MedBayes-Lite: A Clinical Uncertainty Governance Layer for Risk-Aware Medical Decision Support

MedBayes-Lite is a newly introduced uncertainty governance layer designed for transformer-based clinical predictors, which enhances the reliability of clinical language models by integrating Monte Carlo dropout, predictive calibration, and confidence-guided abstention, all without adding trainable parameters. Evaluated on MedMCQA and MedQA-USMLE datasets, it significantly reduces expected calibration error by 0.23 to 0.33 and nearly eliminates high-severity overconfident errors, decreasing them from approximately 21% to near zero. This framework provides practitioners with a practical solution to improve decision-making in high-stakes clinical environments, ensuring that low-confidence predictions are deferred for human review, thereby reducing the risk of harmful errors.