Chapter 36: Coding for Value: HCC Risk Adjustment, Quality Measures, and Value-Based Reimbursement

"The claim describes what we did. The record is supposed to describe who she is. On this file those are two different documents, and only one of them gets read next year." — constructed

Overview

Every payment in this book so far has been triggered by an event. A patient was seen, a code went on a line, the line had an allowed amount, and somebody paid it. Chapter 23 built that arithmetic for the office. Chapter 33 replaced the line with the stay and the fee schedule with a classification. Chapter 34 did it again for the hospital outpatient department. In all three, the money moved because something happened.

This chapter is about money that moves because someone is something. Under risk-adjusted payment, a health plan is paid a fixed amount each month for each member it covers, and that amount is adjusted up or down by how sick that member is expected to be for the coming year. The evidence of how sick she is comes from one place: the diagnosis codes that were documented and reported for her during the previous year. No claim in that system pays for a diagnosis code. The diagnosis codes set the size of the pot before any claim is written.

That is a genuinely different job, and it changes what "correct" means. Chapter 9 §9.7 answered the question this book has carried since Chapter 4 — is E11.9 correct for Account 10-4471's March 14 encounter? Yes, and for a specific reason: Section IV reports conditions that were addressed at the encounter, and that assessment did not address the kidney disease on the problem list. Nothing in this chapter disturbs that answer. This chapter asks the second question, which Chapter 9 explicitly handed here: is E11.9 sufficient?

No. The March 14 code is not wrong; it is incomplete as a description of the patient, and the difference between those two statements is worth money. Getting from the first statement to the second is the work of eleven sections.

You will also meet the other half of "value": the quality measures that decide whether an organization keeps its savings, the Category II codes that report them, and the contracts — shared-savings, upside-only, two-sided, capitated — that put both halves together. And you will meet the reason this is one of the most heavily enforced areas in American healthcare, because a payment system that pays for documented diagnoses creates a permanent incentive to document diagnoses, and everyone involved has known that from the beginning.

In this chapter, you will learn to:

• Explain why a fixed per-member payment has to be risk-adjusted, and what happens when it isn't
• Trace a diagnosis into a condition category and out to a coefficient
• Build and check a risk adjustment factor
• Apply hierarchies, trumping, and interactions correctly
• Recapture chronic conditions under the annual reset — and know what does not recapture them
• Apply MEAT to an assessment and defend the result
• Recognize where chart review crosses the line, and how each failure is caught
• Report process and outcome measures with the right codes
• Read a value-based contract for the terms that decide it

36.1 Paying for a population instead of a visit

Start with the decision on the desk — except this time it is not on your desk yet. The practice manager comes back from a meeting with a contract in hand. A Medicare Advantage plan is offering Northgate Family Medicine an arrangement: the practice keeps billing its usual claims, and on top of that it shares in the difference between what the plan expected its attributed patients to cost and what they actually cost. There is a quality component. There is language about risk scores. Nobody in the building has read a contract like it.

The manager's question to the coder is the right one and almost nobody asks it early enough: "does this change what you do?"

The answer is yes, and the change is not subtle. It is not that the codes are different — they are the same ICD-10-CM codes from the same book. It is that a second reader has appeared. Until now, a diagnosis code on Account 10-4471 had exactly one job: to justify a service to an adjudication system that would pay or not pay a line. Now the same code has a second job, performed a year later, in a building the practice will never see, where nobody is looking at the claim at all.

TWO PAYMENT LOGICS, SAME CODE SET

  FEE-FOR-SERVICE                        RISK-ADJUSTED / VALUE-BASED
  ───────────────                        ───────────────────────────
  a visit happens                        a member is enrolled
        │                                        │
        ▼                                        ▼
  codes describe the visit               codes describe the MEMBER
        │                                        │
        ▼                                        ▼
  a claim is adjudicated                 a year's diagnoses are collected
        │                                        │
        ▼                                        ▼
  a LINE is paid or denied               a RISK SCORE is computed
        │                                        │
        ▼                                        ▼
  money follows the SERVICE              money follows the POPULATION,
                                         monthly, before any service

  UNIT OF PAYMENT:  the line             UNIT OF PAYMENT:  the member-month
  TIMING:           weeks                TIMING:           the FOLLOWING year
  A MISSING CODE:   a denial you see     A MISSING CODE:   nothing you ever see

Read the bottom-right cell slowly, because it is the whole reason this chapter is hard to teach. Every error this book has taught you to catch announced itself. A rejection came back with a reason. A denial came back with a CARC. An underpayment was silent, and Chapter 28 §28.8 had to build a method to find it — but even that method had a number to compare against. Under risk adjustment, an incomplete record produces no artifact at all. There is no denial, no variance, no work queue entry, no report with a line on it. The claim pays exactly as expected. The consequence arrives eleven months later as a payment that is smaller than it should have been, in a system nobody in the practice can see, attached to no particular patient.

The scope is not niche. Medicare Advantage — Chapter 3 §3.6's Part C — now covers a very large share of Medicare beneficiaries, and its payment mechanism is the one this chapter describes. The Affordable Care Act's individual and small-group markets run their own risk-adjustment program. Medicaid managed care programs risk-adjust their capitation rates, by a variety of state-selected methods. Accountable care organizations and commercial value-based contracts risk-adjust their benchmarks. A coder who works only fee-for-service claims today has a strong chance of working in one of these arrangements within a few years, and the transition tends to happen the way it happened in the paragraph above: a manager comes back from a meeting.

📞 On the Phone

The practice manager, after the meeting. "They kept saying 'RAF.' What is a RAF?"

The coder, who has done the reading: "Risk adjustment factor. It's a number attached to each of your patients that says how expensive the plan expects that patient to be next year, relative to an average member. It's built mostly out of two things — who the patient is, meaning age and sex and a few eligibility facts, and what conditions were documented and reported for them during the year. A member at 1.0 is average. Higher means sicker and the plan gets paid more for them; lower means healthier and it gets paid less."

The manager: "So we get paid more if we code more?"

And here is the answer that decides whether this practice does this well or badly: "No. Two separate things are wrong with that sentence. First, the extra money goes to the plan, not to us — what we get is whatever the contract says we get, and I'd like to read that part before anybody celebrates. Second, and this is the part I need you to say out loud in front of the physicians: the job is not to code more. It's to code the patient completely and accurately, which means we will sometimes report a condition that adds nothing and sometimes stop reporting one that was never supported. If anyone in this building starts describing this as a revenue project, we will end up in Chapter 36's second case study rather than its first."

What that models: the failure mode is set by the first sentence anybody says about the program. "Capture what we are already doing" and "find more codes" describe the same activity and produce different organizations.

The theme this chapter carries hardest is the second one. You code from the chart, but you get paid by the contract. In this chapter the two halves have separated further than anywhere else in the book: the code is a statement about a patient, made in a chart, and the payment is a term in an agreement between a plan and a government program that the coder will never read. Both halves have to be right, and they are right for different reasons.

36.2 What risk adjustment is trying to fix

Risk adjustment is easiest to understand as the answer to a failure, and this book has already told the story of the failure. Chapter 2's second case study walked through the collapse of 1990s capitation, and its first named cause was this:

The rate was set without adequate risk adjustment. A group whose attributed panel happened to include a disproportionate share of members with serious chronic illness received the same amount per member per month as a group with a healthy panel, and delivered several times the care.

Hold the mechanism, not the history. Any payer that pays a fixed amount per member has created a problem it did not have before. Under fee-for-service, a sicker patient generates more services and therefore more revenue; the payer absorbs the cost and the arithmetic takes care of itself. Under a fixed per-member payment — PMPM, per member per month — a sicker patient generates the same revenue and more cost. The organization holding that contract now has a financial interest in the health of the people who enroll with it, and the cheapest way to serve that interest is not to make anyone healthier. It is to enroll healthier people.

That behavior has names in the literature, none of them flattering — selection, cherry-picking, avoiding the complicated patient, being slow to return the call from the nursing home. And it is worth being blunt about the stakes, because this is the rare place where a payment formula is a patient-protection mechanism: a payment system that does not adjust for risk pays organizations to avoid sick people. Risk adjustment exists so that the member with heart failure, diabetes, and kidney disease is worth roughly what she costs, and nobody has a financial reason to wish her onto a competitor.

Risk adjustment is the practice of modifying a fixed per-member payment according to the expected relative cost of that member, using documented characteristics — demographics and diagnosed conditions — as the predictors.

Three properties of that definition are load-bearing and are routinely misunderstood.

It predicts cost, not sickness. The models are fit statistically against actual spending. A condition earns a coefficient because people who have it cost more, on average, in the following period — not because it is medically grave. This is why some serious, disabling conditions carry no weight at all and some unglamorous ones carry a lot. A coder who reasons "this patient is very sick, so her score should be high" will be surprised regularly.

It is relative, not absolute. A score of 1.0 means "an average member." The whole model is normalized so the population sits near 1.0, and the payment is a benchmark multiplied by the score. A member at 1.6 is not expensive in dollars; she is expected to cost about 60% more than average.

It is retrospective in evidence and prospective in use. The diagnoses come from last year. The payment is for next year. That single sentence generates §36.6's annual reset, most of §36.7's documentation discipline, and every enforcement action in §36.8.

The tension the design cannot remove

A model that pays according to documented diagnoses pays for documentation. That is not a flaw somebody discovered later; it is inherent in the design, it was foreseen, and the entire audit apparatus in §36.8 exists because of it. Chapter 33's Case Study 2 told the same story one payment system earlier: when a classification prices a stay, the classification becomes an object of attention, and "DRG creep" entered the vocabulary before the system had finished launching. The identical drama is running now, with diagnoses instead of severity tiers, and with a much larger number of participants.

The honest statement of the coder's position inside that tension is short. Under-reporting and over-reporting are the same error with opposite signs, and neither is the safe choice. Chapter 5 §5.8 established that downcoding is not a defense; here it also describes a population as healthier than it is, which distorts the payment, the quality data computed from it, and the practice's ability to answer basic questions about its own patients. Chapter 11's Case Study 2 showed exactly that outcome: a practice that could not say how many of its heart failure patients had systolic versus diastolic disease, because the information was in every chart and in none of the codes.

⚠️ Where Claims Die

The code that changes nothing.

Chapter 6 §6.9 named risk adjustment as the standing counterexample to financial accuracy as an audit metric — the measure that treats an error which happens not to change payment as no error at all. Here is that abstraction with the money attached.

A coder assigns an unspecified code where a specified one was supported. The claim pays. The internal audit scores it as financially accurate. The physician never hears about it. Every control the practice operates reports success. And the patient has been described, for an entire payment year, as someone she is not.

The pattern has a signature you can look for without any special software: a practice whose diagnosis distribution is dominated by the unspecified option. Pull your own top twenty diagnosis codes by volume. If the unspecified variant leads the list in categories where the chart routinely contains the specifics — heart failure type, chronic kidney disease stage, diabetes complications — you have found it, and you found it without a payer telling you.

What the disciplined professional does instead: treat "the claim paid" as evidence about the claim and nothing else. It is not evidence that the code was right, and under risk adjustment it is not evidence that the money was right either.

36.3 Hierarchical condition categories

The machinery that turns a diagnosis into a payment predictor is a grouping model, and the most important one in the United States is the CMS-HCC model — the Centers for Medicare & Medicaid Services' hierarchical condition category model, used to risk-adjust payments to Medicare Advantage organizations.

A hierarchical condition category (HCC) is a group of clinically related diagnoses that predict similar incremental cost, treated by the payment model as a single unit and carrying a single coefficient. "Hierarchical" refers to the ranking within a related family of categories, so that a member with several severities of the same disease is counted once, at the most severe.

The pipeline is short and worth drawing, because most confusion about HCCs comes from collapsing two of these steps into one:

FROM A CHART TO A COEFFICIENT

   THE RECORD          a documented, dated, signed encounter note
        │
        ▼
   ICD-10-CM CODE      assigned by the ordinary rules of Chapters 7-12
        │                  (tens of thousands of codes exist)
        ▼
   MAPPED?             the model's crosswalk asks: does this code map
        │              into a condition category at all?
        │
        ├── NO  ──────► nothing. The great majority of ICD-10-CM codes
        │               map to no category in any risk model.
        ▼
   CONDITION           a diagnostic group, then a condition category
   CATEGORY
        │
        ▼
   HIERARCHY           within a family, the most severe category
   APPLIED             survives; the rest are zeroed (36.5)
        │
        ▼
   COEFFICIENT         a decimal weight, published with the model,
                       revised on a schedule                [schematic]

Four facts about that pipeline decide most real questions.

Most codes map to nothing. This is the first thing to internalize and the last thing coders believe. A risk model is not an inventory of illness; it is a cost-prediction instrument, and it selects the conditions that predict cost with enough reliability to be worth including. Symptom codes generally do not map. Most musculoskeletal conditions do not map. Uncomplicated hypertension does not map. A chart full of accurate codes can produce a score built from two of them, and that is normal rather than a sign that something was missed.

"Unspecified" is where the loss lives. Where the model does distinguish severities, it does so through the code's specificity. E11.9 says type 2 diabetes, no complications. E11.22 says type 2 diabetes with diabetic chronic kidney disease. Those two statements go to different categories in every version of the model, and the with-complications category carries the larger coefficient. The code is not "more valuable." It is more true, and the model is built to price the truth.

There is more than one model, and they disagree. The CMS-HCC model risk-adjusts Medicare Advantage. A separate CMS-HCC ESRD model covers beneficiaries with end stage renal disease. The RxHCC model risk-adjusts Part D drug payments and selects for conditions that predict drug spending, so its categories are not the same ones. The HHS-HCC model risk-adjusts the Affordable Care Act's individual and small-group markets and covers all ages, including infants and pregnancy, which the Medicare model has no reason to. State Medicaid programs use their own, frequently a diagnostic classification developed for Medicaid populations. A condition that carries weight in one model may carry none in another, for entirely legitimate reasons, and a coder who learned "the HCCs" in one setting has learned one model.

And the model itself changes. CMS has run successive versions of the CMS-HCC model — the categories are renumbered, the mapping is revised, conditions are added and removed, and the coefficients are refit. Transitions have been phased in as a blend of two versions across payment years, which means that in a transition year two versions are simultaneously in force with different weights, and the model applicable to a given member depends on the payment year and the program.

The update cycle, once, for this chapter. ICD-10-CM changes every October 1. CPT changes every January 1. HCPCS Level II changes quarterly. And — the addition this chapter makes — risk adjustment models, their category definitions, their code mappings, and every coefficient are revised annually, and the model version in force differs by contract year and by program. Code from the current year's book or encoder. Compute from the model version your contract actually runs on, obtained from the plan or from CMS, and never from a textbook. This one is worth more than the others, because a stale ICD-10-CM code will be rejected by an edit and a stale coefficient will quietly produce a wrong number that nothing checks.

That is why this book prints no HCC numbers and no coefficients as fact. Every figure in §36.4 and §36.11 is a constructed teaching value, labeled where it appears. Learn the structure — demographic component plus disease coefficients, summed, applied to a benchmark — and look up the values. A coder who has memorized a category number has memorized something with an expiration date; a coder who can explain how a chart becomes a coefficient can work in any model in the country.

🔢 Code It

The documentation, from an established-patient office note (constructed teaching example): "Assessment/Plan: 1. Type 2 diabetes mellitus, poorly controlled, A1c 9.4. Increase metformin; add basal insulin, start 10 units at bedtime; diabetes educator referral. 2. Chronic kidney disease, stage 4, secondary to diabetes — eGFR 24, stable from last visit. Continue avoidance of nephrotoxics; nephrology follow-up scheduled. 3. Peripheral neuropathy, diabetic — gabapentin continued, foot exam today, no ulceration."

The path. Three conditions, each addressed with a status and a plan. The diabetes is documented with two named chronic complications; the classification links diabetes to chronic kidney disease and to neuropathy under the "with" convention of Chapter 9 §9.7, and here the provider has also stated the linkage explicitly for the kidney disease, so no presumption is even required. Report the diabetes with the chronic kidney disease manifestation, the diabetes with neuropathy manifestation, and the stage-4 code from the N18.- category — verifying each in the Tabular and honoring its instructional notes about sequencing and additional codes. Then, for risk purposes, notice what happened: several codes, and after the model's hierarchy runs, one diabetes category — the most severe — plus a kidney category.

The plausible wrong answer, and it is common: E11.9 plus a stage-4 kidney code, on the reasoning that "the diabetes is the diabetes and the kidney disease is coded separately." That claim would pay identically. It describes a patient with uncomplicated diabetes who separately happens to have failing kidneys, which is not what the note says, and it sends the model into the wrong diabetes category. The near-miss is not a payment error on this claim. It is a false statement about a person, made in a permanent record, that will be read by a payment system next year.

36.4 The risk adjustment factor and how it is built

The risk adjustment factor (RAF) — also called the risk score — is one number per member per payment year. It is a sum, and the addition is the least interesting part of it; what matters is knowing what the addends are and where each comes from.

HOW A RISK SCORE IS BUILT

  DEMOGRAPHIC COMPONENT        who the member is, from enrollment data
    age band x sex               the coder contributes nothing here and
    Medicaid / dual status       cannot change it
    institutional vs community
    originally-disabled status
                                        +
  DISEASE COMPONENTS           what was documented and reported LAST year
    one coefficient per            THIS is the part the chart controls
    condition category that
    survives the hierarchies
                                        +
  INTERACTIONS                 modeled combinations that cost more
    disease x disease              together than separately
    disease x status
                                        =
  ------------------------------------------------------------------
  RAF (the raw risk score)

  and then, at the PLAN level and not the member level:
    x normalization factor       keeps the average near 1.0 over time
    x coding intensity adjust.   a statutory reduction applied to MA
    (+/-) other program factors  scores

Two things about that diagram surprise people.

The first is how much of the score the coder does not touch. The demographic component comes from enrollment records — age, sex, whether the member is also eligible for Medicaid, whether she lives in an institution, whether she originally qualified for Medicare through disability. For many comparatively healthy members this component is most of the score. Nothing in a chart changes it.

The second is the bottom block. A member's raw score is not what the plan is paid on. Scores are normalized so that the average member stays near 1.0 as the population and the model change, and Medicare Advantage scores are additionally reduced by a statutory coding intensity adjustment — because Congress concluded that diagnoses are coded more completely in Medicare Advantage than in fee-for-service Medicare and directed that the difference be discounted. A payment model with a built-in haircut for coding intensity is a payment model that has already anticipated §36.8, and saying so out loud is a useful antidote to anyone in a practice who describes better coding as free money.

🧮 Run the Numbers

One member, one year. Take a record this book already owns: Account 22-8891, the four-day admission from Chapter 33 — a Medicare beneficiary, 71, whose coded record carries J44.1 (COPD with acute exacerbation), J96.01 (acute respiratory failure with hypoxia), I50.32 (chronic diastolic heart failure), E11.22 (type 2 diabetes with diabetic chronic kidney disease), N18.31 (chronic kidney disease, stage 3a), Z79.4 (long-term insulin use), and L89.153 (stage 3 sacral pressure ulcer, POA = N).

The lens: suppose this beneficiary were enrolled in a Medicare Advantage plan rather than in fee-for-service Medicare. (The Chapter 33 file itself is unchanged; nothing here alters it.)

[constructed illustrative coefficients — verify current coefficients and the model version in force at CMS]

```text DEMOGRAPHIC COMPONENT female, 70-74, community, non-dual, aged entitlement 0.346

DISEASE COMPONENTS (one per category surviving the hierarchies) chronic obstructive pulmonary disease J44.1 0.328 diabetes with chronic complications E11.22 0.302 chronic kidney disease, moderate stage N18.31 0.127 congestive heart failure I50.32 0.331

INTERACTION heart failure + diabetes 0.121

---

RAF 1.555 ```

Check: 0.346 + 0.328 = 0.674 · + 0.302 = 0.976 · + 0.127 = 1.103 · + 0.331 = 1.434 · + 0.121 = 1.555 ✓

Apply it. At a constructed county benchmark of \$1,000.00 PMPM [constructed]:

text 1.555 x $1,000.00 = $1,555.00 per member per month $1,555.00 x 12 = $18,660.00 for the payment year

Check: 1,555.00 × 12 = 18,660.00 ✓

Interpretation. This member is expected to cost about 55% more than an average member, and the plan is funded accordingly before it knows anything about what will actually happen to her. If she has a quiet year, the plan keeps the difference. If she is admitted three times, it absorbs the loss. That is the same prospective bargain Chapter 33 §33.1 described for the hospital, moved up one level and stretched from one stay to one year.

And notice what is not in the sum. Z79.4 is a status code: clinically load-bearing, and it earns no disease coefficient in this model. L89.153 carried POA = N and was therefore denied its severity credit by the DRG grouper — and the risk model does not read POA at all, because it is asking a different question about a different year. The same record, read by two machines, uses two different subsets of it. J96.01 raises a third point, taken up in §36.6: acute conditions can map, they count for the year in which they occurred, and they must not be carried into the next year unless they recur.

One more structural note, because coders ask it immediately. A member's score is built from diagnoses collected across all her acceptable encounters in the collection year — the hospital admission, the specialist visits, the primary care visits — not from one chart and not from one practice. Northgate contributes the diagnoses documented at Northgate. If the nephrologist documents and reports the kidney disease and the primary care practice never does, the condition is still on the member's record. This is why "we lost that condition" is usually the wrong sentence and "nobody in the member's care reported it that year" is usually the right one — and why §36.8's chart review, at its most legitimate, is an exercise in finding conditions that were documented somewhere and reported nowhere.

36.5 Hierarchies, trumping, and interactions

The word hierarchical is the H in HCC, and it does one specific job: it stops the same disease from being counted twice.

A HIERARCHY, SCHEMATICALLY
[structure only - the category names, numbering, and coefficients in any
 real model are a lookup in that model's published tables]

   THE DIABETES FAMILY, most severe at the top

   +--------------------------------------------------+
   |  diabetes with SEVERE ACUTE complications        |
   +--------------------------------------------------+
                    | TRUMPS everything below
                    v
   +--------------------------------------------------+
   |  diabetes with CHRONIC complications    E11.22   |
   +--------------------------------------------------+
                    | TRUMPS
                    v
   +--------------------------------------------------+
   |  diabetes WITHOUT complications         E11.9    |
   +--------------------------------------------------+

   REPORT ALL THREE  ->  the model counts ONE, the most severe
   REPORT ONLY E11.9 ->  the model counts the LOWEST

   The hierarchy protects the payer against double-counting.
   It does not protect the record against under-description.

Trumping is the operation: within a hierarchy, the highest-ranked category the member's diagnoses reach is retained, and every lower category in that same family is set to zero for the year.

Four consequences, and the fourth is the one that changes behavior.

You cannot inflate a score by stacking severities. A record reporting mild, moderate, and severe versions of the same disease earns the coefficient for the severe one and nothing extra. Coders coming from the fee-for-service world sometimes assume the opposite, by analogy to line items. There is no analogy; this is a different arithmetic.

Hierarchies operate within a family, not across families. Diabetes and heart failure are separate families and both count. Chronic kidney disease has its own hierarchy by stage. A member with four unrelated serious conditions accumulates four coefficients. This is also why §36.4's worked example sums four disease terms without any of them cancelling.

Reporting only the least severe version is not conservative — it is inaccurate, and the hierarchy guarantees it costs. If the record supports diabetes with chronic complications and the claim carries E11.9, the hierarchy has nothing to work with. It cannot promote a category that was never reached. The asymmetry is deliberate: the model is designed so that over-reporting gains you nothing and under-reporting costs you everything. That is a genuinely elegant piece of design and it deserves to be taught as such, because it converts "code accurately" from a moral instruction into the obviously correct strategy.

And trumping means a deleted code can matter. If a review determines that a severe category was reported without support and it is removed, the lower category the member's other diagnoses support becomes live again — assuming those diagnoses were reported. A program that only ever adds has no way to discover this, which is §36.8's subject.

Interactions

Some combinations of conditions cost more together than the sum of their parts, and the models say so explicitly with interaction terms. Heart failure with diabetes is the classic example and it is the reason §36.4's worksheet has a fifth line. Some models also carry disease-by-status interactions — a condition combined with disabled status, or with institutional residence.

Two practical points follow. First, an interaction cannot be reported. There is no code for it; it is computed by the model when both member conditions are present, which means the only way to earn it is for both underlying conditions to be documented and reported independently. Second, an interaction is where a single missing condition is worth more than its own coefficient, because dropping it also drops every interaction it participated in. In §36.4's example, losing the heart failure would cost 0.331 plus the 0.121 interaction — 0.452, not 0.331.

Check: 1.555 − 0.331 − 0.121 = 1.103, which is the score without heart failure ✓

🔍 Check Your Understanding

1. A record reports both E11.9 and E11.22 for the same member in the same collection year. How many diabetes coefficients does the model count, and which one?
2. The same member also has documented chronic kidney disease and chronic obstructive pulmonary disease. Do those trump anything, or are they trumped? Why?
3. A colleague proposes reporting the unspecified version of a condition "to be safe" when the specific version is documented. Give the two-part answer: what it does to this claim, and what it does to the score.
4. Why can an interaction term never be coded?

(1) One — the more severe. (2) Neither; they are separate families and each contributes its own coefficient. (3) Nothing to the claim; it under-describes the member and the hierarchy cannot recover the difference. (4) Because it is computed by the model from two conditions that must each be documented and reported on their own.)

36.6 The annual reset and why chronic conditions must be recaptured

Here is the rule that produces more real work than every other rule in this chapter combined:

Every condition resets to zero. A risk score is built from the diagnoses documented and reported during a defined collection period — for Medicare Advantage, essentially a calendar year — and used to set payment for the following payment year. Nothing carries forward. A chronic condition reported in one year and not reported in the next is, for the next year's payment, gone.

THE ANNUAL RESET

  COLLECTION YEAR (year 1)              PAYMENT YEAR (year 2)
  Jan ......................... Dec  ->  Jan ................. Dec
  every acceptable encounter's           the score built from year 1
  diagnoses are gathered                 pays every month of year 2
        |                                        |
        v                                        v
  the record must SAY IT AGAIN,          then year 2's encounters
  this year, on a face-to-face           build year 3's score, and
  encounter                              the clock resets again

  ON JANUARY 1, EVERY CHRONIC CONDITION IS BACK TO ZERO.

The reaction of every clinician hearing this for the first time is a good one and should be respected rather than argued with: the amputation did not grow back. Correct. The leg is still off, the transplant is still transplanted, the paraplegia is still permanent — and the model still requires the condition to be documented and reported this year, because the model is a statistical instrument fit on one year of claims and it has no memory. This is a genuine burden imposed on clinicians for the convenience of a payment methodology, and pretending otherwise is how a coder loses a physician's cooperation in the first meeting. Say it is a burden, say why the rule exists, and then teach the workflow that makes it cheap.

What recaptures a condition, and what does not

The requirement is not "the condition appears somewhere in the chart." It is a documented, face-to-face encounter with an acceptable provider type, in an acceptable setting, with the condition assessed and reported on the resulting claim or encounter submission. That excludes a surprising amount of activity:

Telehealth rules in this area have changed more than once and vary by program; verify the current policy rather than applying a remembered one.

The vehicle: the annual comprehensive visit

The practical answer to the reset is a planned annual visit at which every chronic condition on the member's record is actually assessed. For Medicare beneficiaries the annual wellness visit — G0439 for the subsequent-year version — is frequently the scheduling anchor, because it is a covered, no-coinsurance benefit that a practice can invite a patient to and that includes a health risk assessment and a review of the member's medical and family history.

Two cautions, and both matter. The annual wellness visit is a preventive service, not a problem-oriented one; a condition that is merely listed during it has not been addressed any more than a problem-list entry has. And where the clinician does perform significant, separately identifiable problem-oriented work at the same encounter, that is the ordinary modifier 25 situation from Chapter 14 §14.4 — the same analysis this book already ran on Account 10-4471, with the same requirement that the two services be documented distinctly.

📋 Read the Chart

```text FIGURE 36.1 - "The recapture worklist and what came back" [constructed teaching example] THE DOCUMENT A pre-visit worklist generated for a scheduled annual visit, printed the day before, plus the assessment section of the note that resulted. THE CONTEXT A primary care practice with a Medicare Advantage panel, February of the collection year.

THE WORKLIST PATIENT 1 of 14 - scheduled tomorrow, 9:20 Conditions documented in a prior year and NOT YET documented this year: * chronic diastolic heart failure last: prior yr * type 2 diabetes w/ chronic kidney dz last: prior yr * chronic kidney disease, stage 3a last: prior yr * major depressive disorder, recurrent last: 3 yrs ago "Assess and document each condition that is clinically present today. Document 'resolved' or 'no longer present' where that is the case. This list is a PROMPT TO EVALUATE, not a list of codes to affirm."

WHAT IT SHOWS The note that resulted addresses three of the four: heart failure (weight stable, no edema, continue furosemide), diabetes (A1c 7.1, metformin dose reduced for renal function), CKD 3a (eGFR 51, nephrology follow-up in 6 months). Each has a status and a plan. The fourth reads: "Depression - no symptoms for several years, off medication since 2 yrs ago, PHQ-9 today 1. Not an active problem."

WHAT IT DOESN'T It does not tell you the worklist was right. Three of the four conditions were confirmed by an examining clinician; that is the worklist doing its job. The fourth was refused, which is the worklist doing its other job.

THE DECISION Report the three assessed conditions, at the specificity the note supports. Report nothing for the depression, and record that it was evaluated and found inactive so that next year's worklist is smarter than this year's.

THE LESSON A recapture program is legitimate exactly to the extent that "no" is one of its expected answers. A list that only ever grows is not a clinical tool; it is a revenue instrument wearing a clinical tool's clothes, and 36.8 is about what happens next. ```

Chronic does not mean weighted

Account 10-4471's knee is the counterexample the reader already owns. Chapter 22 closed Q5: the imaging obtained after March 14 supports M17.11, unilateral primary osteoarthritis of the right knee, and M25.561 was correct for March 14 because a later diagnosis does not reach backward. Osteoarthritis is chronic, permanent, frequently disabling, and it must be documented every year for the record to describe the patient — and it is not among the categories the CMS-HCC model weights. Verify against the model in force, but the general shape holds and is the point: the model selects for cost prediction, not for importance. A practice that instructs its clinicians to "document the HCCs" has told them to write a chart optimized for a payment formula. A practice that instructs them to document the patient completely gets the risk-adjustment result as a by-product and gets a usable medical record as well.

And the reset runs in both directions. A condition that resolved must not be carried forward. Account 22-8891's J96.01 — acute respiratory failure during that admission — belongs to that year's score and to no other. A patient's pneumonia last winter is not this year's diagnosis. Chapter 10's Case Study 2 showed the worst version of this failure: a treated malignancy, prompted forward from a prior year and affirmed at the end of a long visit, so that a patient had active cancer on paper for five consecutive years. The recapture worklist and the malignancy that would not die are the same tool. What separates them is whether the clinician is being asked to evaluate or to confirm.

🎓 Exam Watch

Risk adjustment has its own credential — the Certified Risk Adjustment Coder (CRC) from AAPC — and Chapter 39 §39.3 places it among the specialty credentials. It is worth knowing about even if you never sit for it, because it signals that this is a distinct body of knowledge rather than an add-on to CPC work.

Four question shapes recur.

(1) The reset. A stem describes a condition documented in detail two years ago and not since, and asks whether it counts for this year. It does not. Candidates who answer from clinical logic ("the condition still exists") get it wrong; the question is about the collection period.

(2) The hierarchy. A stem reports two severities of the same disease and offers "both" among the choices. One counts — the most severe.

(3) The acceptable encounter. A stem supplies the condition in a laboratory report, a problem list, or a nurse's telephone note and asks whether it can be reported. It cannot. The discriminating fact is always the same: was there a face-to-face encounter with a diagnosing provider who assessed it?

(4) The linkage. A stem documents diabetes and a condition the classification joins under "with," without an explicit linking statement — Chapter 9 §9.7's convention, now with money on it. The presumption applies where the classification creates it and nowhere else, and the exam tests the boundary at least as often as the rule.

36.7 MEAT and the documentation risk adjustment requires

Every chapter of this book has depended on the first theme — if it isn't documented, it didn't happen. This is the section where the profession gave that theme an acronym.

The MEAT criteria are an industry mnemonic for the evidence a note must contain before a condition may be reported as addressed at an encounter:

• M — Monitor. Signs, symptoms, disease progression or regression, values followed over time.
• E — Evaluate. Test results reviewed, medication effectiveness assessed, response to treatment examined.
• A — Assess / Address. The condition is discussed, examined, counseled about, records reviewed, tests ordered because of it.
• T — Treat. Medications, therapies, procedures, referrals directed at the condition.

One of those four is enough. MEAT is not a checklist requiring all four; it is a list of the ways a note can demonstrate that the clinician engaged with the condition at this encounter.

And here is a piece of honesty most treatments of MEAT omit: MEAT is not a federal regulation. There is no CMS rule that says "MEAT." It is a teaching device the risk-adjustment community built to operationalize requirements that are binding — the ICD-10-CM Official Guidelines' Section IV instruction to report conditions that coexist and require or affect patient care, treatment, or management; the general documentation standards; and the payment rules' requirement of a face-to-face encounter with an assessment. Some organizations use a different mnemonic — TAMPER is the common alternative, expanding to treatment, assessment, monitoring or medicating, plan, evaluation, and referral — and it encodes the same underlying requirements. Use whichever your organization uses, and know that the authority is the Guidelines and the payment rules, not the acronym. A coder who cites MEAT to a physician as though it were law has invited a challenge they cannot answer.

MEAT APPLIED - the question the coder is actually asking

   For each condition on the record, this encounter:

   +-----------------------------------------------------------+
   |  Did the note MONITOR it?   (values, symptoms, course)     |
   |  Did the note EVALUATE it?  (results, response, effect)    |
   |  Did the note ASSESS it?    (status stated, plan stated)   |
   |  Did the note TREAT it?     (drug, therapy, referral)      |
   +-----------------------------------------------------------+
                          |
            ANY ONE, documented -> REPORTABLE at the
                                   specificity the note supports
                          |
            NONE            ------> NOT REPORTABLE. Not this year,
                                    not from this encounter, not
                                    because you know it is true.

The failure modes, named

The problem list is not documentation. It is a list. Chapter 9 §9.7 settled this for Account 10-4471 and it settles it everywhere: a condition on a list has been recorded, not addressed.

The history is not the assessment. This is subtler and it is the most common near-miss in real charts. A condition mentioned in the history of present illness, the past medical history, or the review of systems has been narrated. MEAT lives in the assessment and plan, where the clinician states what they concluded and what they are doing. Account 10-4471 already contains a documented instance of exactly this pathology — the note's evidence that conservative therapy for the knee had failed sits in the HPI rather than the assessment, which is documented gap 1 and which Chapter 22 §22.6 worked through for medical necessity. Same disease, different organ: the information exists and is in the wrong section of the note.

"Stable" is not "absent," and it is not "low risk." Chapter 15 §15.5 froze the definition that matters here: for medical decision making, stable means at treatment goal, not "unchanged." A diabetic at goal on metformin, with an A1c ordered and the medication continued, is a stable chronic condition that has been monitored, evaluated, and treated — fully reportable. Coders new to risk adjustment sometimes assume the opposite, that a stable condition is somehow less codable. It is not. And the converse trap is worse: a patient described as "stable" in the medical decision making sense is not thereby a low-risk member in the risk-adjustment sense. Her diabetes is at goal because it is being managed, and the cost of managing it is exactly what the model is pricing.

A medication list is not treatment documentation. Metformin appearing on a reconciled medication list demonstrates that the patient takes metformin. It does not, by itself, document that the diabetes was treated at this encounter. The note has to connect the drug to the condition — which most assessments do naturally, and which is why the fix here is usually a template change rather than a change in behavior.

Copy-forward is a liability, not an asset. Chapter 4 §4.6 defined cloned documentation and Chapter 38 §38.1 owns what to do about it. In this setting it has a specific consequence: six identical annual assessments of the same condition, word for word, are not six years of MEAT. They are one year of MEAT and five years of a copy operation, and an auditor reading them in sequence will say so.

Specificity is documentation too. The gap between "diabetes" and "diabetes with diabetic chronic kidney disease, stage 3a" is not a coding preference. It is the difference between two condition categories, and the coder cannot close it by inference — Chapter 4 §4.7's line has not moved.

⚠️ Where Claims Die

The four sentences that lose a condition, in order of how often you will see them.

1. "Continue current medications." Which condition? Everything and nothing. A blanket continuation statement supports no individual condition, and Chapter 15 §15.10 already showed what it does to an evaluation and management level: the assessment "chronic conditions stable, medications refilled" is arguable at best low on the problems element, because nothing was individually addressed.

2. "History of CHF." In American clinical usage, "history of" is ambiguous between "has, since 2019" and "had, and no longer has." Chapter 12 §12.9 owns the history codes and the distinction matters enormously here: the classification reads "history of" as a past condition, and a coder who reports an active condition from that phrase has supplied a clinical judgment.

3. "CKD 3a" appearing only in the problem list. No status, no plan, no assessment. §36.11 works this one at length, because it is on the book's own anchor file.

4. The condition assessed by a specialist and never by anyone else, in a practice that assumed the specialist reported it. Sometimes true, sometimes not — and "we assumed" is not a documentation strategy. The fix is not to report what you hope the nephrologist reported. It is to have the primary care note say what the primary care clinician assessed.

What the disciplined professional does instead: read the assessment and plan first, condition by condition, and report only what survives that reading. Then, where the record clearly contains the clinical facts and lacks the sentence, use the query — which is §36.11's subject, and whose compliant form belongs to Chapter 38 §38.3.

36.8 Chart review, prospective and retrospective, and where it goes wrong

Reviewing charts to find conditions that were documented but never reported is legitimate and, done properly, necessary. It is also the single most active enforcement area in this chapter's subject matter. Both sentences are true and this section is about holding them together.

A prospective chart review happens before an encounter: the record is reviewed and the clinician receives information — usually a list of previously documented conditions not yet addressed this year — to inform the visit.

A retrospective chart review happens after an encounter: completed records are reviewed to determine whether the diagnoses submitted match what the documentation supports, and submissions are corrected in either direction.

The line, stated once

Teaching a coder to recognize that a chart does not support a condition is instruction. Teaching a coder how to make a chart appear to support one is not, and this book does not do it. Everything below is written from the detection side: what the failure modes are, what they look like in data, and how each one is found.

Where prospective review goes wrong

The mechanism is a prompt that asks a clinician to affirm rather than to evaluate. Figure 36.1's worklist is the good version, and the sentence that makes it good is printed on it: this is a prompt to evaluate, not a list of codes to affirm. The bad version differs in small ways with large consequences — a list presented at the end of a long visit, a default that is already checked, a system that records a click as a physician's diagnosis, an incentive attached to how many prompts are closed, and no mechanism at all for recording that a condition is no longer present.

Chapter 10's Case Study 2 is the fully worked example and it belongs to this section: an annual review process, not improper in concept, that prompted a clinician with an active malignancy code from a prior year and produced a patient who had cancer on paper for five consecutive years. Nothing in that process was a lie. Every step was a small convenience.

Where retrospective review goes wrong

The one-way ratchet. A review program that submits added diagnoses and does not delete unsupported ones. This is the most frequently documented failure pattern in the public record, and it is the easiest to detect from the inside — you simply ask the vendor or the department for its deletion count. A review of thousands of charts that has never once removed a code has told you what it is.

Diagnoses that exist only in the review. A condition that appears on a chart-review submission or an in-home health risk assessment and never on any clinical encounter, never in a medication, never in a referral, never again. The Office of Inspector General (OIG) of the Department of Health and Human Services has published repeatedly on this pattern in Medicare Advantage, and it is the analytic signature auditors look for first, because it requires no chart at all — only the data.

The unsupported code. A submitted diagnosis the record does not support, whatever the process that produced it. Reporting it is a false statement; keeping the money once you know is a separate problem, governed by the sixty-day overpayment rule Chapter 31 §31.9 owns.

📋 Read the Chart

```text FIGURE 36.2 - "The finding a compliance officer wants" [constructed teaching example] THE DOCUMENT A one-page internal summary of a practice's annual review of its risk-adjustment submissions, prepared by the compliance officer for the physician partners. THE CONTEXT A primary care practice, first year of a Medicare Advantage arrangement, 470 attributed members.

WHAT IT SHOWS Charts reviewed: 120 Conditions supported and already reported: 312 Conditions supported and NOT reported: 37 -> submitted, with the encounter cited Conditions reported and NOT supported: 9 -> DELETED, and the deletion documented Net effect on the panel's aggregate score: a small increase, itemized chart by chart Every submitted and deleted code traced to a page of the record and to a signed, dated encounter.

WHAT IT DOESN'T It does not prove the 37 were right; it proves they were traceable. And it does not say the nine deletions were anyone's fault - two were duplicate submissions and four were conditions that had resolved.

THE DECISION File it. Keep the trace. Repeat it annually, and keep reporting the deletion count even in a year when it is embarrassing.

THE LESSON The number that makes this document credible is the NINE. A review that only adds is not a review. An external auditor's first question about any chart review program is what it has ever removed, and a program with a good answer has already survived the hardest part of the conversation. ```

How it is caught

Five mechanisms, roughly in order of how often each is the first to find something:

Data analysis without any chart. A diagnosis that appears from one source and nowhere else. A condition prevalence far from what the population would predict. A score that jumps in the first year of a program. Chapter 21 §21.10 taught pattern detection from claims data; this is that skill in a different data set, and the OIG performs it at national scale.

RADV — risk adjustment data validation. CMS's audit mechanism for Medicare Advantage: a sample of members from a contract, medical records requested to substantiate the diagnoses submitted for them, and unsupported diagnoses removed. CMS finalized its contract-level RADV rule in January 2023, and the two provisions that matter to a coder are these: findings will be extrapolated from the sample to the contract beginning with payment year 2018, and no "fee-for-service adjuster" is applied. The industry's objection was substantial — the argument being that the model is calibrated on fee-for-service data whose own diagnoses are not audited to the same standard — and litigation followed; check the current status of both the rule and the litigation before relying on any description of it, including this one.

Extrapolation is the multiplier, and Chapter 37 §37.6 owns its arithmetic. It is the reason a finding rate in a sample of a few hundred members is not a small problem.

The OIG Work Plan and audit reports. Published, searchable, and the closest thing this field has to an early warning system. Chapter 37 §37.5 places the OIG among the external reviewers.

The False Claims Act, usually with a relator. Chapter 5 §5.3 defined the statute and the qui tam mechanism. Enforcement activity against Medicare Advantage organizations and provider groups over diagnoses that records did not support has been extensive and sustained, and the theory is straightforward: submitting a diagnosis known to be unsupported, for payment, is a false claim. Amounts in this area have run into large figures and this book does not print any of them — read the Department of Justice's own announcements for the current record.

⚖️ Compliance Check

What a defensible risk-adjustment program can show, on the day somebody asks.

1. A written policy describing what is reviewed, by whom, under what credentials, and what happens to a code that is not supported.
2. Bidirectionality, evidenced. Adds and deletes, both counted, both reported, both reviewed by someone whose compensation does not depend on the direction.
3. A trace from every submitted diagnosis to a page of a signed, dated, face-to-face encounter note.
4. Physician ownership of every clinical statement. A coder never adds a diagnosis to a record. An addendum is made by the clinician, under Chapter 4 §4.5's amendment rules, dated when it was made.
5. Queries that are not leading, retained with their responses. Chapter 38 §38.3 owns the form.
6. A refusal path. Somebody, named, can decline to submit a code, and declining is not a performance problem.
7. A deletion and repayment process that connects to the sixty-day rule (Chapter 31 §31.9), because discovering an unsupported submission starts a clock.

And the sentence to carry: the accurate record and the defensible record are the same record. That has been this book's position since Chapter 1 and it has never been more literally true than here.

Requirements in this area change, differ by program and by contract, and are the subject of active rulemaking and litigation. Verify with your compliance officer, the plan's own program documentation, the current CMS guidance, and counsel. Nothing in this chapter is legal advice.

📞 On the Phone

A vendor call, second meeting. "Our retrospective review typically finds three to five additional conditions per chart, and we work on contingency — a percentage of the incremental revenue."

The three questions that should follow, in this order:

"How many codes did you delete last year, across all your clients? Not the rate — the count."

"When your reviewer finds a condition the record does not support, who decides, and does that decision reach the plan?"

"You are paid a percentage of the increase. Walk me through what your incentive is when a chart is ambiguous."

What the answers tell you. A vendor with a real deletion practice will answer the first question from memory, because it is the number their own compliance people ask about. A vendor who cannot answer it has told you the shape of the program. And the third question is not an accusation — a contingency arrangement is common and not per se improper — but it is a question a reasonable buyer asks, and an organization that reacts badly to being asked it has answered it.

36.9 Quality measures: process, outcome, and the codes that report them

Risk adjustment sets how much money is in the arrangement. Quality measurement decides whether the organization keeps any of it, and it is the counterweight that stops "spend less" from being a complete strategy. Chapter 2's Case Study 2 named this as one of the things the second attempt at capitation has that the first attempt lacked.

A quality measure is a standardized, specified calculation applied to a defined population to produce a rate. Its three moving parts are a denominator (who is eligible), a numerator (who met the standard), and exclusions (who is removed for a stated reason).

Chapter 29 §29.7's denominator discipline transfers here without modification. A measure is a definition before it is a number, and two organizations reporting "diabetes control" may be computing different things.

The three families

WHAT A MEASURE MEASURES

  STRUCTURE   Does the organization have the capability?
              (a registry, a certified record system, a care manager)
              Easy to verify, weakly related to what happens to patients.

  PROCESS     Was the right thing DONE?
              Was the A1c drawn? The eye exam done? The statin prescribed?
              + attributable, actionable, fast to move
              - doing the thing is not the same as the patient benefiting

  OUTCOME     Did the patient get a better RESULT?
              Is the A1c controlled? Was there a readmission? Did the
              blood pressure reach goal?
              + it is what anybody actually cares about
              - needs RISK ADJUSTMENT and ATTRIBUTION to be fair

  and alongside them:
  PATIENT EXPERIENCE   survey-based
  BALANCING            what got worse while the target got better

Notice that the outcome column loops back to §36.2. An outcome measure compares organizations whose patients differ, so it has to adjust for how sick those patients were — which means the same diagnosis codes that drive the payment also drive the fairness of the quality comparison. The two halves of this chapter are one system, and a practice that codes incompletely is simultaneously underpaid and unfairly measured, because it looks like a practice with healthier patients and worse outcomes.

The codes that report them

Measures are computed from claims, from electronic records (eCQMs, electronic clinical quality measures), from registries, or from a hybrid of claims plus chart abstraction. Where a claim is the reporting vehicle and the ordinary codes do not carry the needed fact, the reporting mechanism is CPT Category II.

Chapter 13 §13.2 introduced them and deferred the machinery here. They are five characters ending in F, they carry no relative value, they are not paid, and a coder in a practice participating in quality reporting may spend real time assigning codes that generate no revenue at all.

Those two are worth a moment side by side, because they are the clearest illustration in the code set of the distinction this section is built on. 3044F reports a value the patient's body produced. 0001F reports that a clinician did something. One is an outcome measure's numerator; the other is a process measure's. And the outcome one — as Case Study 2 in this chapter argues at length — is the one that can be pursued into harm.

Category II codes also carry their own performance measure exclusion modifiers, which are how a measure records a legitimate reason the standard was not met: 1P for medical reasons, 2P for patient reasons, 3P for system reasons, and 8P for an action not performed with the reason not otherwise specified. These are the mechanism by which a measure is supposed to accommodate the patient for whom the standard is wrong — and Case Study 2 is about what happens when an organization uses them to accommodate its own rate instead. Verify the current Category II code set and its modifiers each January; this section of CPT is revised more often than most.

🔢 Code It

The documentation (constructed teaching example): "Established patient, 66, seen for chronic disease follow-up. 1. Chronic diastolic heart failure — I50.32. Volume status assessed: no orthopnea, no edema, weight stable at 174 lb. Continue furosemide 20 mg daily and lisinopril. 2. Type 2 diabetes mellitus — A1c resulted today at 6.8%. Continue metformin, no change."

The path. The billable service is the office visit at the level the medical decision making supports (Chapter 15). The diagnoses are I50.32 and the appropriate diabetes code at the specificity the record supports. Then, if this practice reports these measures on claims: 0001F, because the note documents that heart failure was assessed with the specified elements, and 3044F, because the most recent A1c is below 7.0%. Both are reported with a charge the practice sets by its own convention — often \$0.00 or a nominal amount — and both will be adjudicated at nothing.

The plausible wrong answer: omitting them because "they don't pay." They do not pay on this claim. Whether the practice reports them may decide a quality rate that decides a shared-savings distribution eleven months from now, which is §36.10. A code with no allowed amount is not a code with no consequence — that sentence is this chapter in miniature and it applies equally to the diagnosis line and the Category II line.

MIPS, and the honest caveat about it

The Merit-based Incentive Payment System (MIPS) is the fee-for-service half of Medicare's Quality Payment Program, created by the Medicare Access and CHIP Reauthorization Act of 2015 (MACRA). Its structure is stable enough to teach and its details are not:

• Eligible clinicians are scored across four performance categories — quality, cost, improvement activities, and promoting interoperability — into a single final score.
• The final score is compared against a performance threshold, and the result is a payment adjustment to Medicare Part B payments applied two years after the performance year. Positive and negative adjustments are scaled to offset one another.
• Clinicians below a low-volume threshold are excluded, which exempts a great many small practices entirely.
• MIPS Value Pathways (MVPs) are a newer reporting route organized around a specialty or condition rather than a menu.
• The alternative track is participation in an Advanced Alternative Payment Model (APM), which routes a clinician out of MIPS and into the arrangement's own terms.

Every weight, threshold, category definition, measure specification, and reporting option in that list is set by annual rulemaking and has changed repeatedly since the program began. Anything this book printed as a number would be wrong within a year. Read the current year's Quality Payment Program resources, and treat any summary — including this one — as an orientation to the structure.

36.10 Shared savings and value-based contracts

The last piece is the contract that ties the two halves together. There is a spectrum, and knowing where an arrangement sits on it tells you most of what you need to know about it.

THE RISK SPECTRUM

  fee-for-service ........................................ full risk
       |            |             |            |              |
   FFS only    FFS + quality  UPSIDE-ONLY   TWO-SIDED     CAPITATION /
               bonus          shared        risk          global budget
                              savings       (up + down)
       |            |             |            |              |
  no risk      tiny upside   share of      share of      you hold the
                             savings       savings AND   whole budget
                             only          of losses

Shared savings is an arrangement in which an organization is measured against a benchmark — the expected cost of its attributed population — and receives a share of the difference if actual cost comes in below it, typically conditioned on meeting quality standards.

A value-based contract is any agreement that ties payment to measured cost, quality, or both, rather than to volume alone.

The reference example is the Medicare Shared Savings Program (MSSP), in which groups of providers form accountable care organizations (ACOs). The mechanics generalize to most commercial versions.

🧮 Run the Numbers

A shared-savings settlement, structurally [constructed teaching figures — every value in a real arrangement is a contract term]:

```text Attributed members 5,000 Benchmark, per member per year $12,000.00

---

BENCHMARK TOTAL 5,000 x $12,000.00 $60,000,000.00 ACTUAL EXPENDITURES $57,600,000.00

---

GROSS SAVINGS $2,400,000.00 as a percentage of benchmark 2,400,000 / 60,000,000 = 4.0%

MINIMUM SAVINGS RATE (a contract term) 2.0% 4.0% exceeds 2.0%, so savings are shared.

SHARING RATE (a contract term) 50% QUALITY MULTIPLIER (the quality score gate) 90%

DISTRIBUTION 2,400,000 x 0.50 x 0.90 $1,080,000.00 ```

Checks: 5,000 × 12,000.00 = 60,000,000.00 ✓ · 60,000,000.00 − 57,600,000.00 = 2,400,000.00 ✓ · 2,400,000 ÷ 60,000,000 = 0.04 ✓ · 2,400,000.00 × 0.50 = 1,200,000.00; × 0.90 = 1,080,000.00 ✓

Interpretation, and the three lines worth arguing about. The minimum savings rate exists because a population of 5,000 has statistical noise in it, and an organization should not be paid for luck; below the threshold, nothing is shared at all — an organization that saves 1.8% receives nothing. The sharing rate is the contract's central number and it is negotiated. And the quality multiplier is the entire reason §36.9 exists: 10% of this distribution — \$120,000.00 — turned on quality performance. Check: 1,200,000.00 − 1,080,000.00 = 120,000.00 ✓

What this does not show is the cost of earning it: the care managers, the data feeds, the after-hours access, the analyst. A distribution is not a margin, and an organization that budgets it as one is Chapter 2's Case Study 2 with better vocabulary.

The four terms that decide whether a contract can be met

Attribution. Which patients are yours? Every other number is computed on this population, and practices routinely sign before understanding it. Attribution may be prospective (the list is known in advance) or retrospective (the list is determined after the fact from where members actually received care), and it is usually assigned by a plurality of primary care services. Retrospective attribution means an organization spends a year managing a population it cannot see, which is a materially harder job.

The benchmark. What are you being compared to? Historical spending for the same population, regional spending, a trended national figure, or a blend — and whether it is risk-adjusted, which is where this section rejoins the rest of the chapter.

The quality gate. Which measures, whose specifications, what data source, and whether the gate is a threshold or a multiplier.

The data. What the plan will send, in what format, and how late. Chapter 2's Case Study 2 stated the rule: a payment model that a practice cannot measure is a payment model that practice should not sign.

Where risk adjustment and shared savings interact, and why it is not what people expect

This is the paragraph most treatments skip and it decides real strategy.

In a risk-adjusted capitated arrangement, more complete coding raises the revenue attached to the population. That is Medicare Advantage, and it is where the incentive in §36.8 lives.

In a shared-savings arrangement, more complete coding usually raises the benchmark you are measured against as well — because the benchmark is risk-adjusted using the same scores. Coding does not manufacture savings; it moves both sides of the comparison. And programs have been designed specifically to prevent it from doing otherwise: the Medicare Shared Savings Program caps how much an ACO's risk score may grow between periods, precisely so that documentation change cannot be mistaken for cost performance.

The consequence for the coder's own conscience is worth stating plainly. In one arrangement, better coding raises revenue; in another, it mostly does not; in a third, it changes what you are compared to. None of that changes the code. The record either supports a condition or it does not, and the reason this book keeps returning to that sentence is that it is the only position that survives a change in the contract — which will happen, roughly every three years, for the rest of your career.

🔍 Check Your Understanding

1. An organization comes in 1.6% below its benchmark against a minimum savings rate of 2.0%. What is its distribution?
2. Name the two things a benchmark must specify before a practice can tell whether the arrangement is achievable.
3. Why does more complete diagnosis coding raise revenue in a capitated arrangement but not obviously in a shared-savings one?
4. A practice is offered an upside-only arrangement with no data feed. What is the objection, in one sentence?

(1) Nothing. (2) What population it is computed on, and whether it is risk-adjusted. (3) Because the shared-savings benchmark is itself risk-adjusted, so the score moves both sides. (4) It cannot be managed, only received — the practice would be accepting a measurement it has no way to reproduce.)

36.11 The diabetes code that was right and incomplete

🗂️ The Encounter

What this chapter contributes: the revisit, and the closing of Q2.

Account 10-4471's line B has read E11.9 since Chapter 10, and this book has told you thirty times that Chapter 36 would come back for it. Here is what happens now, and the first thing that happens is that nothing gets taken back.

Chapter 9 §9.7 was right and stays right. E11.9 was correct on March 14. The reasoning is not being re-argued: the "with" convention would have permitted linking the diabetes and the chronic kidney disease without an explicit provider statement, but a prior question governed — Section IV reports additional diagnoses that are addressed, that affect treatment, or that require management at this encounter — and the assessment did not address the kidney disease. A coder who reported E11.22 and N18.31 from that record on March 15 would have answered the linkage question correctly and skipped the addressed question entirely. That coder would have been wrong, and this chapter does not rehabilitate them.

The question here is different: was E11.9 sufficient? No.

---

Step one: MEAT, applied to the actual note, condition by condition.

```text FIGURE 36.3 - "The March 14 assessment, read under MEAT" [Account 10-4471]

CONDITION M E A T REPORTABLE?

---

Type 2 diabetes yes yes yes yes YES - E11.9 "stable, continue metformin, A1c ordered today" monitored (status stated) - evaluated (A1c ordered to assess control) - assessed (status + plan) - treated (metformin continued)

Essential HTN yes yes yes yes YES - I10 "at goal, continue lisinopril, no change"; BP 132/84 recorded

Hyperlipidemia yes yes yes yes YES - E78.5 "continue atorvastatin, lipid panel ordered"

Right knee pain yes yes yes yes YES - M25.561 history, examination, injection performed; no definitive diagnosis established (that is Ch. 22's story, and Q5 is closed)

CKD, stage 3a no no no no NO - nothing Appears ONCE, on the reviewed problem list. No status. No plan. No mention anywhere in the assessment. No linkage documented in either direction. ZERO OF FOUR.

---

The coder read the record correctly. The record is the problem. ```

Four conditions pass on all four criteria. The fifth fails on all four. That is not a marginal call and there is no argument to be had about it — which is exactly why this file was chosen as the book's anchor. The failure is clean, the coding was right, and the patient is still described incorrectly.

---

Step two: the sentence that is missing, and why it is almost certainly true.

Read two frozen facts of this record together, which no chapter before this one was permitted to do.

The problem list carries chronic kidney disease, stage 3a. The medication list carries metformin 1000 mg BID, continued at that dose by this assessment.

Metformin dosing depends on renal function. A clinician continuing 1000 mg twice daily in a patient whose problem list says stage 3a has, as a matter of ordinary practice, thought about the kidneys. The note does not say so.

And a coder may not say it either. Chapter 4 §4.7's line has not moved an inch: the coder does not know what the provider did or thought; the coder knows what the provider wrote. Inferring that renal function was considered — from a dose, from a drug class, from clinical plausibility — is supplying a clinical judgment, and it is the same error, in a friendlier costume, as the risk-adjustment vendor in Chapter 9 §9.7's compliance note who treats the "with" convention as general permission.

This is the third time this book has found this exact shape of gap, and by now it should be recognizable on sight:

```text Ch. 14 4.4 The note never states the decision to inject was made at this visit. Strongly implied. Absent as a sentence. -> Ch. 30's appeal had to be CONSTRUCTED, not quoted.

Ch. 33 33.10 "Hypoxic" written where "acute respiratory failure with hypoxia" was true. Same patient, same care. -> $1,867.44, and a CDI query is the remedy.

Ch. 36 here Renal function considered, and unwritten. -> a condition disappears from a year's description of the patient.

THE PATTERN: not wrong. UNWRITTEN. Three settings, one disease. ```

---

Step three: the query that should have been sent — and what it may not ask.

The wrong query is the obvious one. "Is the patient's chronic kidney disease related to her diabetes?" is leading in this context, and worse, it is the wrong question: it answers linkage, which the classification already presumes, and it does nothing about whether the condition was addressed. A "yes" to it would not make the March 14 encounter into an encounter that managed the kidney disease.

The defensible query asks the clinical question the record actually raises, and offers "no" as a real answer:

"The assessment continues metformin 1000 mg BID. The reviewed problem list includes chronic kidney disease, stage 3a. Was renal function evaluated or considered in the management of the diabetes at this encounter? If so, please document. If the kidney disease was not addressed at this encounter, please indicate that as well."

Four things to notice about it. It quotes only the relevant excerpt (minimum necessary). It offers the negative answer with equal weight. It asks about this encounter, not about the patient's history in general. And it does not name a code, suggest a diagnosis, or mention payment. The form of a compliant query — the elements, the options, the retention, and the line past which a query becomes leading — belongs to Chapter 38 §38.3. What this chapter owns is the occasion: risk adjustment is why this query is worth someone's fifteen minutes on a record that has already been billed and paid.

And now the limit, which matters more than the query. Suppose the answer is "yes — I reviewed her eGFR before continuing metformin at this dose." An addendum under Chapter 4 §4.5's rules — dated when it is made, signed, identified as an addendum — now documents evaluation, and the encounter supports E11.22 + N18.31: type 2 diabetes with diabetic chronic kidney disease, and the stage. Chapter 11's Encounter checkpoint noted that two separate conventions would each have supported a more complete code here, since the classification also presumes a relationship between hypertension and chronic kidney disease; the I12.- category is the second of them, and the same addressed question governs it.

Suppose the answer is no. Then the code stays E11.9, the record is correct as written, and nothing has been lost except the illusion that a query is a tool for getting a better code. A query cannot manufacture an encounter. If the kidney disease was not managed on March 14, no answer to any question makes it reportable for March 14, and the real remedy is prospective: the practice should be surfacing unaddressed chronic conditions before the visit, which is §36.6's worklist and §36.8's legitimate half.

---

Step four: the consequence, and it is not about this claim.

The patient was seen twice earlier in the same calendar year and again on March 14 — three encounters, none of which addressed the kidney disease. The claim was denied, appealed, paid, and closed at zero balance on day 100, in late June. Nothing about the money on Account 10-4471 changes by one cent.

What changes is the description of the patient for the year. If no encounter in the remaining months addresses it, the collection year closes with this patient recorded as having type 2 diabetes without complications and no kidney disease at all. The practice's own data then cannot answer how many of our diabetics have kidney disease — Chapter 11's Case Study 2, on a different organ. And the nephrology referral that a stage-3a diabetic should be discussed for is not on anyone's list, because no list knows.

---

Step five: what it is worth.

The honest scoping first. This patient is 58 and covered by Northfield Mutual, a commercial PPO. She is not in a Medicare Advantage plan and no risk score is being computed for her by anyone. The arithmetic below is a lens, not this file's money — the same device Chapter 33's checkpoint used for the hip fracture and Chapter 26 §26.9 used for the provider-based clinic.

[constructed illustrative coefficients — verify current coefficients and the model version in force at CMS]

```text THE SAME CHART, IN A RISK-ADJUSTED PANEL - one member, one year

AS CODED (the record as written) demographic component 0.323 diabetes without complications E11.9 0.105 ----- RAF 0.428

AS THE PATIENT ACTUALLY IS (the record after the addendum) demographic component 0.323 diabetes with chronic complications E11.22 0.302 chronic kidney disease, moderate stage N18.31 0.127 ----- RAF 0.752

DIFFERENCE 0.324

at a constructed benchmark of $1,000.00 PMPM: 0.324 x $1,000.00 = $324.00 per member per month $324.00 x 12 = $3,888.00 for the payment year ```

Checks: 0.323 + 0.105 = 0.428 ✓ · 0.323 + 0.302 + 0.127 = 0.752 ✓ · 0.752 − 0.428 = 0.324 ✓ · 324.00 × 12 = 3,888.00 ✓

One member. One year. One unwritten sentence. And whether stage 3a chronic kidney disease is itself payment-eligible has differed between versions of the CMS-HCC model, which is its own lesson: a coder who learned "stage 3 doesn't count" and stopped reporting it was wrong the year the model changed. Report what the record supports; let the model decide what it weighs.

Now scale it the way a practice must. Suppose Northgate has 470 patients enrolled in Medicare Advantage plans [constructed], and a review of a sample finds the same pattern — a chronic condition on the problem list that no encounter this year addressed — in 25 of them:

text 25 members x $3,888.00 = $97,200.00 for one payment year

Check: 3,888.00 × 25 = 97,200.00 ✓

And immediately, the two honest qualifications. First, that money goes to the plan, not to Northgate — what Northgate receives is whatever §36.10's contract says, and in an upside-only shared-savings arrangement the answer may be nothing at all. Second, and this is the argument that actually matters: the practice's commercial charts and its Medicare Advantage charts are written by the same clinicians using the same templates and read by the same coder. A documentation habit is not payer-specific. The March 14 note is evidence about how this practice documents, and it is evidence collected on a chart where nothing was at stake.

---

What this settles. Q2 is CLOSED. The diabetes code was correct on March 14 (Chapter 9) and insufficient as a description of the patient (this chapter). Both halves are true, neither cancels the other, and the difference between them is worth money. The record, not the coder, is what needs to change — by a query answered honestly, or better, by a visit that addresses the condition.

What it does not settle. Whether the practice will do anything about it. Nothing in the payment system will tell them to; there is no denial, no variance, no report. This finding exists only because somebody read the chart against the problem list and noticed that a condition had been carried for years and addressed by no one.

Open questions. Q4 — could the denial have been prevented? — remains open and belongs to Chapter 40. Every other question on this file is now closed: Q1 at Chapter 14, Q3 and Q5 at Chapter 22, Q6 at Chapter 23, and Q2 here.

Summary

Risk-adjusted payment pays for a member, not for a visit. A plan receives a fixed amount per member per month, adjusted by a score built from the diagnoses documented and reported for that member during the previous year. No claim in that system pays for a diagnosis; the diagnoses set the pot before any claim exists. An incomplete record produces no denial, no variance, and no report — which makes this the only failure in the book that generates no artifact at all.

Risk adjustment exists to remove the incentive to avoid sick people. Any fixed per-member payment creates that incentive; adjusting the payment by expected cost is the countermeasure. The models predict cost, not sickness; they are relative, normalized near 1.0; and they take their evidence from last year to pay for next year. The design pays for documentation, which is why §36.8's enforcement apparatus exists — and why Medicare Advantage scores carry a statutory coding intensity reduction before anyone is paid.

A hierarchical condition category groups diagnoses that predict similar incremental cost. Most ICD-10-CM codes map to nothing. Where the model does distinguish severities it does so through the code's specificity, and there are several different models — CMS-HCC, the ESRD model, RxHCC, HHS-HCC, state Medicaid models — that legitimately disagree with one another. Category definitions, mappings, and every coefficient are revised annually and the version in force differs by contract year, which is why this book prints no HCC number and no coefficient as fact.

The risk adjustment factor is a sum: a demographic component the chart cannot touch, plus one coefficient per condition category surviving the hierarchies, plus modeled interactions — then normalized and adjusted at the plan level. Applied to a benchmark, it produces a monthly payment. Hierarchies mean the most severe category in a family trumps the rest, so stacking severities gains nothing while under-specifying loses everything, and an interaction can never be coded because it is computed from two conditions that must each be reported.

Every condition resets to zero on January 1. A chronic condition must be documented and reported each collection year on a face-to-face encounter with a diagnosing provider; a problem list, a laboratory result, a refill, and a copied-forward note do not recapture it. The annual comprehensive visit — often anchored on G0439 — is the practical vehicle, and the reset runs both ways: a resolved condition must not be carried forward.

MEAT — monitor, evaluate, assess, treat — is the evidence a note must show, and any one of the four suffices. It is a mnemonic, not a regulation; the authority is Section IV of the Official Guidelines and the payment rules. Its failure modes are consistent: the problem list is not documentation, the history is not the assessment, "stable" means at treatment goal and is fully reportable, a medication list is not treatment, copy-forward is a liability, and specificity is documentation the coder may not supply by inference.

Chart review is legitimate and it is the most enforced activity in this chapter. Prospective review fails when a prompt asks a clinician to affirm rather than evaluate. Retrospective review fails as a one-way ratchet, as diagnoses that exist only in the review, and as codes the record does not support. It is caught by data analysis, by RADV — whose 2023 final rule extrapolates findings from payment year 2018 and applies no fee-for-service adjuster — by the OIG, and by False Claims Act actions. A defensible program can show its deletions.

Quality measurement is the counterweight. Process measures ask whether the right thing was done; outcome measures ask whether the patient got a better result and therefore need risk adjustment and attribution to be fair — which is where the two halves of this chapter turn out to be one system. CPT Category II codes report measures and are never paid: 0001F reports that heart failure was assessed (process), 3044F reports an A1c below 7.0% (outcome), and modifiers 1P/2P/3P/8P record why a standard was legitimately not met. MIPS is real, structured in four categories, adjusts Part B payment two years later — and every one of its numbers changes annually.

Shared savings pays a share of the difference between a benchmark and actual cost, gated by quality. Attribution, the benchmark, the quality gate, and the data feed are the four terms that decide whether a contract is achievable. And the interaction people expect is backwards: in a capitated arrangement complete coding raises revenue; in a shared-savings arrangement it usually raises the benchmark too, which is why programs cap risk score growth. None of it changes the code.

And Account 10-4471's diabetes line closes the question it opened in Chapter 4. Under MEAT, four conditions on the March 14 note pass on all four criteria and the fifth — chronic kidney disease, stage 3a — fails on all four, appearing once, on a problem list, addressed by nobody. The metformin dose says the clinician thought about the kidneys; the note does not, and no coder may say it for them. E11.9 was correct and E11.9 was insufficient, and the difference between those two sentences is what this chapter has been about.

Key Terms

Risk adjustment — modifying a fixed per-member payment according to a member's expected relative cost, using documented demographics and diagnoses as predictors; the mechanism that removes the financial incentive to avoid sick people. (Ch.36)

Hierarchical condition category (HCC) — a group of clinically related diagnoses that predict similar incremental cost, carrying a single coefficient in a risk model and ranked within its disease family. (Ch.36)

Risk adjustment factor (RAF) — the risk score: one number per member per payment year, built as demographic component + disease coefficients + interactions, normalized near 1.0 and multiplied by a benchmark to produce payment. (Ch.36)

Demographic coefficient — the portion of a risk score derived from enrollment facts: age band, sex, dual/Medicaid status, institutional versus community residence, originally-disabled status. Nothing in a chart changes it. (Ch.36)

Disease coefficient — the weight attached to a condition category that a member's reported diagnoses reached during the collection year. (Ch.36)

Hierarchy — the ranking of condition categories within one disease family, from most to least severe. (Ch.36)

Trumping — the operation by which the highest-ranked category a member reaches is retained and every lower category in the same family is zeroed for the year. (Ch.36)

Interaction — a modeled combination of conditions (or a condition and a status) that costs more together than separately, computed by the model and never codable. (Ch.36)

Annual reset — the rule that every condition returns to zero each collection year and must be documented and reported again on a face-to-face encounter, however permanent the condition. (Ch.36)

MEAT criteria — monitor, evaluate, assess/address, treat: the industry mnemonic for the evidence a note must contain before a condition is reportable as addressed; any one suffices, and the binding authority is the Official Guidelines and the payment rules, not the acronym. (Ch.36)

Prospective chart review — review before an encounter, producing information for the clinician; legitimate when it prompts evaluation, improper when it prompts affirmation. (Ch.36)

Retrospective chart review — review after an encounter to reconcile submitted diagnoses to the documentation; defensible only when it deletes as well as adds. (Ch.36)

Quality measure — a specified calculation over a defined population producing a rate, with a denominator, a numerator, and stated exclusions. (Ch.36)

Process measure — a measure of whether the right action was taken (the test ordered, the exam performed): attributable and actionable, weakly linked to outcome. (Ch.36)

Outcome measure — a measure of the result achieved (the value controlled, the readmission avoided): what people care about, and unfair without risk adjustment and sound attribution. (Ch.36)

MIPS (Merit-based Incentive Payment System) — the fee-for-service track of Medicare's Quality Payment Program under MACRA: four performance categories, one final score against a threshold, and a Part B payment adjustment applied two years later. Details change annually. (Ch.36)

Shared savings — an arrangement paying a share of the difference between a benchmark and actual expenditures for an attributed population, typically gated by quality performance. (Ch.36)

Value-based contract — any agreement tying payment to measured cost, quality, or both rather than to volume alone. (Ch.36)

Attribution — the method by which patients are assigned to an organization for measurement and payment; prospective or retrospective, usually by plurality of primary care services. (Ch.36)

Per member per month (PMPM) — the unit of payment and of measurement in capitated and risk-adjusted arrangements. (Ch.36)

Spaced Review

1. State, in two sentences, the difference between the question Chapter 9 §9.7 answered about Account 10-4471's diabetes code and the question this chapter answered. Then say why both answers can be correct at once.

2. (Chapter 35) A specialty practice tells you its coders "don't do risk adjustment, we're procedural." Give the two reasons that is only half true, using this chapter's rules about where diagnoses are collected from.

3. (Chapter 10) An annual review process prompts a clinician with an active malignancy code from a prior year, and the clinician affirms it at the end of a long visit. Name the prompt design that caused it, the MEAT criterion that was never met, and the one sentence Figure 36.1's worklist carries that would have prevented it.

4. (Chapter 11) A note documents "hypertension" and "heart failure" with no linking language, and separately carries chronic kidney disease on the problem list without addressing it. Which combination codes are and are not available, and which question governs before any linkage convention is reached?

5. A member's record supports diabetes with chronic complications, chronic kidney disease, and heart failure. A colleague also reports diabetes without complications "for completeness." Using the constructed coefficients in §36.4, compute what the model counts, and state what would happen to the score if the heart failure were dropped.