SaCal: An Efficient Saliency-Guided Causal Framework for Interpretable Healthcare Analytics

Multi-modal Electronic Health Records (EHRs) enable comprehensive patient assessment and facilitate multi-task prediction in healthcare. Current relevant studies mainly focus on three aspects: multi-modal learning (MML), multi-task learning (MTL), and multi-modal multi-task learning (MM-MTL). MML approaches attempt to extract latent information for multimodal fusion, but the process remains opaque and provides limited flexibility of adjustment over modality contributions, with potential underemphasis of task-relevant modalities. MTL methods typically feed features directly into models, without taking the spurious correlations between features and tasks into account, which hampers performance. While existing MM-MTL methods combine MML and MTL, they did not address the problems of obscure multi-modal fusion and multi-task learning with spurious correlations. In this paper, we propose a saliency-guided causal framework named SaCal for interpretable MM-MTL healthcare analytics. Specifically, a saliency-based mechanism that can quantify modality importance and adjust modality proportion accordingly is proposed to guide the modality fusion. With the aim of separating spurious correlations from truly causal features for MTL, we propose an innovative Task-to-Module Graph Learning structure based on causal theory for improved prediction. Considering the high dimensionality and redundancy in multi-modal data, we design the de-correlation regularization to ensure that the learned representations are diversified and discriminative. Moreover, an adaptive activation and graph sparsity strategy is designed to enhance overall model efficiency and reduce computational overhead. SaCal is capable of offering saliency-driven interpretation for different medical prediction tasks. Experimental results on four MIMIC series clinical databases and five healthcare prediction tasks confirm SaCal's effectiveness in improving predictive performance compared to seven state-of-the-art baselines, also demonstrating its efficiency and interpretability.