Lightweight, Explainable, and Efficiency-Aware Deep Learning for Early Sepsis Prediction from Structured Electronic Health Record Time-Series

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Early detection of sepsis in Intensive Care Units (ICUs) is critical, as delayed intervention significantly increases mortality risk. This study investigates early sepsis prediction using multivariate physiological time-series data from the PhysioNet 2019 dataset, comprising 40 structured Electronic Health Record (EHR) variables recorded hourly. We evaluate lightweight deep learning models, including 1D-CNN, GRU, TCN, and a lightweight CNN–ViT hybrid, under a unified protocol targeting prediction up to 6 h before sepsis onset. This prediction horizon follows the PhysioNet 2019 benchmark and represents a clinically meaningful early-warning window that balances actionable intervention time with sufficient predictive signal. The results show that effective temporal modeling is key for early detection. The TCN achieves the most favorable balance between predictive performance and efficiency, with the highest recall (0.778), tied highest F1-score (0.495), highest AUPRC (0.50), and highest AUC (0.907 ± 0.010). It also records the lowest inference latency (0.62 s total over the test set under a CPU-only setting), supporting real-time applicability. In contrast, CNN–ViT achieves competitive performance (AUC 0.906 ± 0.011) but incurs substantially higher computational cost (44.34 s). Explainability using SHAP highlights clinically plausible indicators, including blood pressure, heart rate, oxygen saturation, and lactate. Overall, lightweight temporal models offer a practical and clinically meaningful solution for early sepsis prediction under real-time constraints.

This publication has 13 references indexed in Scilit: