RealICU (RAG) - Action Recommendation: leaderboard
RealICU (Shen et al., 2026) replays MIMIC-IV intensive-care stays in 30-minute windows and scores an LLM against hindsight labels from senior ICU physicians on the 930 windows of RealICU-Gold. Action Recommendation: Hit@5 (0-1), the share of windows whose top five recommended actions include one the physicians labelled, higher is better. RAG protocol: the model sees the current window and the five most relevant windows of the stay so far, retrieved with PubMedBERT embeddings.
Metric: Hit@5 (0-1; share of windows whose top-5 recommended actions include a physician-labelled action, PubMedBERT semantic matching; RAG strategy, the top-5 prior windows retrieved with PubMedBERT). Source: arxiv.org. Saturation forecast: Around April 2028. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.5 | #91 |
| 2 | GPT-5.4 | 0.48 | #121 |
| 3 | Qwen 3 235B A22B | 0.45 | #622 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=realicu-rag-action-recommendation · How It Works · Data refreshed daily, snapshot 2026-11-12.