Trustworthy Learning in Medical AI

Clinical AI should meet trustworthiness demands now set out in the GDPR and the EU AI Act: a deployed system should explain its decisions, indicate when those decisions are unreliable, avoid unnecessary centralization of patient data, and permit the removal of an individual’s data influence on request. These demands are usually addressed as an afterthought, recovered from a model once it is already trained, through explanation, repeated inference, or retraining applied after the fact. This thesis investigates a different commitment: that trustworthiness, operationally framed by these four demands, is built into the structure of the system from the start. Three chapters develop the argument. Knowing applies the commitment to interpretability through uncertainty-aware risk classification and gaze-driven prompting of a frozen segmentation model. Doubting applies it to uncertainty through federated learning where uncertainty-filtered attention maps replace parameter aggregation as the unit of inter-client exchange. Forgetting applies it to unlearning through prompt-removal and passport-embedded LoRA adapters with authority-mediated audit. Across this work, where two demands are met through a shared structural commitment, the combination yields value that neither delivers separately.