SMILE: Self-Explainable Multimodal Information Bottleneck for Medical Diagnosis

In September 2026, researchers proposed a self-explaining multimodal information bottleneck method called SMILE, aimed at addressing the problem of interpretability in medical diagnosis. This method constructs a unified learning paradigm within the framework of information bottleneck, by identifying the information elements that contribute most to diagnosis across different modalities, and jointly optimizing both prediction performance and modality-specific interpretability. The study utilized Renyi’s α-order entropy function based on matrices for stable optimization, and conducted extensive experiments on representative medical datasets covering various modalities. Results showed that this method increased the absolute accuracy on the iCTCF dataset by 9.1 percentage points, while providing transparent and modality-aware insights into feature correlations, significantly enhancing the interpretability and generalization ability of diagnosis.