BenchToward

Medical AI, evaluated on real-world clinical data

Evaluated once, reused everywhere. Open evaluations of medical AI on Korean clinical datasets released by Seoul National University Hospital. Every result ships with scores, item level predictions, and exact model versions, ready for researchers to cite and for agents to query.

8 benchmarks · 21 models · 42 results · 38 reuses

BUILDING · benchmark builder Building Intraoperative hypotension, the authors' unbiased dataset Build Watch →

For AI agents: curl -s https://benchtoward.com/AGENTS.md · more

Clinical text

Korean clinical knowledge (KMLE-level MCQ)

ClinicalQA: Korean Medical Knowledge QA Benchmark · accuracy

  1. 1GLM-5.3 753B-A40B0.909
  2. 2Gemini 3.8 Flash0.908
  3. 3Qwen3.8 Max 2.4T-A95B0.906

Reference: Majority class (always B) 0.343

+ 16 more models

Updated 31h ago
Korean surgical robot voice commands

Korean Surgical Robot Voice-Command Benchmark · exact_match

  1. 1Qwen3.8 Max 2.4T-A95B0.947
  2. 2DeepSeek V4.1 Flash 552B-A16B0.929
  3. 3Qwen3.8 Flash Next 180B-A6B0.921

Reference: Prior (ignores the input) 0.133

+ 1 more model

Updated 43h ago

Biosignal

ICU cardiac arrest within 0.5-24 h from a 5-min ECG epoch

SCOPE: ICU ECG and PPG before cardiac arrest, death or discharge · auroc

  1. 1LightGBM on 33 HRV measures (SCOPE authors' training script)0.826

Reference: Prior (ignores the input) 0.500

Updated 32h ago
Abnormal rhythm in 5-second intraoperative ECG (lead II)

VitalDB Arrhythmia Database · auroc

  1. 1Gradient boosting on RR-interval features0.961

Reference: Prior (ignores the input) 0.500

Updated 2d ago
Intraoperative hypotension 5 min ahead (numeric MAP)

VitalDB · auroc

  1. 1Gradient boosting on window summary features0.910
  2. 2Logistic regression on last-minute MAP0.852
Updated 2d ago

Speech

Korean operating-room speech recognition

Synthetic K-OR Speech Audio · cer

  1. 1Qwen3-ASR 1.7B0.073
  2. 2Nemotron 3.5 ASR 0.6B0.136
Updated 2d ago

Structured EHR

30-day postoperative in-hospital mortality

INSPIRE · auroc

  1. 1Logistic regression, mean imputation (INSPIRE authors' script)0.936
  2. 2XGBoost (INSPIRE authors' script)0.924
  3. 3ASA class alone (INSPIRE authors' baseline)0.818
Updated 2d ago
Synthetic-SNUH Text-to-SQL (OMOP CDM, Korean and English)

Synthetic-SNUH v4 · rs0

  1. 1DeepSeek V4.1 Flash 552B-A16B0.663
  2. 2gpt-oss 20B0.223

Reference: Prior (ignores the input) 0.178

Reference: Constant query (SELECT 0 for every question) 0.000

Updated 18h ago
Text-to-SQL over OMOP CDM Coming soon

Synthetic-SNUH v4 (10,000 synthetic patients)

Recently evaluated