Clinical & Coding Depth: Edits, Criteria, DRGs, and Risk Adjustment

Lesson 1: NCCI edits — the procedure-to-procedure rulebook

The National Correct Coding Initiative (NCCI) is a CMS program that publishes edits to prevent improper payments. There are two main edit types:

NCCI Procedure-to-Procedure (PTP) edits

PTP edits identify code pairs that should not be billed together on the same day for the same patient by the same provider. Each pair has a Column 1 code (the one that gets paid) and a Column 2 code (the one that gets denied as bundled).

Each pair carries a modifier indicator:

NCCI Medically Unlikely Edits (MUE)

MUEs cap the number of units of a single code billable for the same patient on the same date of service. They're set per code, by setting (practitioner, outpatient hospital, DME). Each MUE has a MAI (MUE Adjudication Indicator):

MAI 2 is the unforgiving one. If your AI generates or suggests claims, it must enforce MAI-2 MUEs as hard constraints, not soft signals.

How to use NCCI in an AI pipeline

1. Download the quarterly NCCI PTP file and the MUE file from CMS. They update every quarter.
2. Load into a queryable store (Postgres, DuckDB, or a key-value store keyed on code pairs).
3. Pre-bill: for every claim, check all line-pair combinations against PTP, and every line's units against MUE.
4. Surface the specific edit code and the modifier indicator so a human (or your appeals model) knows whether unbundling is even possible.

Lesson 2: LCDs and NCDs — what counts as medically necessary

Edits catch billing errors. Coverage policies determine whether a service is even covered for a given patient and condition. CMS publishes two kinds:

National Coverage Determinations (NCDs)

NCDs apply to all Medicare beneficiaries nationwide. There are roughly 300 active NCDs covering specific services. Each NCD lists the conditions (ICD-10s), frequencies, settings, and other criteria under which a service is covered.

Example: NCD 220.6.17 (PET scan for oncology) lists the cancer types for which FDG-PET is covered, frequency limits, and clinical criteria.

Local Coverage Determinations (LCDs)

LCDs are issued by Medicare Administrative Contractors (MACs) for their geographic jurisdiction. Where no NCD exists, the local MAC's LCD applies. LCDs are more numerous (thousands) and more granular than NCDs.

Each LCD has a paired Local Coverage Article (LCA) that lists the specific CPT and ICD-10 codes that satisfy or fail the coverage criteria. This is the part your AI consumes.

Where Medicare Advantage fits

Per the MA Final Rule (2024), MA plans must follow Traditional Medicare NCDs and LCDs as a floor. They cannot use proprietary criteria more restrictive than the applicable NCD/LCD. This was a major shift — many MA plans had to retire proprietary policies that were tighter than CMS standards.

Lesson 3: InterQual versus MCG — the criteria sets behind utilization management

For acute admissions and post-acute care, payers don't just check codes — they apply clinical criteria. Two products dominate the market:

InterQual mechanics (simplified)

InterQual criteria are structured as decision trees. For an inpatient admission, you generally walk:

1. Severity of Illness (SI) — does the patient's clinical picture meet thresholds (vitals, labs, imaging findings)?
2. Intensity of Service (IS) — does the planned care require inpatient resources (telemetry, IV meds requiring monitoring, etc.)?
3. If both pass — admission is supported. If only one passes — observation status is more appropriate. If neither — outpatient.

MCG mechanics (simplified)

MCG presents care guidelines: an admission criteria block, expected goal length of stay, and clinical indications for continued stay or discharge. Reviewers document met/not-met against the criteria language.

Lesson 4: DRGs — how inpatient hospital reimbursement actually works

Inpatient hospital claims don't get paid per-procedure. Under Medicare's Inpatient Prospective Payment System (IPPS) and most commercial inpatient contracts, the hospital gets one lump-sum payment per admission determined by which DRG (Diagnosis-Related Group) the case falls into.

The current Medicare system is MS-DRG (Medicare Severity DRGs), with about 770 distinct DRGs. Each DRG has:

• A relative weight (heavier DRGs pay more)
• An expected geometric mean length of stay
• An expected arithmetic mean length of stay
• Inclusion criteria based on principal diagnosis, secondary diagnoses, procedures, and discharge disposition

How a case groups into a DRG

1. Principal diagnosis — the condition determined after study to be the reason for admission. ICD-10-CM.
2. Major Diagnostic Category (MDC) — mapped from the principal diagnosis (e.g., MDC 4 = diseases & disorders of the respiratory system).
3. Surgical vs. medical partition — if a relevant OR procedure was performed, the case goes to a surgical DRG; otherwise medical.
4. Severity adjustment — secondary diagnoses are checked for CC (Complication or Comorbidity) or MCC (Major Complication or Comorbidity) status. Many DRG triplets exist as "with MCC / with CC / without CC/MCC."

DRG validation: where audits happen

Payers conduct DRG validation audits (pre-pay or post-pay) to verify the principal diagnosis, secondary diagnoses, and procedures coded support the DRG billed. The most common DRG-shift findings:

• Principal diagnosis changed (sequencing) — moves the case to a different MDC entirely.
• An MCC was coded but documentation doesn't support it — case drops from "with MCC" to "with CC" or below.
• A CC was coded but the secondary diagnosis isn't valid as a CC — same drop.
• Procedure coding error changing surgical vs medical partition.

AI for DRG validation reads the discharge summary and full chart, identifies what diagnoses are supported by documentation per ICD-10-CM Official Guidelines, and compares to what was actually coded. The legal framework is significant: improperly upcoded DRGs can trigger False Claims Act exposure.

Lesson 5: HCC coding and risk adjustment — how MA plans get paid

Medicare Advantage plans receive a per-member, per-month (PMPM) capitation payment from CMS. That payment is risk-adjusted: sicker members generate higher payments. The mechanism is HCC (Hierarchical Condition Category) coding and the RAF (Risk Adjustment Factor) score.

How HCCs work

CMS maintains a mapping from ICD-10-CM diagnosis codes to HCC categories. The current CMS-HCC v28 model has about 115 HCCs. Each HCC has a coefficient (its relative weight). For each member, the plan sums:

• Demographic coefficients (age, sex, Medicaid status, etc.)
• Coefficients for each HCC the member had documented in the data year
• Interaction coefficients (e.g., diabetes + heart failure has an interaction term)

The sum is the member's RAF score, normalized so a typical FFS Medicare beneficiary equals 1.0. A member with a RAF of 1.7 generates 70% more PMPM revenue than one at 1.0.

Hierarchy

HCCs hierarchical — if a member has both HCC 17 (Diabetes with Acute) and HCC 19 (Diabetes without), only HCC 17 counts. The more severe code in the same hierarchy trumps the less severe.

Why this matters for AI

Risk adjustment is one of the largest dollar levers in MA. Major use cases:

• Suspect identification — model reads charts and historical claims to surface HCCs that are clinically present but not yet coded for the current year.
• Documentation validation — verifies that diagnoses coded are actually supported by valid documentation per the Monitoring, Evaluating, Assessing, or Treating (MEAT) criteria.
• RADV preparation — Risk Adjustment Data Validation audits by CMS retroactively. If documentation doesn't support a coded HCC, CMS extrapolates and recoups. Massive financial exposure.

Certifications to layer in during Step 2

At least one engineer or clinical informaticist per pod should pursue these. They translate the lessons above into operational fluency:

Order of priority depends on use case: CPC first if you're broad, CCS first if you're inpatient-heavy, CRC first if you're MA risk adjustment, CPMA first if you're payment integrity.