200,000-trajectory ICU medication-dosing & sepsis-triage safety benchmark on de-identified MIMIC-IV + open-weight SFT/RLVR clinical abstention guardrails.
Summary
Frontier language models are rapidly being adopted by clinicians at the bedside, yet in high-stakes intensive care unit (ICU) and perioperative medicine, silent reasoning failures kill: models routinely fail to adjust vasopressor, opioid, sedative, or anticoagulant doses under acute renal/hepatic dysfunction, misread longitudinal acid-base/hemodynamic trajectories, and confidently guess instead of abstaining when critical laboratory values are absent.
Led by Dr. Furkan Bas, MD (Specialist Physician in Anesthesiology & Intensive Care, Hacettepe University Faculty of Medicine; ORCID: 0009-0002-2787-2709), this 100% open-source (Apache-2.0) technical AI safety project builds:
1. 200,000+ Trajectory ICU Safety & Abstention Benchmark: Built exclusively on credentialed, fully de-identified public critical care cohorts (PhysioNet MIMIC-IV v3.1 and eICU-CRD) plus synthetic ASA/ESAIC perioperative crisis scenarios (zero identifiable patient PHI).
2. Cross-Frontier Zero-Data-Retention Safety Audit: Systematic evaluation of frontier reasoning models on dose-error sensitivity, calibration, and sycophancy under time-pressured clinical prompts.
3. Open-Weight Clinical Guardrail & Abstention Models: Multi-node H100 SFT + DPO/RLVR alignment of open-weight 8B-70B clinical verification models paired with a deterministic pharmacokinetic rule-verifier.
Budget Breakdown ($50,000 Target / $10,000 Minimum)
- $30,000: Zero-Data-Retention Frontier LLM API evaluation across 200,000+ longitudinal multi-turn ICU stays and adversarial counterfactuals.
- $20,000: Dedicated NVIDIA H100 cluster compute for SFT + RLVR training of open-weight clinical abstention & dose-verification guardrails, containerization, and public leaderboard hosting (0% overhead / 100% direct compute & API spend).
All evaluation harnesses, benchmark datasets, and guardrail weights are released publicly under Apache-2.0.
There are no bids on this project.
Comments
No comments yet. Sign in to create one!