EHRBench - MIMIC-III: leaderboard

EHRBench (2026) tests LLMs on clinical decision questions generated from real patient records and verified against biomedical knowledge bases, posed as multiple-choice questions with 4 to 6 options. This board is accuracy in percent on the questions built from MIMIC-III records; higher is better.

Metric: Accuracy (%) on the questions built from MIMIC-III records, EHR-grounded multiple-choice clinical decision questions (4, 5 and 6 options) built from MIMIC-III, MIMIC-IV and PROMOTE encounter trajectories with knowledge-base verification, JSON-constrained answers, deterministic decoding; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 31 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.2 Instant71.5#304
2GPT-4.170.59#505
3GPT-5 Chat70.16#449
4GPT-5 Mini69.08#364
5Llama 3.3 70B Instruct68.74#1385
6GPT-4.1 Mini68.58#789
7Qwen 3 32B68.48#972
8Mistral Small 367#1796
9GLM-4 32B (0414)66.27#1505
10EHRBench llama3-70B (checkpoint unspecified)65.39
11Qwen 2.5 32B65.22#1344
12m1-32B-1k62.87#1240
13Qwen 3 8B62.26#1873
14Qwen 3 4B62.01#2348
15doctor-r1-8B61.94#1624

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=ehrbench-mimic-iii · How It Works · Data refreshed daily, snapshot 2026-11-21.