Explainable ICU Mortality Prediction with Temporal Attention and Clinically Constrained Feature Attributions

Abstract

The rapid digitization of healthcare infrastructure has resulted in the proliferation of Electronic Health Records (EHRs), providing a fertile ground for data-driven clinical decision support systems. Among the most critical applications is the prediction of mortality in Intensive Care Units (ICUs), where early identification of deteriorating patients can significantly influence survival outcomes. While Deep Learning (DL) models, particularly Recurrent Neural Networks (RNNs) and Transformers, have demonstrated superior predictive performance compared to traditional scoring systems, their deployment is frequently hindered by a lack of interpretability. This paper introduces a novel architecture that integrates Temporal Attention mechanisms with Clinically Constrained Feature Attributions to predict ICU mortality. Unlike standard interpretability methods that provide post-hoc explanations, our approach incorporates domain knowledge directly into the training process via a regularization term that penalizes physiologically implausible feature associations. We evaluate our model on the MIMIC-III dataset, demonstrating that it achieves state-of-the-art predictive performance while generating explanations that align with clinical consensus. The results indicate that enforcing clinical constraints does not degrade accuracy; rather, it improves the model's robustness and trustworthiness, facilitating safer integration into clinical workflows.