AI

The market is moving beyond output volume toward a more mature standard.

Not whether a system can find something.

But whether it can support action with enough clarity that people can trust it, use it, and defend it.

In the RADV era, healthcare AI traceability is no longer optional.

It is not a feature. It is not a preference. It is not a philosophy.

It is part of risk management.

And it is increasingly what separates high-performing workflows from those that simply create more work.

What This Means Now

In this environment, healthcare organizations should expect more than intelligent output.

They should expect workflows that make evidence visible, reduce validation burden, and support defensible decisions across teams.

That is the practical meaning of healthcare AI traceability.

And in the RADV era, it is one of the clearest indicators of whether a workflow is truly ready for what comes next.

TL;DR

As of 2026, the CMS-HCC V28 model is fully in effect for Medicare Advantage organizations. Within CMS-HCC V28 risk adjustment, risk scores are now calculated using 100% of the 2024 model. Fewer ICD‑10 codes drive payment, and RADV audits are expanding.

The takeaway is simple: Finding more diagnoses is no longer the goal.

What matters now is the ability to identify evidence-backed, model-relevant opportunities that are specific enough to support submission and defensible enough to withstand audit.

Organizations that continue to rely on broad suspecting and opaque workflows will feel increasing operational and financial strain. Those that adapt will focus on prioritization, traceability, and workflow discipline.

For years, the CMS-HCC V28 transition was discussed as something approaching the horizon. Now it is here.

For Medicare Advantage organizations, CMS has completed the phase‑in. Risk scores for 2026 are calculated using 100% of the 2024 CMS-HCC model. This is not a technical footnote. It is a structural change in how risk adjustment works day to day. (CMS)

Many organizations are already experiencing the effects. Risk scores look different. Longstanding assumptions are less reliable. Diagnoses that once “counted” now require greater specificity, clearer linkage, and stronger documentation logic.

The most important question in 2026 is no longer, “How do we find more codes?” It is, “Which opportunities are real, defensible, and worth acting on?”

CMS-HCC V28 Workflow Changes More Than Just Coding Logic

V28 is often described as a coding update. That framing is incomplete.

CMS restructured condition categories, incorporated more recent fee‑for‑service diagnosis and cost data, and introduced clinical revisions aimed specifically at conditions with high coding variation. The stated goal was to improve predictive accuracy. The operational consequence is tighter payment logic. (CMS Factsheet)

The code-mapping change alone is telling. The 2020 CMS-HCC model mapped 9,797 ICD‑10 codes to payment. The finalized 2024 model maps 7,770.

That reduction fundamentally changes the math behind broad suspecting. When fewer diagnoses drive payment, imprecision carries a higher cost. Workflows that surface large numbers of weak or outdated suspects now create noise before they create value.

That noise is not theoretical. It shows up as reviewer fatigue, repeated chart review, rework, and delayed attention to cases that are actually defensible. V28, in practice, is as much a workflow problem as it is a reimbursement one.

The Problem Is No Longer Detection Alone

For years, many risk adjustment workflows were built around a single promise: more opportunity.

Under V28, that promise is no longer sufficient.

The more relevant question is whether a workflow helps a clinician, coder, or operations leader quickly answer four things:

• Is this diagnosis still relevant under current model logic?
• Does the documentation support it with enough specificity?
• Can the supporting evidence be verified without excessive chart navigation?
• Is this worth acting on now?

This is where many “intelligent” solutions break down. Generating a long list of suggestions does not improve outcomes if users still have to manually piece together context, hunt for support, and decide whether a recommendation is usable.

In the current environment, a smaller set of evidence‑linked, defensible opportunities is far more valuable than a large queue of weak suspects. The market is shifting from code discovery to evidence‑guided action.

In a Tighter Audit Environment, Traceability Matters

At the same time V28 tightens payment logic, the audit environment has become less forgiving.

CMS’s 2023 RADV rule allowed audit findings to be extrapolated starting with payment year 2018. A federal court vacated that rule in September 2025 and HHS has appealed, so extrapolation is paused for now. Audits continue, though, and if extrapolation returns, unsupported diagnoses would expose organizations to risk well beyond individual charts.

This shift materially raises the bar for V28 documentation requirements.

Accuracy alone is no longer enough. Organizations must be able to explain why a diagnosis surfaced and where the supporting evidence lives. Documentation must be specific, traceable, and aligned with current model logic. If a recommendation cannot be traced directly to a clinician note, lab result, or other source documentation, the burden of validation remains manual.

This is why opacity is becoming operationally dangerous. A system that produces output without visible evidence may look productive on the surface, but it does not create confidence. Instead, it shifts validation work—and audit risk—back to already stretched clinical and operational teams.

In 2026, accuracy without traceability is not sufficient.

What Strong Workflows Need to Do Now

If V28 has exposed anything, it is that many workflows were designed for a less demanding environment.

Going forward, effective systems will not be the ones that generate the most activity. They will be the ones that help teams focus on the right work. Practically, that means answering five questions well:

• Is the logic current? Teams should not have to guess whether recommendations reflect the finalized V28 model. This is a present‑tense requirement.
• Is the evidence visible? Users should be able to move directly from a suggestion to its supporting documentation without excessive navigation.
• Is the opportunity actionable? The goal is not a longer queue. It is a clearer prioritization.
• Does it reduce review burden—or just relocate it? Real improvement means less chart‑chasing, not chart‑chasing in a different place.
• Can it support trust across stakeholders? Clinicians, coders, operational leaders, and executives all need to understand the reasoning behind what gets acted on.

These are no longer differentiators. They are operating requirements.

Why Workflow Visibility Matters More Under V28

One of the most expensive failures in risk adjustment is not lack of effort. It is a lack of visibility.

In many organizations, data exists. The notes exist. The review queues exist. Yet leaders cannot clearly see where work is stalling, which suspects are weak, or where documentation gaps are emerging until the problem is already costly.

Under V28, that delay is harder to absorb.

Organizations need visibility into the workflow itself: where high‑value opportunities are getting stuck, where weak suspects consume time, and where rework is being created upstream. That visibility allows teams to intervene earlier, rather than cleaning up later.

This is not a reporting convenience. It is a management tool.

The Real ROI Is Revenue Integrity

Risk adjustment is still too often framed as a volume exercise. That framing no longer fits reality.

The more relevant measure today is revenue integrity: improving confidence in what gets submitted while reducing wasted effort and downstream correction.

That includes fewer weak opportunities entering the workflow, less time hunting for justification, tighter alignment between documentation and coding action, and stronger defensibility when decisions are questioned later.

In SNF and LTC environments, especially, operational discipline matters as much as technical capability. The value of a strong workflow is not simply that it identifies opportunity. It is that it helps teams act with confidence and restraint.

A Clearer Way to Think About the Transition

The Medicare Advantage V28 transition makes one thing clear: the strategy of “finding everything and sorting it out later” is no longer sustainable.

What organizations need now is not more output. They need a better signal.

They need workflows aligned to current model logic, grounded in visible evidence, usable by real clinical teams, and capable of withstanding scrutiny.

V28 did not just change the rules. It made evidence, traceability, and prioritization non-negotiable.

And it did not just change the map. It clarified what will actually hold up in today's risk adjustment environment.

At ParaDocs, this is how we think about the V28 shift: the question is no longer whether a system can surface opportunity, but whether it can surface the right opportunity—supported by clear evidence and presented in a way teams can actually act on.

In the current phase of risk adjustment, that distinction is no longer subtle. It is decisive.

Want to see what evidence-linked workflow visibility looks like in practice?

Contact ParaDocs to learn how we think about V28-ready workflows.

Skilled Nursing Facilities (SNFs) and long-term care (LTC) providers live and die by documentation and coding. Under the Patient Driven Payment Model (PDPM), Medicare payment is tied tightly to ICD-10 diagnosis coding, comorbidities, and functional status captured on the MDS, rather than therapy minutes.

At the same time:

• SNF residents carry heavy chronic disease burdens (hypertension, vascular disease, dementia, depression, GERD, and more).
• Coding errors and omissions in SNFs are common enough that coding societies now publish SNF-specific error lists and training programs.
• Studies continue to show undercoding of comorbidities even in hospital ICD-10 data, with authors explicitly calling for automation to help coding teams.

Layer on staffing shortages, high turnover, complex MDS rules, and year-over-year PDPM refinements, and it’s no surprise that coding intensity and accuracy are under a microscope.

This is where AI—used correctly—can help SNFs and LTC operators improve coding accuracy, reduce risk, and protect revenue, without asking already-burned-out staff to do even more manual chart review.

1. Why coding accuracy is uniquely hard in SNFs & LTC

1.1 Complex residents, complex rules

SNF and LTC residents are often the sickest, most complex patients in the Medicare population:

• Common comorbid clusters include hypertension, vascular disease, dementia, arthritis, depression, and reflux.
• Many have long histories across multiple hospitals and specialists, with documentation scattered across systems and formats.

Under PDPM, that complexity is monetized through:

• A primary diagnosis reported on the MDS, mapped into one of the PDPM clinical categories for PT/OT/SLP components.
• Comorbidities that drive Non-Therapy Ancillary (NTA) and SLP components, using ICD-10 mapping tables and MDS items (e.g., Section I for diagnoses, Section O for services).

The more clinically accurate and complete the coding, the better PDPM reflects true resident acuity—and the more sustainable your margins.

1.2 Real-world coding problems in SNFs

Industry guidance and audits highlight a familiar pattern of SNF coding issues:

• Missing or nonspecific ICD-10 codes for the primary reason for the SNF stay
• Incomplete capture of surgical history and prior hospital conditions that drive PDPM groups
• Failure to code NTA comorbidities that are clearly documented in the record
• MDS items and ICD-10 codes that don’t match the clinical chart

Research on ICD-10 validity more broadly shows that secondary conditions and comorbidities are systematically undercoded, especially when coding teams are stretched thin—prompting calls for automation and better tools.

On the other side of the spectrum, overcoding and upcoding expose facilities to recoupments, penalties, and reputational risk. Guidance for providers repeatedly warns about the financial and legal risks of both undercoding and overcoding.

1.3 PDPM and “coding intensity” pressure

Early and ongoing analyses of PDPM’s impact show that:

• PDPM implementation is associated with increased coding intensity and higher Medicare expenditures per stay, even without clear changes in mortality or readmissions.
• SNFs have historically shown lower coding intensity compared with hospitals, but PDPM shifted incentives sharply toward capturing more—and more specific—diagnoses.

Regulators and researchers have flagged the need for continued monitoring to ensure PDPM incentives drive accurate—not inflated—coding.

In other words: SNF/LTC coding must now be both complete and defensible, under much closer scrutiny.

2. What “AI for coding” actually means

“AI” has become a buzzword, so let’s be specific. When we talk about AI improving coding accuracy in SNFs and LTC, we’re usually referring to a combination of:

• Natural language processing (NLP): Extracts diagnoses, procedures, and clinical clues from free-text notes, discharge summaries, consults, and scanned documents.
• Machine learning models: Learn patterns linking documentation, labs, meds, and prior codes to likely ICD-10 and HCC codes, then surface candidates for human review.
• Workflow engines: Embed these suggestions directly into coder and clinician workflows so that chart review becomes an exercise in verification, not hunting for evidence.

Real-world implementations show that AI-assisted coding and chart review can:

• Increase coding throughput without lowering accuracy
• Reduce manual chart review time by cutting down evidence search
• Improve risk score accuracy and coding quality in risk-adjusted contracts

The key is how you use these tools in a SNF/LTC context.

3. How AI actually improves coding accuracy in SNFs & LTC

3.1 Surfacing “hidden” comorbidities and PDPM drivers

AI systems can scan entire longitudinal charts, including:

• Hospital discharge summaries
• Specialist consults
• Therapy and nursing notes
• Lab results and medication lists

…to identify conditions that are documented but not coded, such as:

• Chronic kidney disease stages based on labs and nephrology notes
• Neurologic conditions (stroke, aphasia, dysphagia) that drive SLP and NTA components
• Surgical history and complex wounds that impact PDPM classification

NLP and AI-based risk adjustment tools are already used to:

• Extract additional diagnoses from unstructured documentation
• Map those diagnoses to ICD-10 and HCC codes
• Flag gaps where chart evidence supports a code that is missing in claims

For SNFs/LTC, the same approach can be tuned to PDPM:

• Aligning AI-identified diagnoses with PDPM clinical categories and NTA comorbidities
• Highlighting residents who appear under-classified relative to their documented complexity

3.2 Reducing human error and variability

Manual coding in SNFs is vulnerable to:

• Fatigue and cognitive overload
• Inconsistent familiarity with PDPM rules and mapping tables
• Varied experience among MDS coordinators and coders

AI-assisted coding tools help by:

• Auto-suggesting primary and secondary diagnoses based on chart patterns
• Checking that codes used on the MDS are actually supported by the documentation
• Standardizing application of PDPM mapping rules and local coding policies

NLP-driven automation has been shown to reduce missed diagnosis codes and manual errors in other settings, and vendors report improved coding accuracy when AI suggestions are systematically reviewed by coders instead of starting from a blank slate.

3.3 Making chart review faster and more focused

In many SNFs, coding accuracy hinges on one thing: whether someone has time to read the chart carefully.

Recent work on AI-assisted chart review shows that:

• “Computer-assisted chart review” lets skilled human reviewers work faster and more accurately, using the AI to pre-locate relevant evidence while humans adjudicate.
• In value-based contexts, AI-powered workflows have cut Annual Wellness Visit coding time while maintaining or improving HCC capture.

For SNFs and LTC, similar workflows can:

• Pre-assemble evidence bundles per resident (e.g., all HF evidence, all diabetes evidence)
• Present candidate codes side-by-side with evidence so coders and MDS nurses can confirm or reject with a click
• Generate PDPM impact previews (how a code change affects case-mix groups) for internal QA and education.

The result is more consistent coding decisions in less time, without turning the process into a black box.

3.4 Ensuring alignment between MDS, ICD-10, and documentation

PDPM hinges on tight alignment between:

• MDS diagnosis fields (e.g., I0020B, I8000)
• ICD-10 codes chosen for those fields
• The underlying medical record (hospital discharge, physician orders, assessments)

Coding and billing experts stress that incorrect or incomplete diagnosis capture leads to lower-paying PDPM groups and audit exposure.

AI can continuously cross-check:

• Does every PDPM-relevant diagnosis on the MDS have traceable support in the chart?
• Are there chart-documented conditions that should be present on the MDS but aren’t?
• Are surgical and procedural histories accurately reflected, especially when they change PDPM classification?

Instead of relying solely on sporadic manual audits, facilities can run daily or weekly AI checks to keep MDS–ICD-10–chart alignment tight.

3.5 Monitoring undercoding, overcoding, and intensity trends

Because PDPM and risk-based models are under scrutiny for coding intensity, facilities need to monitor not just individual codes but trends:

• Are certain clinicians or units consistently undercoding comorbidities compared with documented burdens and peers?
• Are there sudden shifts toward higher-paying PDPM groups without corresponding clinical changes—a signal of possible upcoding or documentation drift?
• Are known risk areas (e.g., NTAs, SLP comorbidities) being coded consistently and backed by strong evidence?

AI-enabled analytics and risk adjustment platforms already provide these kinds of coding accuracy dashboards for health plans and ACOs; the same principles can be adapted to SNF/LTC operations to support internal compliance and education.

4. Design principles for safe AI coding in SNFs & LTC

AI can absolutely make coding more accurate. It can also create new failure modes if used poorly. A few non-negotiables for the SNF/LTC environment:

4.1 AI as copilot, not autopilot

Evidence and expert commentary around AI-assisted chart review converge on the same model:

• Skilled human reviewers, assisted—not replaced—by AI.

For coding in SNFs/LTC, that means:

• AI suggests codes and evidence; coders and clinicians confirm
• MDS nurses remain accountable for final assessments
• Coding staff can override suggestions and provide feedback to improve the system

4.2 Grounding in the full record, not just text snippets

Safe AI coding must be grounded in:

• Structured data (labs, meds, vitals, orders)
• Unstructured notes (hospital discharge, nursing, therapy)
• MDS items and ICD-10 mappings

PDPM itself assumes that ICD-10 choices on the MDS reflect the full clinical picture, not just phrases in a single note.

AI coding tools should mirror that assumption, not shortcut it.

4.3 Transparent “receipts” for every suggested code

Best-in-class AI risk adjustment and coding tools emphasize:

• Clear evidence lists (“this diagnosis is suggested because…”), including note excerpts and data points
• Simple UIs for coders to accept, modify, or reject each suggestion
• Audit logs showing who accepted which codes and when

For SNFs, this transparency is essential if you ever need to:

• Defend coding choices in the face of an audit
• Demonstrate that PDPM changes reflect genuine acuity, not undocumented upcoding

4.4 Governance: coding, compliance, and clinical all at the table

Finally, AI coding in SNFs/LTC needs shared governance:

• Coding and HIM leaders (accuracy, guidelines, audit-readiness)
• Clinical leadership (appropriateness of diagnoses, PDPM ethics)
• Compliance and legal (regulatory risk, vendor contracts)
• IT/analytics (data integration, monitoring, drift detection)

Their job is to define guardrails:

• Where AI can and cannot make suggestions
• Minimum evidence thresholds for codes to be surfaced
• Monitoring for error patterns, drift, or unsafe shortcuts

5. A pragmatic roadmap for SNF & LTC operators

If you’re considering AI to improve coding accuracy, you don’t have to “boil the ocean.” A realistic progression:

• Baseline measurement
• Quantify current denial rates, audit outcomes, and PDPM case-mix accuracy.
• Sample charts to measure undercoding of common comorbidities vs documentation.
• Start with AI-assisted chart review for high-value residents
• Post-acute, medically complex, or high-NTA residents where coding has the biggest margin impact.
• Use AI to pre-assemble evidence and suggested codes; coders validate.
• Expand to continuous MDS–ICD–chart alignment checks
• Daily/weekly AI runs to flag mismatches and missing PDPM drivers.
• Add analytics for intensity and compliance monitoring
• Track coding intensity trends, case-mix shifts, and undercoding patterns by unit or provider.
• Iterate with training and feedback
• Use AI findings to guide coder and MDS nurse education.
• Feed coder feedback into model refinement.

Done well, AI in SNFs & LTC is not about “letting the machine code everything.” It’s about giving your clinicians and coders superpowers: the ability to see the full story in the chart, consistently, under PDPM’s pressure—without burning out.

If you work in Medicare Advantage (MA), ACO REACH, or any value-based program, CMS’s move to the HCC v28 risk-adjustment model is not just a payer problem—it directly affects how your documentation turns into dollars and how your sickest patients are “seen” by the system.

Here’s a 5-minute, clinically grounded walkthrough you can use with your admin, coding, and finance teams.

Minute 1 – What is the HCC model (and why should I care)?

CMS’s Hierarchical Condition Category (HCC) model is the engine that turns diagnoses into a risk score (RAF). That RAF score adjusts per-member-per-month payments in:

• Medicare Advantage (MA)
• ACO REACH and other CMMI models
• Certain risk-sharing and total cost of care contracts

Key points:

• Average Traditional Medicare beneficiary = RAF 1.0.
• Higher clinical complexity → more HCCs → higher RAF → higher benchmark/payment.
• HCCs are not every ICD-10 code; they are a curated set of clinically similar condition groups (e.g., “Diabetes with chronic complications”).

Historically, the model (v24 and earlier) has been criticized because plans and provider groups could increase payment through aggressive documentation and coding practices, even when true disease burden didn’t change.

That’s the backdrop for HCC v28.

Minute 2 – What changed in HCC v28?

CMS rebuilt the model and is phasing it in from 2024–2026.

Major shifts:

• Directly ICD-10–based model
• v24 used an ICD-10 → ICD-9 → condition → HCC pipeline.
• v28 maps directly from ICD-10 to HCCs, improving clinical specificity and alignment with current coding practice.
• More granular HCC structure
• The number of condition indicators increased (from 86 payment HCCs in V24 to 115 in V28), providing finer stratification of disease categories—especially for cardiovascular disease, kidney disease, mental health, and diabetes.
• Removal / down-weighting of “high-leverage, low-cost” conditions

CMS explicitly targeted HCCs that had become magnets for coding intensity rather than true cost:

• Protein-calorie malnutrition
• Certain angina and atherosclerosis codes
• Some conditions with highly variable documentation practices These are suppressed or significantly down-weighted in v28.
• Constrained coefficients for certain hierarchies
• Diabetes HCCs (no complications, chronic complications, acute complications) share constrained coefficients to reduce outsized payment differences driven by coding detail alone.
• Similar constraints are applied to some heart failure categories.
• Calibration to more recent FFS data
• Coefficients are recalibrated on more recent Traditional Medicare data, which affects relative weights across chronic conditions and interactions.
• In plain language: v28 is stricter, more granular, and explicitly designed to blunt the impact of “upcoding-heavy” diagnoses.

Minute 3 – How and where is v28 being phased in?

CMS is not flipping a single switch—it’s a multi-year blend:

• 2024: Start of v28 phase-in for MA (partially blended with v24).
• 2025: CMS applies a 67% / 33% blend (67% v28, 33% v24) for MA, and similar blending for ACO REACH/Kidney Care Choices with their Part C risk adjustment alignment.
• 2026: Expected to be fully v28-based for MA (barring policy changes).

In parallel, MedPAC estimates that MA payments are ~20% higher than FFS in 2025 (about $84B more), driven by coding intensity and favorable selection. CMS and MedPAC explicitly point to model reforms like v28 and coding intensity adjustments as key levers to rein this in.

For you, that means:

• Your RAF scores may drift down even if you code the same way as last year.
• Benchmarks in ACO REACH/Kidney Care models are also being recalibrated with Part C (v28) alignment and caps on risk score growth.

Minute 4 – What does v28 mean for providers on the ground?

From a provider and coding perspective, v28 changes what “counts,” how much it counts, and how risky coding gaps become.

1. Some diagnoses “matter less” for payment

Conditions that historically yielded meaningful RAF bumps (e.g., malnutrition, some vascular codes, angina) may now:

• Map to non-payment HCCs, or
• Have substantially lower coefficients.

This doesn’t mean they’re clinically unimportant—but they will drive less revenue signal.

2. Complexity and combinations matter more than isolated labels

Because of constrained coefficients and more granular hierarchies, the model better differentiates:

• Truly complex chronic disease (e.g., diabetes + advanced CKD + heart failure)
• Versus “diagnosis stacking” without consistent evidence in the chart.

That increases the importance of:

• Longitudinal evidence (labs, imaging, notes)
• Clear links between ICD-10 codes and the assessment & plan

3. Coding intensity is under a microscope

A growing body of literature shows that differential coding practices in MA drove billions in additional payments:

• One study in Annals of Internal Medicine estimated $33B in extra payments in 2021 from coding differences, with CMS explicitly citing v28 as a tool to reduce the MA–FFS risk-score gap.
• MedPAC finds that most MA enrollees are in contracts where coding intensity exceeds CMS’s adjustment, and cites multiple Health Affairs / HSR studies documenting persistent upcoding.

Translation: CMS expects risk scores to come down—and is okay with that. If your organization’s RAF performance depends on aggressive rather than accurate documentation, v28 will expose that.

4. SNFs, LTC, and complex-care providers are in a “barbell” zone

For SNFs, LTC, and high-acuity primary care:

• Patients often truly have high disease burden (CKD 4–5, CHF, dementia, frailty, multimorbidity).
• But documentation is fragmented across EHR encounters, hospitalizations, and post-acute transitions.

In v28, missing a single key condition (e.g., advanced CKD) in a year can have a larger relative impact on RAF than before, because the model leans more heavily on accurate capture of genuinely expensive chronic illness and less on edge-case diagnoses.

Minute 5 – How should providers respond?

Here’s a pragmatic playbook you can act on now.

1. Re-educate teams on which conditions “drive the model”

Don’t treat v28 as just a payer issue. Your clinical, coding, and quality teams should know:

• Which chronic condition families are most impactful in v28 (e.g., CHF, CKD, diabetes, pulmonary disease, serious mental illness).
• Which historically “favorite” codes (e.g., protein-calorie malnutrition) are now much less impactful.

Use short, condition-focused refreshers rather than long coding seminars—anchored in real patients and charts.

2. Tighten the link between documentation and ICD-10

In a stricter model, sloppy documentation is more expensive:

• Make sure Assessment & Plan clearly supports each HCC-relevant diagnosis.
• Reduce over-reliance on problem list carry-forward; reconcile active vs historical conditions.
• Encourage clinicians to document severity and chronicity (e.g., CKD stage, HF with reduced EF, insulin-use in diabetes) when clinically accurate.

3. Shift from “chasing codes” to clinical reasoning + evidence

Risk adjustment under v28 rewards:

• Coherent clinical stories
• Longitudinal follow-up of chronic disease
• Evidence of active management (labs, meds, referrals, plans)

That’s why we focus on clinical reasoning AI rather than just NLP:

• Instead of surfacing every possible code, we prioritize conditions where documentation and evidence clearly support persistent, costly disease.
• We align nudges and gap closure recommendations with v28’s structure, so you’re not optimizing for a model that’s already going away.

4. Stress-test your RAF and revenue under v28 assumptions

Work with analytics / actuarial partners to:

• Simulate RAF under v24 vs v28 using your existing population.
• Identify which conditions, practices, or facilities drive the largest delta.
• Prioritize high-impact cohorts: multi-morbid patients in SNFs, LTC, complex primary care panels, dual-eligibles.

This turns v28 from a source of “rate shock” into a roadmap for documentation and care improvement.

HCC v28 is here and phasing in through 2026—it’s not theoretical anymore.

The model is more granular and stricter, directly ICD-10–based and explicitly targeting high-leverage coding intensity.

Expect RAF scores and payments to compress, especially where coding intensity was driving revenue.

Clinically coherent documentation, longitudinal evidence, and accurate capture of true complexity will matter more than ever.

If you’re in SNF/LTC or complex primary care, v28 can help you—if you have the data and workflows to surface and document the real burden of disease consistently.

Every clinician knows what administrative drag feels like:

• Clicking through 12 EHR tabs to confirm what you already know about the patient
• Re-typing the same story into different templates
• Hunting for one lab value, one consult line, one sentence in a 20-page PDF to prove a diagnosis or justify a code

Studies keep confirming the same story:

• Primary care physicians spend a median 36.2 minutes in the EHR for each 30-minute visit, including about 6.2 minutes of after-hours “pajama time” per encounter.
• Documentation and clerical work are primary drivers of EHR-related burnout.
• Ambient AI “scribe” tools have already been shown to cut documentation time by roughly 20–30%, and reduce burnout and after-hours EHR use in early studies.

That’s a step forward—but it mostly tackles note writing.

The bigger opportunity is to apply clinical reasoning AI to the rest of the work: the chart review, evidence hunting, and measure abstraction that sit around the note.

When you do that, you’re not just shaving 20–30% off; you can realistically push 99% of the manual search and assembly work onto the machine for very specific workflows—leaving humans to review, correct, and sign off.

This article breaks down:

• Where administrative drag really comes from
• What we mean by “clinical reasoning AI” (beyond dictation or a scribe)
• How it can remove ~99% of manual hunting in targeted workflows
• How to design it safely, without creating new risks
• A practical roadmap for providers

1. The real shape of administrative drag

Most discussions of “documentation burden” focus on the note itself. The literature paints a broader picture:

• EHR time includes note entry, order entry, inbox management, and extensive chart review.
• Clinicians report that documentation burden and EHR usability are major contributors to burnout and intent to reduce clinical hours.

When you decompose a typical clinic day, “admin drag” shows up in four places:

• Note composition
• Turning a complex visit into structured documentation, often under time pressure.
• Evidence gathering inside the EHR
• Scrolling through years of labs, imaging, consult notes, hospital summaries, and scanned PDFs to confirm what is already clinically clear.
• Abstracting the same information for other stakeholders
• Quality measures, registries, care gaps, risk adjustment, prior auths, internal audits.
• Re-typing or re-formatting for other systems
• Copy-pasting into forms, portals, or spreadsheets because the EHR doesn’t speak the language those programs need.

Classic EHR-related burnout papers stress that the burden is not just typing; it’s the cognitive overhead and fragmentation of doing all this in tiny, disconnected steps.

2. What “clinical reasoning AI” actually is

Most of the public conversation so far has focused on ambient AI scribes—tools that listen to a visit and draft the note. These are important, and early studies show they:

• Reduce time in notes and documentation,
• Improve self-reported burnout and well-being, and
• Are widely perceived by clinicians as helpful.

But clinical reasoning AI goes further. It combines:

• Structured data understanding
• Labs, meds, vitals, problem lists, orders, flowsheets, prior codes.
• Unstructured text understanding
• Free-text notes, consults, discharge summaries, radiology and echo reports.
• Temporal reasoning
• Seeing trajectories (e.g., eGFR decline, weight changes, worsening dyspnea) instead of isolated snapshots.
• Task-aware summarization and evidence selection
• Knowing which facts matter for this exact task: risk adjustment, CCM eligibility, quality measure, prior auth, etc.

And crucially:

Clinical reasoning AI is not just “writing notes”; it is assembling proof for whatever clinical or administrative question you are trying to answer.

This is aligned with recent commentary and guidance on generative AI in medicine, which emphasize task-specific, context-aware AI that augments clinical judgment rather than replaces it.

3. Where the “99%” comes from: targeting the right workflows

The 99% reduction is not a claim that all admin work disappears. It’s about narrow, high-friction workflows where:

• 95–99% of the effort is search and assembly,
• 1–5% is clinical review and judgment.

In those workflows, clinical reasoning AI can do almost all of the first part.

3.1 Example 1 – Evidence packets for coding & risk adjustment

Today, building evidence for a suspected chronic condition often looks like:

Manually searching the chart for:

• Diagnosis mentions
• Corresponding labs/imaging
• Relevant meds and specialty notes
• Copying snippets into a template or note
• Double-checking dates, ICD codes, and attribution

Chart review and evidence gathering at scale is so expensive that a recent J Gen Intern Med paper explicitly argues for computer-assisted chart review, with AI doing the pre-work so humans can focus on decisions.

With clinical reasoning AI, you can invert the workflow:

• The AI pre-scans the entire longitudinal record for structured + unstructured evidence related to a condition (e.g., CKD, HF, COPD, diabetes with complications).

It assembles a task-specific bundle:

• Key labs and trends
• Pertinent imaging findings
• Chronological note excerpts that support the diagnosis
• Links back to the original sources in the EHR

The coder or clinician’s job becomes:

• Scan a curated bundle, correct anything off, and approve or reject.

If your prior process was 200 clicks and 10–15 minutes of hunting to build each packet, and the new process is 5–10 clicks and a 60–90 second review, you’ve offloaded well over 90% of the “drag” onto the AI.

3.2 Example 2 – Clinical documentation for routine visits

Ambient AI documentation platforms already show:

• ~20–30% reduction in documentation time,
• Significant reductions in after-hours note completion,
• Improved clinician experience.

Clinical reasoning AI builds on top of that by:

• Auto-linking each line in the note to evidence already in the chart (labs, meds, imaging, prior notes).
• Suggesting problem list updates and codes aligned with the story you just told.
• Preparing downstream artifacts (e.g., documentation snippets for CCM enrollment, care gaps, risk codes) at the same time.

Again, if the old workflow required:

• Writing the note,
• Then going back to search for evidence for a separate gap form,
• Then re-typing data into another system…

…a reasoning-based AI that drafts the note and pre-populates the rest can dramatically reduce the total admin overhead per encounter, not just note time.

3.3 Example 3 – Manual abstraction for quality and registries

Scoping reviews on documentation burden and chart abstraction describe:

• Manual abstraction as time-consuming, costly, and variably reliable,
• With hundreds of hours per metric per year in some organizations.

Clinical reasoning AI can:

• Read the entire chart,
• Apply measure-specific logic (e.g., inclusion/exclusion criteria, time windows),
• Pre-fill abstractions and highlight ambiguous cases.

Here again, humans stay in charge:

• They adjudicate edge cases,
• Correct misclassifications,
• Provide feedback to improve the system.

But they no longer spend their day doing needle-in-haystack searches. For a large fraction of charts, the AI can take on essentially all of the “find and assemble” work.

In mature implementations, this is where you see the 99% reduction in “drag”: most of the clicks and hunting disappear; the judgment stays.

4. Why this is different from “just another AI scribe”

Ambient AI scribes are a powerful first step—but they’re mostly focused on transcription and summarization of the current encounter.

Clinical reasoning AI adds three critical capabilities:

• Longitudinal memory
• It understands how today’s encounter fits into the entire chronic disease story, not just what was said in the room.
• Task-specific rigor
• It knows what counts as evidence for CKD progression vs. HCC coding vs. CCM vs. a quality metric, and it selects information accordingly.
• Bi-directional workflow integration

It doesn’t just produce notes. It also pre-populates:

• Problem lists and codes (for clinician review)
• Care gap artifacts
• Registry fields
• Risk/evidence packets

Systematic reviews of AI in documentation emphasize that the most valuable tools structure data, annotate notes, identify trends, and detect errors, not just dictate.

Reasoning-based systems are built to do exactly that.

5. Safety, risk, and human control

Of course, there are real risks:

• Over-reliance on AI summaries
• Hallucinated or misattributed evidence
• Subtle biases in which conditions or codes the AI “sees” first

Recent commentaries and reviews on AI scribes and generative AI in medicine underscore the need for:

• Human-in-the-loop control
• Clinicians and reviewers must be the final authority.
• Transparent links back to source data
• Every suggested code, gap, or registry field should come with “receipts” (e.g., which note, which lab, which line).
• Clear scope limits
• Define what the AI can’t do (e.g., no autonomous ordering, no silent code changes).
• Continuous monitoring and governance
• Track error rates, disagreement patterns, and user feedback over time.

The goal is not to automate judgment. It is to automate the grunt work that surrounds judgment.

As one perspective in JAMA Network Open on documentation burden and ambient AI put it, the most promising path is AI that restores human time and attention to the parts of care where it matters most—not AI that tries to be the clinician.

6. A practical roadmap for providers

If you’re a provider organization looking at “clinical reasoning AI,” a realistic roadmap might look like this:

• Quantify your baseline drag

Use EHR logs and surveys to quantify:

• Time in notes
• Time in chart review
• After-hours EHR time
• Align this with burnout data and internal feedback.
• Start with one narrow workflow
• Example: generating evidence packets for a single high-value chronic condition or for a specific registry.
• Pilot an AI system that pre-assembles evidence; keep humans firmly in charge of final decisions.
• Instrument the before/after
• Measure clicks, time per case, and error rates before and after AI support.
• Track perceived burden and burnout for directly affected staff.
• Scale along “reasoning-heavy, typing-light” axes
• Move from one disease to multiple, from one registry to a cluster of related metrics, from one payer program to others.
• Build governance and training into the expansion
• Create policies around validation, auditing, and incident response.
• Make AI literacy part of ongoing clinical education.

Done well, the destination looks like this:

• Clinicians and clinical ops teams still make the calls. The system does almost all of the searching, assembling, and formatting around them.

In that world, a 99% reduction in “drag” for the worst offender workflows isn’t magical. It’s just what happens when you finally stop asking humans to be search engines.

Large language models (LLMs) are suddenly everywhere in healthcare.

Vendors promise auto-generated notes, “one-click” HCC documentation, and “AI scribes” that quietly listen in the background. Early studies and pilots show that AI scribes can reduce time spent on documentation and improve perceived efficiency and satisfaction, but they also raise questions about accuracy, safety, and over-reliance.

For clinicians exhausted by clicks and templates, this all sounds tempting.

But there’s a hard truth: LLMs are very good at writing — and very bad at owning clinical responsibility.

If you treat them like a smarter dictation tool, you’re mostly fine. If you treat them like a clinical brain, you’re in trouble.

This article breaks down what LLMs systematically get wrong about documentation, why that matters in value-based care, and what a safe provider-facing AI stack actually needs.

1. LLMs don’t “understand” the record — they predict text

An LLM doesn’t see a patient; it sees tokens.

Given a prompt, it produces the most statistically likely continuation based on its training data. That’s why it can write fluent, human-sounding notes. But:

• It doesn’t truly “know” whether a patient has CKD stage 3b or 4.
• It doesn’t inherently care about temporal consistency (diagnosis dates, chronic vs resolved, annual recapture).
• It can hallucinate diagnoses, medications, or prior history that were never documented.

Recent work on LLMs in medicine and note generation has shown that even strong models can generate false but plausible clinical details (“hallucinations”), including unsupported diagnoses or findings, which pose real safety risks when outputs are not tightly constrained.

For narrative tasks (like drafting patient letters), that risk is manageable. For problem lists and HCC documentation, it’s unacceptable if used naively.

2. “Pretty summaries” are not clinically faithful notes

LLMs are fantastic at turning a messy chart or transcript into a clean paragraph. But clinical documentation is not just “the story”; it’s the data backbone for:

• Care continuity
• Quality measurement
• Risk adjustment (HCC / RAF)
• Legal and regulatory audit trails

Recent evaluations of AI-generated clinical notes emphasize that high-quality drafts still require careful physician review, and that field-specific requirements (precision, attribution, temporal accuracy) are easy to violate when models are optimized mainly for fluent language.

Typical failure modes when you ask an LLM to “write the note” from raw text or transcripts include:

a. Loss of critical qualifiers

“CKD” vs “CKD 3b due to diabetic nephropathy”

“Depression” vs “Recurrent major depressive disorder, in partial remission”

“History of” vs active conditions

LLMs tend to simplify, because their objective is fluent text, not coding precision. That simplification can decrease specificity, which hurts both care quality and revenue integrity in risk-based contracts.

b. Confusing past and present

Models can easily:

• Pull in old diagnoses as if they are current.
• Repeat historical problems that have since resolved.
• Miss the subtleties of “suspected vs ruled out vs confirmed.”

In value-based care, that means overstating some risk and missing others — a nightmare from both compliance and actuarial perspectives.

c. Making up “linkages” that never existed

To satisfy the style of a good SOAP note, an LLM might assert relationships like:

• “The patient’s CHF is likely due to long-standing uncontrolled hypertension…”
• even if the chart never actually said that.

To a human, it reads reasonable. To a regulator or malpractice lawyer, it’s a statement of clinical judgment attributed to you.

3. LLMs don’t know your contracts, your coding rules, or your risk model

Even if an LLM is fine-tuned on clinical notes, it still doesn’t:

• Know your payer-specific rules (MA vs MSSP vs commercial ACO).
• Align with your internal coding compliance policies.
• Understand HCC v28 vs v24, local coverage determinations, or recent audit findings.

If you simply ask:

• “Generate the assessment and plan and list all relevant diagnoses.”
• You’ll get something fluent, but not necessarily compliant or contract-aligned.
• You need logic layers outside the LLM that:
• Enforce which codes can be suggested and when.
• Check that documentation supports any suggested chronic condition.
• Respect visit type, encounter context, and policy constraints.

The LLM can draft; it must not be the source of truth for what’s allowed.

4. Garbage in, fluent garbage out

Most EHR data is:

• Fragmented across encounters and systems
• Full of copy-forwarded problem lists
• Riddled with typos, legacy codes, and outdated diagnoses

An LLM that ingests all of this unfiltered and then “summarizes” it will:

• Echo existing errors
• Amplify outdated information
• Bury what little signal you had under even more text

Studies on documentation burden and EHR usability show that clinicians already face information overload, long notes, and complex navigation, all of which increase cognitive load and error risk.

If you simply add generative AI on top of noisy, inconsistent data, you don’t get “smart notes” — you get faster wrongness in a more polished paragraph.

The model isn’t broken; the data pipeline and governance are.

5. The invisible cost: trust and cognitive load

The promise is: “AI will reduce documentation burden.” But if clinicians:

• Don’t know what the AI looked at
• Can’t see why it suggested a diagnosis
• Have to re-read everything to be sure nothing is fabricated

…then you’ve simply moved the burden from “typing notes” to “auditing notes.”. We already know that EHR work and after-hours charting are strongly associated with burnout, information overload, and diagnostic risk.

Two predictable outcomes with naive LLM deployment:

• Overtrust – Clinicians accept AI-generated notes with minimal review.
• Risk: silent propagation of hallucinated diagnoses and false linkages.
• Undertrust – Clinicians don’t believe the AI, so they redo the work manually.
• Result: more clicks, more burnout, no ROI.

The only sustainable path is calibrated trust: clinicians know exactly where the AI is strong, where it’s weak, and what they remain responsible for.

6. What “good” looks like: LLMs as assistants, not authors

For provider organizations, the question isn’t “LLMs: yes or no?” It’s how you use them.

A safer pattern looks like this.

a. Structured data first, language second

Pipeline order should be:

• Ingest and normalize data: meds, labs, vitals, claims, problem lists, notes.

Apply deterministic and ML logic for:

• Phenotypes (e.g., suspected CKD based on eGFR over time)
• Quality and risk rules (e.g., open gaps, recapture needs)

Then let the LLM turn that structured, auditable evidence into:

• Drafted assessment language
• Patient-facing explanations
• Follow-up letters or summaries

The LLM explains and rephrases; it does not invent the underlying facts.

b. Evidence-linked suggestions

Every AI-driven diagnosis suggestion should come with receipts:

• “Suggested: CKD 3b”

Because:

• eGFR 42, 44, 40 over last 18 months
• UACR 60 mg/g
• Existing diabetes + hypertension codes

LLM can then draft:

• “Chronic kidney disease stage 3b, likely secondary to long-standing diabetes and hypertension, with stable but reduced eGFR (40–44) and albuminuria (UACR 60 mg/g).”
• But the decision to accept that diagnosis remains with the clinician, who can verify the evidence in one glance.

c. Tight scope: from “write the note” to “draft this part”

Instead of giving the model the entire chart and saying “write everything,” constrain it:

• “Draft the HPI from this transcript.”
• “Draft the patient letter explaining new diagnosis X.”
• “Draft the provider-to-provider summary for this referral.”
• “Turn this evidence bundle into a concise assessment sentence.”

This reduces hallucination risk and keeps the model focused on what it does best: turning clear inputs into clear language.

d. Clear human-in-the-loop workflows

Design the system so that:

• Clinicians see every suggested change to the problem list and diagnoses.
• Acceptance is explicit (check boxes, “apply all,” etc.).
• You log what was accepted, modified, or rejected — and use that feedback to continuously tune prompts and policies.

Safety frameworks now being proposed for clinical LLM use emphasize exactly this kind of error taxonomy, safety assessment, and interface design, rather than just “turn it on and see what happens.”

Without this, you’re flying blind on both safety and performance.

7. Questions providers should ask any “AI documentation” vendor

Before you roll out an AI scribe or LLM-based documentation tool, ask:

• What exactly does the LLM do?
• Does it only draft language, or does it also pick diagnoses/codes?
• Where are deterministic rules vs. where is free-text generation?
• How do you prevent hallucinations?
• Do you enforce that all diagnoses come from a structured evidence set?
• Can the model introduce a condition that isn’t supported by chart data?
• Can I see the evidence behind each suggestion?
• Is there a one-click view of “why this diagnosis is suggested”?
• Are you using phenotypes, risk rules, or just text pattern matching?
• How do you handle my specific contracts and coding policies?
• Can the system be tuned for my HCC version, my payers, my internal guidelines?
• Who is on the hook if the AI encourages noncompliant coding?
• What data do you train on?
• Is my data used to train your models? How is PHI protected?
• Can we keep a private, tenant-specific model boundary?
• How do clinicians stay in control?
• Is there an easy way to decline or edit AI suggestions?
• What does the audit trail look like?

If a vendor can’t answer these clearly, they’re not ready to handle your documentation.

8. The real opportunity: documentation as a byproduct of good care, not more typing

Used well, LLMs can help:

• Turn conversations into high-quality draft notes.
• Turn complex evidence into succinct, guideline-aligned assessments.
• Turn population-level analytics into patient-specific nudges.

But that only works when:

• The clinical facts are governed outside the model (data layer, rules, phenotypes).
• The model’s job is language, not judgment.
• Clinicians are in the loop, not out of it.

Documentation in value-based care is too important to outsource to a black box that “sounds smart.”

The future isn’t “AI that writes your chart.”

It’s AI that quietly does the data work, surfaces the right facts, and lets you be the doctor — while making the documentation fall out naturally from that process.

That’s the bar we should hold LLMs — and their vendors — to.

Most clinicians know that RAF scores matter for Medicare Advantage and other value-based contracts. Fewer have a clear mental model of what a RAF score actually is and what it means for day-to-day clinical and operational decisions.

In under 10 minutes, here’s how to think about RAF like a vital sign for your clinic—not just a finance metric.

1. RAF in one sentence

CMS uses risk adjustment to translate a patient’s diagnoses and demographics into a risk score that predicts their expected cost of care.

A RAF score of 1.0 ≈ the average Traditional Medicare beneficiary.

A RAF of 2.0 ≈ roughly twice the expected cost; a RAF of 0.5 ≈ half.

Each year, CMS publishes the HCC (Hierarchical Condition Category) model and coefficients that turn ICD-10 codes plus demographics into that score.

So your clinic’s RAF profile is not just an abstract number—it’s CMS’s best guess of how complex and costly your panel is.

2. How RAF turns into dollars

A RAF score is used differently depending on the program, but the logic is similar:

• CMS estimates a baseline cost for a “1.0” beneficiary in a given county/year.
• That baseline is multiplied by the patient’s RAF score (plus other factors like demographics and geography).

The resulting adjusted amount flows into:

• Medicare Advantage payments to plans,
• Benchmarks and capitation in ACO REACH and other CMMI models,
• And, downstream, into shared savings, PMPMs, and bonuses for providers.

This has three practical implications for your clinic:

• If your panel is sicker than average but your RAF scores are low (because conditions are under-documented), you are underpaid relative to your true workload.
• If your RAFs are high only because of aggressive coding, you may be temporarily “overpaid” and sitting in the crosshairs of audits.
• Over time, CMS keeps renormalizing the model so the average FFS risk score stays at 1.0, even as coding and populations shift.

RAF is not “free money for codes.” It’s how the system decides how much resource should follow your patients.

3. What a RAF score tells you about your panel

Think of RAF in three layers:

3.1 Patient-level RAF

At the individual level, RAF is a proxy for multimorbidity and expected cost:

• More serious chronic disease (HF, CKD, diabetes with complications, COPD, serious mental illness) → higher RAF.
• Age, disability, and dual-eligibility status also contribute.

For a front-line clinician, this is a rough marker of complexity, not a clinical risk score like CHADS₂-VASc or MELD. But when RAF is much lower than your clinical impression, it’s a signal that documentation and coding aren’t reflecting reality.

3.2 Clinic-level RAF (your “panel profile”)

Aggregated across your panel, RAF starts to say things like:

• “Our average patient is 1.4× as complex as the average Medicare beneficiary.”
• “Our SNF panel has a much higher RAF than our community panel, which should be reflected in staffing and funding.”

High, accurate RAF at the clinic level should correlate with:

• Higher actual disease burden in your population,
• More intensive clinical work, and
• Appropriately higher per-patient revenue to support that work.

3.3 System-level RAF and policy

From CMS and MedPAC’s perspective, RAF scores also reveal:

• Coding intensity—how much of the difference between MA and FFS spending is driven by documentation patterns rather than real morbidity.
• The need to adjust models (e.g., HCC v28) and apply coding intensity adjustments to prevent overpayment.
• That’s why RAF is under growing regulatory scrutiny, not just actuarial interest.

4. The good, the bad, and the ugly of RAF for your clinic

4.1 The good: when RAF works in your favor

When used as intended, risk adjustment can:

• Protect clinics caring for sicker patients from financial penalty, by ensuring payments rise with disease burden.
• Support investment in care management, social work, pharmacy, and behavioral health for high-RAF panels.
• Make quality comparisons fairer, because outcomes are judged after adjusting for baseline risk.

For SNFs, LTC, and complex PCPs, accurate RAF is often the difference between:

• “These patients are too complex to manage under our budget”

…and “We’re paid enough to build a real chronic-care program.”

4.2 The bad: coding intensity and warped incentives

The same mechanism can distort behavior. Evidence from MedPAC, Health Affairs, and Annals of Internal Medicine shows that:

• MA plans and some provider groups have used chart reviews and home-visit programs to maximize documented diagnoses, generating tens of billions of dollars in extra payments.
• This “differential coding” is not always tied to true increases in morbidity and has triggered calls for model reforms and stricter auditing.
• RAF can create pressure to “find more codes,” even when the clinical value is questionable.
• Investigative journalism has illustrated the extreme end of this spectrum—e.g., one-hour home visits used primarily to harvest diagnoses that raise RAF and payments.

4.3 The ugly: under-coding and missed complexity

On the other side, many clinics—especially safety-net, rural, and busy primary care—face the opposite problem:

• Multimorbid, high-need patients
• Incomplete documentation and problem lists
• Limited coding support

That combination leads to RAF scores that are systematically too low given true patient complexity, which:

• Undercuts revenue,
• Masks the need for staffing and infrastructure, and
• Makes it look like your population is “simpler” than it is.

In other words: RAF can hurt you both when it’s abused and when it’s ignored.

5. So what does a “good” RAF strategy look like?

For a clinic, “good RAF” is not “the highest possible numbers.” It’s:

• Accurate, defensible scores that match your real clinical workload— and stand up to audit.

Practically, that means:

• Tight alignment between documentation and codes
• Every chronic diagnosis that drives RAF is clearly supported by the history, exam, labs/imaging, and plan.
• Annual re-assessment of chronic conditions
• Conditions that are still active are evaluated and addressed; resolved problems are clarified as historical.
• Focus on whole-patient complexity, not “favorite codes”
• Prioritize accurate capture of major drivers of cost: multimorbidity, advanced organ disease, serious mental illness, frailty.
• Using data and AI to surface real gaps, not to game the system

Clinical reasoning AI should:

• Scan longitudinal records for documented-but-uncoded conditions,
• Assemble evidence packets for clinician review,
• Help you avoid both under-coding and unsupported diagnoses.

It should not suggest conditions without chart evidence or bypass clinician judgment.

Understanding your RAF in context

Compare your panel’s RAF to:

• Clinical measures of multimorbidity,
• Hospitalization and ED rates,
• Social risk factors.

Large mismatches are a sign that your documentation and risk model are out of sync.

6. How to talk about RAF with your team (in plain language)

When you’re explaining RAF to clinicians, nurses, or operators, you might say:

• “RAF is how Medicare decides how sick our population looks, and how much money follows them.”
• “A RAF of 1.0 is average; our patients are often more complex than that. If we don’t document that complexity clearly, we’re under-resourced.”
• “The goal isn’t to ‘chase codes,’ it’s to tell an honest, defensible story about our patients’ chronic disease—and let the risk model do its job.”

When your team sees RAF this way, it becomes less about gaming the system and more about aligning dollars with reality so you can care for complex patients sustainably.

Chronic disease is now the default, not the exception. Yet millions of patients with diabetes, hypertension, and kidney disease are either undiagnosed or poorly documented in the record, even when they are regularly touching the health system.

In the U.S., an estimated 8.7 million adults have diabetes but don’t know it—about 23% of all adults with diabetes.

More than 11 million Americans are thought to have undiagnosed hypertension.

Chronic kidney disease (CKD) is often present in labs but missing from the problem list or billing codes, leading to under-treatment and lack of follow-up.

For clinicians, this is not news. What’s less discussed is why this happens and what kind of data infrastructure actually fixes it—without adding yet another inbox or alert.

This article breaks that down from a provider-operations point of view.

What does it mean to “miss” a chronic condition?

When we say conditions are “missed,” we’re usually talking about three related but distinct failures:

Undiagnosed

The patient clinically meets criteria (e.g., repeated A1c ≥ 6.5%, eGFR < 60, multiple elevated BPs), but no formal diagnosis has been made.

Example: A patient repeatedly has hypertensive-range blood pressures in urgent care, but “hypertension” is never added to the problem list.

Undocumented (or under-documented)

The condition is recognized by the clinician but not consistently documented in the problem list, assessment, or claims.

CKD is a classic case: in studies of primary care records, only a small minority of patients with lab-confirmed stage 3–4 CKD had CKD on the problem list, and those without it were less likely to receive guideline-recommended care.

Under-coded (or mis-coded)

The condition is present in free text or labs but coded non-specifically or not at all, which undermines quality measures, risk adjustment, and care coordination.

For value-based care and risk-bearing contracts, this is where missed conditions turn into revenue leakage and performance penalties.

All three ultimately come from the same root issue: our data and workflows were not designed for continuous, longitudinal detection of chronic disease.

Why providers miss chronic conditions

None of this is about individual competence. The evidence shows it’s a system problem made up of several layers.

1. Silent or subtle diseases

Many chronic conditions are asymptomatic or nonspecific early on:

• Hypertension: Often no symptoms until late complications; large proportions of adults with hypertension are unaware of it.
• Type 2 diabetes: Can progress for years with vague complaints like fatigue or blurry vision.
• CKD: Frequently discovered only through routine labs; even then, staging and documentation are often missed or incorrect.
• In busy primary care or SNF settings, it is easy for these “quiet” conditions to remain in the background.

2. Fragmented and noisy data

Clinicians rarely see a single, clean longitudinal story:

• Labs in one system, hospital notes in another, outside consults scanned as PDFs, and claims data invisible at the point of care.
• Important patterns (e.g., three elevated BPs across three sites, or progressive eGFR decline over 18 months) are statistical signals spread across multiple visits and settings, not obvious in a single encounter.

Research on diagnostic error repeatedly highlights information availability and integration as major contributors: critical results not seen, not communicated, or not synthesized in time.

3. Cognitive overload and time pressure

Primary care and LTC clinicians manage:

• 15–20 minute visits, often with multiple complaints
• Complex medication lists and social issues
• Performance pressures around access, throughput, and documentation

Studies of diagnostic error in primary care show that task overload, competing demands, and cognitive biases (anchoring, premature closure) are common drivers of missed or delayed diagnosis, even for common conditions like CHF, pneumonia, or renal failure.

When the visit is dominated by acute issues, chronic risk signals in the chart become background noise, not actionable insights.

4. EHR problem lists and coding gaps

Even when conditions are recognized, they’re often not reliably represented in the data:

• One CKD study found that low recognition in the EHR problem list was associated with lower use of CKD-specific processes of care (e.g., ACEi/ARB prescriptions, urine protein measurement).
• CMS has highlighted mis-recognition of CKD stage among primary care providers; in one test case, stage 4 CKD was correctly recognized only ~63% of the time.

In value-based models, this translates into:

• Inaccurate risk scores (HCC / RAF)
• Misaligned panel acuity
• Under-resourced care management for those who are actually high risk

If the EHR can’t “see” the condition, none of the downstream systems can either—quality, risk, CCM, or population health.

Why missed chronic conditions matter

1. Clinical outcomes

Undiagnosed or under-documented conditions mean:

• Delayed treatment (e.g., no statin or ACEi/ARB where indicated)
• Faster progression to complications (ESRD, MI, stroke, heart failure)
• More emergency visits and unplanned admissions
• Evidence around CKD is particularly stark: early recognition and interventions can significantly delay progression, yet under-recognition is common.

2. Quality and value-based performance

For organizations in risk-bearing arrangements:

• Conditions that are not captured, coded, and revisited annually may not be reflected in RAF or other risk models.
• Quality measures (HEDIS, Stars) often depend on having both a diagnosis and appropriate follow-up documented in the measurement year.

The result: sick populations that look “average” on paper—and contracts priced or evaluated as if that were true.

3. Operational burden

Ironically, poor data quality creates more work:

• Manual chart reviews to “hunt” for conditions ahead of audits or submissions
• Back-and-forth between coders, clinicians, and care managers
• Fire drills at the end of the year to close gaps that could have been surfaced earlier

Fixing chronic-condition visibility is therefore not just a clinical or financial issue; it’s a workflow and burnout issue.

What “better data” actually looks like

Better data is not just “more fields” or “more alerts.” For providers, better data means the right clinical story, at the right time, in the right workflow.

1. Longitudinal, integrated patient timelines

Instead of each visit being a separate island, high-functioning systems create:

• A single longitudinal view aggregating vitals, labs, medications, problem list, consults, and claims.

Simple, clinically meaningful signals:

• “3+ hypertensive readings in last 12 months without a hypertension diagnosis”
• “eGFR < 60 for ≥ 3 months without CKD staging”
• “Multiple oral steroids + wheeze + rescue inhaler use, no COPD/asthma diagnosis”

This is exactly what many national disease management programs do at scale—using integrated data to systematically detect high-risk or undiagnosed patients.

2. Computable phenotypes and registries

A computable phenotype is a machine-readable definition of a disease using structured data: lab thresholds, diagnosis codes, medications, time windows, etc. Examples:

• CKD phenotype defined by repeated eGFR and albuminuria values rather than codes alone.
• Diabetes phenotype defined by A1c, fasting glucose, and anti-diabetic medication use, not just ICD-10.

When you define these phenotypes and run them routinely across your data:

• You build live registries (e.g., CKD, diabetes, heart failure) that don’t depend solely on billing codes.
• You can compare coded vs. phenotype-positive cohorts to find patients who are “hiding in plain sight.”

3. Clinical decision support that respects workflow

The literature is clear: clinical decision support (CDS) can improve chronic disease care when it’s integrated thoughtfully into the EHR.

Effective CDS for chronic conditions typically:

• Surfaces fewer, higher-yield signals (e.g., “suspected CKD based on labs”) rather than a flood of alerts.
• Shows concise context: key labs, prior diagnoses, meds, and relevant guidelines.

Supports one-click actions:

• Add/update diagnosis
• Order confirmatory testing
• Refer to nephrology/endocrinology
• Enroll in CCM/TCM program

In other words, it turns data into decisions and decisions into orders, without forcing clinicians into a separate tool.

4. Closing the loop: from alerts to accountability

Better data systems also track:

• Whether the suggested action was accepted, modified, or dismissed, and why.
• Which conditions are consistently being overlooked, and in which clinics or subpopulations.
• Trends over time in undiagnosed vs. diagnosed and documented vs. phenotype-positive patients.

This “CDS exhaust data” can be analyzed to identify knowledge gaps, workflow bottlenecks, and training needs, not just individual errors.

A practical roadmap for provider organizations

You don’t need a greenfield tech stack to start fixing missed chronic conditions. Here’s a pragmatic sequence many organizations follow.

Step 1: Define your top 5 conditions and phenotypes

Start with conditions that:

• Drive the most cost and risk (e.g., diabetes, hypertension, CKD, heart failure, COPD).
• Are common sources of underdocumentation and quality gaps.

For each condition, define:

• The ICD-10 basis (what counts as “documented”).
• The phenotype basis (labs, vitals, meds, timelines).
• What “good management” looks like: key labs, meds, follow-up intervals.

Step 2: Build “hiding in plain sight” lists

Using your existing EHR + data warehouse (or a partner):

• Identify patients who meet the phenotype but lack a diagnosis code.
• Identify patients with a diagnosis but missing key elements of care (e.g., no A1c in 12 months, no ACEi/ARB in albuminuric CKD).

Prioritize:

• High-risk segments (older adults, multiple comorbidities).
• Value-based contract panels (ACO, Medicare Advantage, SNF/LTC aligned populations).

Step 3: Deliver insights into native workflows

Avoid standalone dashboards that nobody checks. Instead:

• Surface suspected conditions inside the EHR workflow (in-basket, chart banner, problem list assistant, or care management queue).
• Give clinicians short, traceable evidence: “eGFR 52, 49, 47 over 13 months; UACR 45 mg/g.”
• Offer pre-built order sets and documentation shortcuts so that acting on the signal takes seconds, not minutes.

Step 4: Engage care managers and pharmacists

Not every gap needs to be solved in a 15-minute visit:

• Create care management worklists for outreach: BP recheck visits, lab follow-ups, high-risk medication reviews.
• Allow non-physician team members (RNs, pharmacists, NPs, social workers) to advance the workup and education between physician visits.
• This team-based model is essential for closing chronic care gaps at scale.

Step 5: Measure and iterate

Track:

• Reduction in phenotype-positive but uncoded patients per condition.
• Improvements in time to diagnosis, and in adherence to key care processes (labs, meds, follow-ups).
• Impact on utilization and contract performance, recognizing that clinical outcomes may lag.

Use these data to refine phenotypes, CDS rules, workflows, and training.

Where AI fits—and where it doesn’t

AI and advanced analytics can accelerate this entire process, but they are not magic.

Useful applications include:

• NLP on free text to surface unstructured evidence of chronic conditions that never made it into discrete codes.
• Risk models that prioritize which phenotype-positive patients are most likely to benefit from outreach.
• Automated evidence summaries that explain to clinicians why a condition is being flagged, in a way that is auditable and clinically transparent.

However:

• AI models must be trained on representative, high-quality data and regularly validated to avoid bias.
• The real value is not just predicting “who is sick,” but making it easy for teams to do the right thing once those patients are identified.

Takeaway: Missing conditions are a data and workflow problem, not a character flaw

Providers are not “missing” chronic disease because they don’t care or don’t know. They’re doing the best they can in a system where:

• The signals are diffuse and spread across time and systems.
• The tools are often built for billing, not for longitudinal reasoning.
• The workload and cognitive load are already beyond reasonable.

The good news: this is fixable. With better data integration, computable phenotypes, thoughtful CDS, and team-based workflows, organizations can:

• Find the patients who are “hiding in plain sight.”
• Treat conditions earlier and more consistently.
• Align documentation, risk, and resources with the true burden of disease in their population.

If you’re designing your provider strategy around chronic disease management, the most important question is no longer “Do we have data?” but “Can our system recognize and act on the chronic disease signals we already have?”

That’s the gap worth closing.

Electronic health records (EHRs)—often still called EMRs—were supposed to give us a complete, computable picture of each patient.

In reality, when you look closely at problem lists, diagnosis codes, and multimorbidity measures, a tough pattern emerges:

• The typical EHR view misses a huge chunk of chronic-condition complexity — easily 30–50%, depending on how you measure it.

Studies show that:

• Problem list completeness for common chronic diseases varies widely, with only ~60–80% of known conditions actually present in the problem list across institutions.
• For some high-impact diseases like stage 3–4 chronic kidney disease, only 16% of patients with lab-defined CKD had CKD documented on the problem list.
• Reviews of multimorbidity measurement note that EHR systems underestimate the burden of multiple chronic conditions because of missing or out-of-date data.
• Validation work on ICD-10 administrative data finds under-coding of co-existing conditions, especially asymptomatic ones, with sensitivities as low as ~40% for some comorbidities.

Put together, it’s reasonable for a health system to assume that on the order of ~40% of chronic-condition “signal” is invisible if you rely only on problem lists and billing codes in the EHR.

This article unpacks why that happens, what “40%” practically means, and what providers should demand from their data and AI tools going forward.

1. EHRs capture encounters; patients live in trajectories

Most EHR data models were built around encounters:

• Each visit gets diagnoses, meds, orders.
• Problem lists accumulate items over time.
• Claims data records what was billed.

But chronic disease complexity lives in trajectories:

• Kidney function drifting down over five years
• HF symptoms waxing and waning with changing meds
• COPD, obesity, depression, and diabetes interacting in a feedback loop

Multimorbidity research using large EHR datasets repeatedly finds that:

• Chronic conditions cluster in complex ways, and
• These patterns are hard to see if you only look at one code list at a time.

EHRs absolutely store the raw ingredients of these trajectories—labs, vitals, meds, notes—but the structured “front window” (problem list + last set of visit diagnoses) is often a blurry, partial summary.

That’s the first way EHRs miss complexity: they weren’t designed to reason over whole patient timelines.

2. Problem lists: better than nothing, far from complete

Problem lists are supposed to be the single source of truth for a patient’s major diagnoses.

The reality is more complicated.

2.1 What the data say about problem list completeness

Across multiple studies:

• A multi-site analysis of EHRs at 10 health systems found problem list completeness for chronic diseases ranged from 60.2% to 99.4%, with an average around 78%.
• In outpatient settings, common chronic diseases showed both missing diagnoses and duplications on problem lists, with completeness and quality varying by condition.
• For chronic kidney disease, only 16% of patients with lab-defined stage 3–4 CKD had CKD recorded on the problem list; multiple studies have described CKD as under-recognized and under-documented in EHRs.

A 2022 study of patients with heart failure in the emergency department found that the EHR problem list captured multimorbidity with moderate-to-good accuracy, but still misclassified an average of two comorbidity domains per patient when compared with a reference list based on full records.

In other words: problem lists are directionally useful, but they systematically lose detail and drop conditions, especially:

• “Background” conditions that feel less urgent (e.g., CKD, anemia)
• Historical conditions that no one is sure whether to keep
• Issues owned by other specialties

2.2 Why problem lists fall behind

Qualitative work on problem list use points to familiar issues:

• Unclear ownership (“Who maintains this? Primary care? Specialist? Hospitalist?”)
• Time pressure during visits
• Fear of cluttering the list, making it unusable
• Lack of feedback or incentive tied directly to maintaining a high-quality list

When no one “owns” the problem list as a shared clinical asset, it decays. Chronic-condition complexity ends up scattered across:

• Old notes
• Scanned hospital records
• Lab trends
• Imaging reports

…while the official snapshot looks deceptively simple.

3. Codes and claims: the second blindness

If problem lists understate complexity, maybe claims data makes up for it?

Not quite.

3.1 Under-coding of comorbidities in administrative data

Recent evaluations of ICD-10 administrative data compared with chart review show:

• Sensitivity for individual comorbidities ranging from ~40–90%, with some conditions consistently under-coded.
• Asymptomatic or lower-priority conditions (e.g., some chronic diseases that weren’t the main reason for admission) are particularly likely to be missed.
• Under-coding of co-existing conditions has persisted or worsened over time in some settings, even as coding systems and financial incentives evolved.

A 2017 analysis of co-existing conditions in administrative data explicitly highlights under-coding as a major issue, especially for conditions that don’t drive immediate billing or are overshadowed by the primary diagnosis.

3.2 Why claims data underestimates multimorbidity

Claims are optimized for payment, not clinical nuance:

• Only a handful of diagnoses are coded per encounter.
• Coders are under time pressure and prioritize principal diagnoses and high-impact comorbidities.
• Historical context (e.g., “HF with preserved EF vs reduced EF,” CKD staging, diabetic nephropathy vs “diabetes + CKD”) may never be fully expressed in codes.

Multimorbidity researchers emphasize that large datasets based only on codes systematically underestimate the true number and complexity of chronic conditions in patient populations.

So if you stack:

• Imperfect problem lists, and
• Under-coded claims

…you can easily end up blind to 30–50% of the “long tail” of chronic disease complexity in your panels.

4. What “40% of complexity” actually looks like

“40%” is not a universal constant. It’s a pragmatic shorthand for what multiple lines of evidence are telling us:

• Problem list completeness across systems averaging ~78% (so ~22% of conditions missing), with much worse documentation for certain diseases (e.g., CKD).
• Multimorbidity reviews and methods papers warning that EHR-based counts underestimate disease burden because of missing or outdated entries.
• Administrative data validation studies finding substantial under-coding of co-existing conditions, especially those that are asymptomatic or less central to the index encounter.
• Disease-specific work (e.g., CKD, HF) showing dramatic gaps between objective signal in labs/imaging and diagnoses formally recorded in the EHR.

If you’re a health system that:

• Relies mainly on problem lists and diagnosis codes to understand your population, and
• Hasn’t invested in longitudinal, phenotype-based analytics across labs, meds, and notes

…it is entirely realistic that 30–50% of your patients’ chronic-condition interactions, severities, and etiologies are effectively invisible to your dashboards and AI tools.

Concretely, that 40% might include:

• CKD that is evident from eGFR trends but never labeled
• Diabetic neuropathy, retinopathy, or nephropathy documented narratively but not coded as such
• HFpEF vs HFrEF distinctions buried in echo reports
• Depression, frailty, or cognitive impairment that shows up in free text and functional assessments but not in structured codes
• The patient is complex; your EHR summary says “HTN, DM2, OA.”

5. Why EHRs miss complexity: five structural reasons

5.1 Fragmented data within the EHR

A 2023 study of chronic condition identification in Swiss primary care EHRs showed that no single EHR component (problem list, diagnoses, prescriptions, or notes) was sufficient to capture prevalence; accurate estimation required integrating multiple components.

In real systems, chronic disease evidence is split across:

• Problem lists
• Encounter diagnoses
• Labs and imaging
• Orders and medications
• Free-text notes and scanned documents

Most analytics pipelines still pull from one or two of these, not all.

5.2 Asymptomatic and “background” conditions

Conditions like early CKD, anemia, or mild heart failure often don’t drive the encounter agenda. Research repeatedly shows:

• CKD is under-recognized and under-documented despite objective lab evidence.
• Co-existing conditions without acute symptoms are especially likely to be missing from administrative data.

If it’s not the “chief complaint” or the main driver of billing, it often doesn’t show up in the structured summary.

5.3 Time and incentives

Problem list and code quality improve when organizations:

• Provide gap reports and feedback to clinicians
• Link problem lists to billing and quality incentives
• Make problem-oriented charting an integral part of workflow

But in many settings, clinicians receive:

• Little direct feedback on problem list completeness, and
• Stronger signals about throughput and relative value units (RVUs)

So complexity remains in the chart, not in the fields your analytics see.

5.4 Legacy data and migration issues

Longitudinal studies of chronic disease in EHRs note that:

• Conditions diagnosed before a certain date (e.g., before HIE go-live) may not be present in the EHR dataset at all.
• Migrations and system changes can strand older diagnoses in “historic” or unstructured formats.

So your EHR may only truly reflect the last few years of complexity, even for 20-year disease trajectories.

5.5 The combinatorial explosion of multimorbidity

As multimorbidity network studies point out, once patients have 5–10 chronic conditions, the number of possible combinations explodes. Many combinations appear in only a handful of patients.

Most EHR interfaces and registries simply aren’t designed to represent these complex patterns in ways that are easy to see or query, so signals get flattened.

6. Why this matters: care, contracts, and AI

6.1 Clinical care and population health

If your EHR misses 40% of chronic condition complexity, you risk:

• Under-recognizing high-risk patients who actually have dense multimorbidity.
• Missing opportunities for disease management (e.g., CKD or HF clinics) because patients don’t meet coded criteria.
• Designing population health programs on an incomplete denominator, which dilutes impact.

Trials of EHR-based population health management for CKD emphasize the need to identify high-risk patients using labs and care gaps, not just diagnoses on problem lists.

6.2 Risk adjustment and value-based contracts

Under-captured complexity has direct financial implications:

• Risk scores that don’t reflect actual multimorbidity can lead to under-payment for truly complex patients.
• Conversely, naive efforts to “code harder” on a shaky data foundation risk over-coding or mis-coding, which auditors are increasingly scrutinizing.

If you’re negotiating or operating under risk-bearing contracts while your EHR only sees 60% of the true chronic disease landscape, you are making strategic decisions with partial information.

6.3 AI and clinical decision support

Most first-generation AI tools (including some LLM-based solutions) are trained or evaluated on the structured fields that underrepresent complexity:

• Problem lists
• Claim codes
• Simple comorbidity counts

Reviews of multimorbidity and EHR-based AI warn that these limitations can lead to biased models that underestimate risk, especially for patients whose complexity lives in under-documented or fragmented data.

If we don’t fix the data layer, “AI for chronic disease management” just learns a sharper version of the same blind spots.

7. What providers should demand from their data (and AI)

Closing the 40% gap is not about telling clinicians to “code more.” It’s about building infrastructure that can see reality.

At minimum, providers should expect that any serious population health, risk, or AI solution:

• Builds patient-level phenotypes, not just counts of codes
• CKD defined by eGFR trends + albuminuria + diagnoses, not codes alone.
• Integrates multiple EHR components
• Problem lists + encounter diagnoses + labs + meds + notes, as demonstrated in chronic disease prevalence work using EHR components.
• Audits itself against chart review or enriched reference standards
• Periodically comparing phenotype outputs with clinician-reviewed samples, as ICD-10 validation and multimorbidity studies do.
• Surfaces “silent” complexity
• Explicitly highlights patients whose labs, meds, or notes suggest additional chronic conditions or greater severity than codes reflect.
• Keeps longitudinal context front and center
• Designs interfaces and AI explanations around trajectories and timelines, not just static snapshots.
• Respects clinical workflow and ownership
• Aligns with how clinicians think about the problem list—as a shared, longitudinal clinical resource—and gives them tools (not nagging) to keep it meaningful.

The goal isn’t a perfect EHR; it’s an accurate enough longitudinal model of each patient to support good care, fair payment, and trustworthy AI.