Sano Centre for Computational Medicine

Applying machine learning and artificial intelligence in medicine faces specific obstacles: medical data are hard and costly to obtain, limited in volume, often incomplete, and sensitive enough to require strict privacy protection, while clinicians need to understand and verify why a model reached a decision. This talk presents interpretable and privacy-preserving methods for medical data analysis that offer concise, understandable explanations as an alternative or complement to black-box models. Topics include learning from small medical datasets with abstractions that support rule induction, rule-based approaches to federated learning, a rule-based reasoning mechanism for selectively acquiring missing information relevant to a decision, and rule-based representations of knowledge extracted from non-tabular data modalities. The common thread is making effective use of limited, sensitive and fragmented medical data by identifying relationships within it and presenting them as clear, actionable insights for decision-making.