We tuned a compact clinical forecaster from GPT-OSS-120B to predict clinical events from raw MIMIC-III notes, using later patient records to resolve the outcomes. The paper reports a 27% Brier Skill Score, about 70% lower calibration error than the base model, and a slightly better Brier score than GPT-5.
What we did
- Converted time-ordered MIMIC-III notes into questions about future clinical events.
- Resolved labels from what appeared later in the patient record, without a hand-labeled dataset.
- Tuned from GPT-OSS-120B on the generated clinical prediction set and benchmarked against the base model and GPT-5.
Read more
Primary artifacts for this case study.