Interpretable prognosis prediction system (IPPS) framework: the disease experts learning through health records

Abstract

In the medical engineering domain, accurate mortality risk prediction from electronic health records (EHRs) is critical for early warning and timely intervention. However, with the disease heterogeneity, content sparsity, and limited interpretability, it is always uneasy to use health data effectively. In the paper, we propose an Interpretable Prognosis Prediction System (IPPS) framework with a disease expert model to predict mortality of patients, through the integration of core components: (1) similarity-guided input enhancement, (2) disease-specific modeling, and (3) multi-task optimization via auxiliary disease supervision. Regarding data sparsity issue, IPPS employs a patient similarity graph based on disease co-occurrence patterns to retrieve top- k similar patient trajectories. A query attention mechanism dynamically integrates these auxiliary sequences using data of target patient, and produces context-enriched inputs for downstream modeling. For disease heterogeneity, IPPS introduces a modular expert architecture. Each expert specializes in a distinct disease group to capture group-specific progression patterns. Each expert adopts a plug-and-play sequential backbone SeqModel for temporal feature extraction and outputs disease-related signals, which are aggregated to support the primary mortality prediction task. If Mamba is deployed, the IPPS-Mamba model is created. Finally, for prognosis interpretability and prediction, IPPS runs a joint multi-task training strategy to serve as auxiliary supervision for reinforcing mortality estimation through disease-aware signals. Upon conducting experiments on MIMIC-III/IV datasets, IPPS-Mamba model offers outstanding prediction accuracy and interpretability performance. This shows that IPPS framework offers transparent and performant disease-aware prognosis prediction.