Definition
Statistical comparison of the organization's DRG/diagnosis acuity distribution before and after AI-driven CDI deployment, controlling for actual patient population changes, measuring whether AI coding recommendations are systematically shifting the acuity distribution upward independent of genuine clinical changes.
The Measurement Gap
The Decision That Legacy Metrics Cannot Support
Is the CMI increase the organization is seeing a real reflection of sicker patients, or is the AI CDI tool systematically recommending higher-acuity codes in ways that create audit and compliance risk?
What Goes Unmeasured Today
AI-driven Clinical Documentation Improvement (CDI) tools query physicians about missing or underspecified diagnoses, prompting documentation that supports higher-acuity coding. When a CDI tool suggests that a patient's "pneumonia" should be documented as "sepsis due to pneumonia," the resulting DRG change increases reimbursement. Healthcare organizations celebrate the CMI increase as revenue improvement. What they do not measure is whether the increase reflects genuine clinical documentation improvement or systematic upcoding driven by the AI's optimization target.
Legacy CDI metrics (query response rate, CMI improvement, coding accuracy rate) measure whether the CDI process is working as designed. But "working as designed" for an AI CDI tool may mean systematically identifying ambiguous documentation and recommending the highest-acuity interpretation. This is not coding fraud in the traditional sense (the documentation may technically support the code) but it represents a systematic shift in coding behavior that creates audit exposure.
The measurement gap is the inability to distinguish between three sources of CMI change: genuine patient population changes (sicker patients), legitimate documentation improvement (capturing conditions that were present but under-documented), and AI-driven acuity inflation (interpreting ambiguous documentation as higher acuity). Without controlling for patient population characteristics, the organization cannot separate these effects.
The Invisible Failure Mode
AI-driven CDI systematically shifts the coding distribution toward higher acuity. The shift is gradual, difficult to distinguish from legitimate improvement, and invisible in aggregate CDI metrics that treat all CMI increases as positive outcomes. The organization's audit risk increases over time without any internal metric flagging the problem: until a Recovery Audit Contractor identifies the pattern externally.
Formula
Primary Formula
ACDD = (CMI_post − CMI_expected) / CMI_expected
Variable Definitions
| Variable | Definition | Unit |
|---|---|---|
| CMI_post | Actual case mix index in the post-AI CDI period | Index value |
| CMI_expected | Expected CMI based on patient population characteristics using pre-AI CDI coding patterns | Index value |
| CMI_expected derivation | Apply pre-AI coding patterns to the current patient population (age, comorbidity index, admission source, payer mix) | Regression model |
Data Requirements
| Data Element | Source System | Classification |
|---|---|---|
| DRG assignments and weights (pre and post AI CDI) | Claims/billing system, encoder | NATIVE |
| Patient demographics and clinical complexity | EHR: admission data, comorbidity indices | NATIVE |
| AI CDI query activity (suggestions made, accepted, rejected) | CDI platform | NATIVE |
| Pre-AI baseline coding patterns (12+ months) | Historical claims data | NATIVE |
Edge Cases and Exclusions
Some CMI increase from CDI is appropriate: physicians were under-documenting conditions that were present. The challenge is distinguishing "captured more of what was there" from "interpreted ambiguous documentation as higher acuity." Pre-AI baseline needs minimum 12 months for stable coding patterns. Control for payer-driven coding incentives separately from AI-driven behavior. Minimum sample: 1,000 discharges per measurement period.
Proposed formula, pending validation pilots.