As AI systems become increasingly integrated into clinical practice, a number of significant challenges remain. This thesis focuses on two critical challenges: i) the distinctive characteristics and complexities inherent to clinical data, and (ii) the opacity of deep neural networks.
The first part of this work explores the integration of multimodal, sparse and heterogeneous clinical data sources. Using two cardiovascular disease use-cases, this part demonstrates how data sources selection (e.g., administrative versus clinical modalities) significantly influences downstream dataset characteristics. A notable contribution is the development of datasets with electrocardiograms (ECG) labeled with diagnostic derived from multiple clinical modalities according to established clinical guidelines. Based on this analysis of clinical data characteristics which highlight labeled data sparsity, two strategies are investigated to address this challenge for ECG modeling: (i) self-supervised learning, and (ii) a lightweight, interpretable classification model based on multivariate autoregressive coefficients.
Part II focuses on explainability for clinical applications, emphasizing its value for clinical safety, AI adoption, and scientific insight. This part includes a review of recent explainability methods and discusses the challenge of assessing explanation quality. To address this, a novel evaluation framework for post-hoc explainability methods is introduced, identifying Shapley Value Sampling as the most faithful interpretability method across various architectures and datasets. The last contribution of this thesis is a novel, self-explainable and scalable architecture leveraging frozen visual foundation models for image classification tasks. This architecture optimizes classification accuracy and interpretability simultaneously, matching or exceeding state-of-the-art interpretable models in performance while maintaining the number of trainable parameters similar to standard linear probing techniques. Critically, it facilitates systematic model auditing across large datasets, revealing spurious or unintended classifier behaviors in both general and clinical applications.
Collectively, the contributions presented in this work support the current efforts to develop clinically relevant and safe AI systems, enhancing trustworthiness and facilitating their adoption in healthcare.
