← Back to Projects

Case Study

ClaimGuard-AI

Pre-submission healthcare claim denial-risk engine trained on real CMS CERT audit data (ROC-AUC 0.745, PR-AUC 0.295 vs 0.141 base rate). CARC/RARC crosswalk, expected-recovery-$ queue, model-card page, 76 tests.

Executive Summary

ClaimGuard-AI is a pre-submission risk engine for healthcare revenue cycle management, started at the AIxBio Hackathon and since rebuilt around real data. The model is trained and evaluated on public CMS CERT improper-payment audit files with a temporal split (train 2021-2023, validate 2024, test 2025) — the honest deployment scenario of predicting next year from past years.

Problem & Constraints

Claims get denied over documentation gaps, procedure mismatches, and payer policy violations. Manual scrubbing is slow and ignores expected financial recovery when ordering the review queue.

Architecture

Claim fields + physician note → LLM JSON extraction (validated, with fallback) → XGBoost denial probability → CARC/RARC denial-reason crosswalk → expected-recovery-$ knapsack queue (DuckDB) → Next.js auditor worklist + model-card page.

Methodology

  • XGBoost classifier on real CERT claims with temporal train/val/test split — no group leakage
  • CARC/RARC denial-reason crosswalk derived from the audit data
  • Review queue ranked by expected recovery dollars; the overturn-rate assumption (0.5) is labeled as an assumption in the UI
  • Per-claim risk drivers from XGBoost pred_contribs (no external explainability dependency)
  • Appeals letters cite the specific driver evidence and carry a human-review banner
  • Public model-card page documents data provenance, split, metrics, and limitations

Results & Metrics

MetricResult
ROC-AUC (2025 test)0.745
PR-AUC0.295 vs 0.141 base rate
Brier score0.1096
Test set163,940 real CERT claims
Tests76 (pytest), CI green

These numbers are modest and honest — the public CERT feature set is thin by construction, and the model card says so.

Tech Stack

Next.js 16, FastAPI, XGBoost, DuckDB, Nebius/Groq LLM APIs, Pydantic

Future Work

Supabase persistence, EHR FHIR R4 integration, richer payer-policy features.

Links