CommuniCare Medical Coding AI | Digital Scientists Case Study

Overview

Closing the documentation-to-coding gap at scale.

In Medicare Advantage, Risk Adjustment Factor (RAF) scores determine reimbursement rates. Missed Hierarchical Condition Category (HCC) codes mean missed revenue—and inaccurate patient risk profiles that affect care planning. CommuniCare Health Services, one of the nation's largest post-acute care providers, had a problem at scale: diagnoses documented in clinical notes weren't making it to claims.

Digital Scientists built an AI-powered coding engine that ingests data from six clinical sources—diagnoses, health risk assessments, medications, progress notes, lab results, and claims—to identify missed ICD-10 codes with confidence scoring and clinical evidence linking. The platform reviews each patient in roughly 60 seconds, compared to the hour-long manual audits it replaced.

The engine composition is 79% SQL-driven queries against structured clinical data, 20% generative AI analysis using GPT-4, and 1% DBSCAN clustering for pattern detection. Every recommendation links directly to supporting documentation—audit-defensible from day one.

Client

CommuniCare Health Services

Industry

Healthcare, Post-Acute Care

Services

AI/ML, Data Engineering, Product Design

Engagement

Aug 2023 – Present

Method

Discover Engineer

Capabilities

AI / ML Data

The Challenge

Millions in missed risk adjustment from incomplete coding.

CommuniCare's providers were documenting conditions in clinical notes, but those diagnoses weren't consistently being captured as coded claims. With 20,000+ patients across 7 states and 245 providers, there was no systematic way to ensure HCC codes matched what was documented in the chart.

The gap wasn't competence—it was capacity. Manual chart audits took roughly an hour per patient. Chronic conditions from prior years weren't being recaptured annually. And every uncoded diagnosis created audit exposure on top of lost revenue.

Before Digital Scientists

1 hour per patient audit

Manual review of clinical documentation, medications, lab results, claims history

Documentation-to-coding gap

Diagnoses in notes not making it to claims—HCC codes left on the table

Chronic condition recapture failure

HCCs from prior years not being recaptured annually as CMS requires

Audit exposure at scale

Coding without clear documentation linkage across 20K+ patients and 7 states

What We Built

An AI engine that reviews patients in 60 seconds.

The platform analyzes six clinical data sources per patient to surface missed ICD-10 codes, each with a confidence score and direct link to supporting documentation. Coders review, accept, or modify—never black-box automation.

ICD code recapture flow showing prior year vs current year diagnosis comparison

AI Diagnosis Engine

Ingests diagnoses, health risk assessments, medications, progress notes, lab results, and claims data. Identifies missed ICD-10 codes by comparing documented conditions against what's been coded.

Confidence scoring per recommendation
Prior-year HCC recapture identification
Clinical evidence linking for every code

Coder Workbench

Purpose-built interface for medical coders. Each patient shows AI-suggested ICD codes with confidence scores, plain-language reasons derived from clinical notes, and one-click edit actions.

Patient info + crosswalk IDS in one view
AI reasons explain why each code was suggested
HCC, coefficient, and comorbidity data per code

RAF Calculation view showing Base RAF 3.068 projected to 5.096 with ICD code breakdown

RAF Score Analytics

Real-time RAF calculation showing base vs. projected scores per patient. The dashboard breaks down each ICD code's contribution to the RAF score with HCC category, coefficient, and comorbidity multipliers.

Base vs. projected RAF side by side
ICD-to-HCC coefficient mapping
Facility-level and provider-level comparisons

How It Works

Six data sources. Four AI models. One prioritized recommendation list.

The engine pulls from every relevant clinical data source available for each patient, runs four parallel analysis models, and produces a ranked list of missed ICD-10 codes—each with a confidence score and audit-ready evidence chain.

Data Sources

Diagnoses

ICD-10 codes

Health Risk Assessments

HRA records

Medications

Rx data

Progress Notes

Clinical narratives

Lab Results

Diagnostic labs

Claims Data

Prior-year claims

AI Models

Model 01

DBSCAN Clustering

Unsupervised clustering (eps=0.62, k=7) identifies patient cohorts with similar clinical profiles but different coding patterns—surfacing systematic gaps.

Model 02

GPT-4 Analysis

Generative AI reviews clinical narratives to identify implied diagnoses—conditions described but not explicitly coded. Provides plain-language reasoning for each suggestion.

Model 03

NLP Notes Mining

Natural language processing extracts structured clinical signals from unstructured progress notes, identifying conditions referenced in free-text documentation.

Model 04

Claims Gap Analysis

Compares current-year coded claims against prior-year HCC history to identify chronic conditions that should be recaptured annually per CMS requirements.

60 sec

Per-patient review time (vs. ~1 hour manual)

~40%

Of patients had additional valid ICD-10 codes identified

~0.2

Average RAF score increase per impacted patient

Our Approach

Two phases. Data foundation first, then AI.

We didn't start with the AI model. We started with the data—building a secure pipeline from CommuniCare's on-premise MSSQL databases to our AWS environment via VPN tunnel, then validating data quality before training began.

Phase 1

Aug – Nov 2023

Data Pipeline & Model Development

Secure VPN tunnel from on-premise MSSQL to AWS
7 tables across 3 databases mapped and normalized
DBSCAN clustering and GPT-4 model training
HIPAA and CMS RAPS compliance validation

Phase 2

Nov 2023 – Present

Platform Build & Production Deployment

React/TypeScript frontend with Material UI coder workbench
Python/FastAPI backend on AWS Kubernetes (EKS)
Gehrimed, PointClickCare, and Acclivity EHR integrations
Rollout across 7 states, 245 providers, 20K+ patients

The Results

$10M+ in RAF improvements. $2K+ per impacted patient per year.

These numbers come from actual RAF score improvements tracked across CommuniCare's Medicare Advantage population. Roughly 40% of patients reviewed had additional valid ICD-10 codes identified by the engine.

Financial Impact

Total RAF Improvements $10M+

Annualized revenue impact from more accurate risk adjustment coding.

Per-Patient Impact >$2K/yr

Additional annual reimbursement per impacted patient from corrected RAF scores.

Avg RAF Score Increase ~0.2 pts

Mean RAF score improvement per impacted patient—from codes the engine identified that manual review missed.

Operational Impact

Review Speed 50X faster

60-second AI-assisted audit vs. ~1 hour manual chart review per patient.

Patients with New Codes ~40%

Of patients reviewed had additional valid ICD-10 codes identified by the AI engine.

Diagnosis Accuracy 90%+

AI recommendations validated against clinical documentation with high accuracy.

Integrations

Wrapped around existing EHR systems.

No rip-and-replace. The platform integrates directly with CommuniCare's existing EHR ecosystem, pulling clinical data for analysis and surfacing recommendations within the coder's workflow. No separate login, no duplicate data entry.

Production Integrations

Standards & Compliance

HIPAA Compliant

End-to-end encryption, VPN tunnel to on-premise data, AWS HIPAA-eligible services

CMS RAPS Compliance

Coding recommendations align with CMS Risk Adjustment Processing System requirements

Audit-Defensible

Every recommendation linked to clinical documentation with full evidence chain

Gated Learning

Deliberate, validated model improvements—not automated retraining. Every update reviewed before deployment.

Why It Worked

Responsible AI, not black-box automation.

Healthcare AI fails when it's treated as a black box. This platform was built for explainability, auditability, and human oversight at every step.

79% structured queries, 20% AI

Most of the engine runs on deterministic SQL queries against structured clinical data. Generative AI handles the hard part—unstructured notes and implied diagnoses—but it's 20% of the workload, not 100%. This makes the system explainable and predictable.

Every code is audit-defensible

Each AI recommendation links directly to the clinical documentation that supports it. When auditors ask "why was this coded?"—the evidence chain is one click away. No unsupported codes. No guessing.

Gated learning, not continuous

The model doesn't retrain itself automatically. Improvements are deliberate and validated—every update reviewed by clinical and engineering teams before production deployment. This prevents drift and ensures consistency.

Related Work