Access to MERIT requires PhysioNet credentialed access to MIMIC-IV
MERIT was trained on MIMIC-IV. Requests are reviewed manually and granted only to credentialed PhysioNet users who have signed the MIMIC-IV data use agreement.
Log in or Sign Up to review the conditions and access this model content.
MERIT: risk-aware ICD-10 coding with evidence
MERIT reads an inpatient discharge summary and returns ICD-10-CM/PCS codes, a verbatim evidence quote for every diagnosis that maps to a CMS-HCC V28 payment HCC, the HCCs after hierarchies, the disease risk-adjustment factor (RAF), and HCC candidates to route to a human coder with a conformal guarantee on missed HCCs.
This repository is the system: shortlist head, PLM-ICD classifier, fusion rules, per-HCC decoders, conformal thresholds, public CMS V28 and CDC tables, the inference pipeline, an HTTP API and Docker kit, and the research report draft (public page, PDF). The coder language models are in MERIT-Coder-4B-GRPO (default) and MERIT-Coder-4B-CRISP.
Author: Nagur Shareef Shaik, Georgia State University.
How it works
- A frozen Qwen3.5-4B encodes the note; a label-wise attention head proposes up to 96 candidate codes.
- The coder (Qwen3.5-4B, LoRA SFT + GRPO) writes one
CODE | "quote"line per code. - PLM-ICD (biomedical RoBERTa, 8,192-token input) scores all 7,940 codes; a validation-tuned rule fuses both.
- Calibrated per-HCC thresholds decide the HCCs; V28 hierarchies, interactions and coefficients give the RAF.
- Conformal risk control lists extra HCC candidates so that at most 10% of HCCs are missed in expectation.
Results (MIMIC-IV ICD-10, Edin et al. test split, 19,799 notes)
| System | Micro-F1 | Macro-F1 | HCC-F1 | RAF MAE | Missed HCC | Unsupported HCC | Verbatim quotes |
|---|---|---|---|---|---|---|---|
| PLM-ICD 8k (our reproduction) | 59.9 | 23.3 | 77.1 | 0.368 | 26.0% | 19.5% | n/a |
| MERIT, GRPO coder (default) | 60.2 | 24.2 | 79.3 | 0.335 | 22.6% | 18.6% | 99.2% |
| MERIT, CRISP coder | 59.9 | 24.3 | 79.1 | 0.340 | 22.8% | 19.0% | 99.0% |
| Best published (Wang et al. 2026) | 59.6 | 23.7 | n/a | n/a | n/a | n/a | n/a |
Paired against PLM-ICD 8k on the same notes, the default system adds +2.23 [2.02, 2.43] HCC-F1 and lowers RAF error by 0.033, with 3.37 points fewer missed HCCs (95% bootstrap intervals). Conformal routing holds the missed-HCC rate at 10.1% while sending 2.7 codes per note for review (ฮฑ = 0.10), and at 5.1% with 9.2 codes per note (ฮฑ = 0.05). Full results, ablations and limitations are in the report.
Intended use
- Research on automated ICD-10 coding and CMS-HCC risk adjustment, by credentialed MIMIC-IV users.
- Decision support only: every code, HCC and RAF value must be reviewed by a qualified coder.
Out of scope: autonomous coding, billing, payment or audit decisions; use on notes without a lawful basis for processing; any attempt to recover training data. MERIT was trained on inpatient discharge summaries from one US academic centre (MIMIC-IV) and has not been validated elsewhere.
Self-hosting
Requires one GPU with 40 GB or more (tested on an A100-80GB) and a Hugging Face token with granted access.
huggingface-cli download nagur-shareef-shaik/MERIT --local-dir MERIT
cd MERIT && docker build -t merit -f merit_serve/Dockerfile .
docker run --gpus all -p 8000:8000 -e HF_TOKEN=$HF_TOKEN -e MERIT_API_KEY=change-me merit
curl -s localhost:8000/v1/code -H "Authorization: Bearer change-me" -H "Content-Type: application/json" \
-d '{"note": "Discharge Diagnoses: ...", "age": 78, "sex": "F"}'
Set MERIT_CODER=crisp and MERIT_CODER_MODEL=nagur-shareef-shaik/MERIT-Coder-4B-CRISP for the CRISP coder.
Without Docker: pip install vllm==0.32.0 -r merit_serve/requirements.txt and run
uvicorn merit_serve.server:app from the repository root. Notes are processed in memory and never stored or logged.
Data and privacy
Trained on MIMIC-IV 3.1 codes and MIMIC-IV-Note 2.2 discharge summaries under the PhysioNet credentialed data use agreement. This repository contains no note text and no per-note outputs. A memorisation audit (200 training and 200 test notes, greedy continuation of the first 256 tokens) found no difference between training and test notes beyond shared note templates (see the coder cards). The CMS V28 and CDC ICD-10-CM tables are public US government files.
Citation
@techreport{shaik2026merit,
title = {MERIT: Risk-Aware ICD-10 Coding with Evidence-Citing Language Models and Code-Level Credit Assignment},
author = {Shaik, Nagur Shareef},
institution = {Georgia State University},
type = {Research Report Draft},
number = {Risk Adjustment Research Series, Report 1},
year = {2026}
}