Clinical Reasoning and Information-Aware Multimodal Representation Learning for Biochemical Recurrence Prediction

Recent advances in multimodal foundation models have demonstrated the potential to integrate cross-modal information across cellular, organ-level, and clinical domains for precision medicine. In prostate cancer, accurate prediction of biochemical recurrence (BCR) is critical for enabling timely intervention, informing treatment decisions, and improving patient risk stratification. However, effectively integrating heterogeneous clinical modalities remains challenging because of modality imbalance and the difficulty of preserving complementary information across representations. In this work, we propose a unified multimodal framework for BCR prediction that integrates multi-parametric magnetic resonance imaging (mpMRI), hematoxylin and eosin (H&E) histopathology images, and structured clinical information. To enhance semantic understanding of patient-level clinical context, we introduce an expert-guided clinical reasoning module that transforms structured clinical variables into enriched textual representations through structured prompting before multimodal prediction. This enables clinical information to be represented in a more semantically meaningful and reasoning-oriented form for integration with visual modalities. Furthermore, we introduce a Mutual Information Neural Estimation (MINE)-based learning strategy to maximize the mutual information within individual modalities and derive an information-aware prior for enhancing their feature representations before cross-modal integration. This prior promotes more informative modality-specific representations, allowing the subsequent fusion process to better leverage the information encoded in each modality. Experiments demonstrate that the proposed framework achieves a C-index of 0.9366 and consistently outperforms unimodal and bimodal baselines, with improvements ranging from 2\% to 10\% across different modality combinations. These results demonstrate the value of combining clinical reasoning with information-aware multimodal representation learning for robust biochemical recurrence prediction, highlighting the potential of heterogeneous clinical and imaging data integration for clinically relevant risk stratification.