Farah E. Shamout

MedGuards achieves significant improvements in medical error detection and correction by leveraging a multi-agent system that enhances interpretability and robustness without retraining LLMs.

Data-driven priors, leveraging cross-modal similarity and modality-specific corruptions, can substantially improve both the accuracy and reliability of uncertainty estimates in multimodal clinical risk prediction.

Selective prediction, a proposed safeguard for AI in clinical settings, can backfire dramatically due to class-dependent miscalibration, leading to worse performance than simply trusting the model.