Parallel Deep Learning Framework with Feature-Level Fusion for Early Sepsis Prediction in ICU Settings: Toward Real-Time Clinical Decision-Making

Abstract

Sepsis remains a major cause of ICU mortality, requiring accurate early prediction for clinical decision-making. Conventional deep learning (DL) architectures based on feature-level fusion, even when high-performing, are limited in capturing dynamic feature dependencies and may not ensure real-time applicability. We propose the Parallel Neural Network Fusion Predictor of Sepsis (PNNFPS), an ICU-oriented early prediction framework that employs two Deep Neural Networks (DNN1, DNN2) to process distinct clinical modalities in parallel, with their outputs fused and fed into an artificial neural network (ANN) to estimate sepsis risk. The approach is structured into two complementary phases: (1) offline development and optimization, using 5-fold cross-validation with class-imbalance handling, feature extraction, hyperparameter optimization, automated feature selection, and post-hoc explainability; and (2) online simulation, where the best offline model is deployed in a simulated ICU environment under four alarm policies (low, medium, high, multi-threshold) combined with an 8-hour silencing mechanism to reduce redundant alarms. The resulting alarm streams are evaluated through patient-wise detection, window-wise predictive behavior, alarm burden, alarm reduction, and lead-time analysis. Using the 2019 PhysioNet/Computing in Cardiology Challenge dataset, PNNFPS achieved a competitive offline U-score of 0.673, outperforming previously reported approaches. SHAP-based explainability confirmed clinical consistency, identifying Temperature, Heart Rate, HCO3, pH, WBC, Lactate, and Calcium as key predictive contributors. In online evaluation, the medium-threshold policy provided the most balanced operating point, preserving high patient-wise sensitivity (90.6%) with improved specificity (40.8%) and PPV (18.6%) over the low-threshold policy, an extended lead time of 89 hours before sepsis onset, and substantial alarm reduction (85.64%) via silencing. These findings indicate a favorable balance between patient safety, alarm reliability, and clinician workload, supporting PNNFPS for real-time ICU monitoring and clinical decision support.