MedJev study reports calibration gains for laptop-based medical predictions

Researchers report in a preprint that MedJev, an open medical decision model that runs locally on a laptop, can adapt predictions to local patient populations using 50 to 200 labeled examples. Their two-parameter recalibration method reduced expected calibration error to 0.007. MedJev answers structured questions in a single forward pass without generating text. Across ten medical benchmarks, its in-hospital mortality prediction achieved an AUROC of 0.926, exceeding XGBoost's 0.902. The researchers also released their code and evaluation harness. Performance varied by task: MedJev underperformed supervised models on emergency-department triage admission, scoring 0.745 versus 0.809. Patient-level predictions also showed substantial miscalibration before adjustment, with in-hospital mortality expected calibration error of 0.286.