Holtercare-Bench - Event Timing (Text): leaderboard
Holtercare-Bench (2026) asks multimodal models closed multiple-choice questions about long-term dynamic ECG (Holter) recordings of 13 to 24 hours from real hospital cases. The event timing task asks for the millisecond timestamps of every occurrence of an abnormal event, among close distractors; the ECG lead is given as text values in millivolts. Score is accuracy in percent; higher is better.
Metric: Accuracy (%; multiple-choice closed QA on the millisecond timestamps of every occurrence of an abnormal event, among close distractors; the ECG lead given as text, values in millivolts to three decimals, truncated where it exceeds the input limit; zero-shot on the held-out test cases of Holtercare-23K, 20 percent of 788 clinical Holter recordings of 13-24 hours). Source: arxiv.org. Saturation forecast: Around April 2028. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 62.5 | #164 |
| 2 | GPT-5 Mini | 33.54 | #348 |
| 3 | Phi-4 Mini Instruct | 33.23 | #6419 |
| 4 | Claude Haiku 4.5 | 31.71 | #568 |
| 5 | MedGemma 1.5 4B | 30.79 | #2562 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=holtercare-bench-event-timing-text · How It Works · Data refreshed daily, snapshot 2026-11-13.