TimeSHAP-LSTM: An Interpretable Temporal Deep Learning Framework for Early Sepsis Prediction Using EHR Data

Early detection of sepsis remains a major clinical challenge due to the complex and dynamic nature of physiological signals recorded in Intensive Care Units (ICUs). This study proposes an interpretable temporal deep learning framework, TimeSHAP-LSTM, designed to predict sepsis onset using Electronic Health Record (EHR) time-series data while providing clinically meaningful explanations. The model combines a Long Short-Term Memory (LSTM) architecture with the TimeSHAP interpretability algorithm to quantify both feature-level and time-dependent contributions to the prediction outcome. Experiments were conducted using the MIMIC-IV database, comprising more than 60,000 ICU admissions. The proposed model achieved competitive predictive performance (AUROC = 0.87, AUPRC = 0.67), outperforming conventional black-box approaches while maintaining transparency. Temporal interpretability analysis revealed that physiological features such as heart rate, lactate, and respiratory rate significantly contributed to predictions, particularly within 6–12 hours before sepsis onset. A clinician evaluation involving 12 ICU physicians confirmed that TimeSHAP visualizations were clear, relevant, and trustworthy, suggesting strong potential for clinical adoption. These findings demonstrate that interpretable artificial intelligence can achieve state-of-the-art accuracy while offering time-resolved insights that align with medical reasoning. The proposed framework contributes toward building reliable, explainable, and ethically deployable AI systems for critical care environments.