<![CDATA[Purpose: Early detection of sepsis can reduce mortality by enabling timely antibiotics, fluids, and escalation of care, yet prediction is challenging because clinical deterioration is heterogeneous and data are fragmented across electronic health records (EHR) and bedside monitoring. Methods: We present a comparative study of multi-modal machine learning approaches that integrate EHR time-series (labs, medications, diagnoses, and nurse-charted vitals) with high-frequency vital-sign waveforms (heart rate, respiratory rate, and oxygen saturation). Using a de-identified, ICU-like synthetic cohort, we define sepsis onset by a proxy clinical rule and evaluate prediction at 3–12 hours before onset. We compare classic baselines (logistic regression, random forest, XGBoost) to deep temporal models (LSTM over EHR sequences, convolutional encoders for waveforms) and a late-fusion Transformer that combines modality-specific embeddings. Results: Across horizons, the late-fusion Transformer achieves the best discrimination and precision-recall, improving AUROC by 0.02–0.04 and AUPRC by 0.04–0.08 over the strongest single-modality model. Calibration improves with temperature scaling and modality dropout. Explainability analyses using attention rollout and SHAP suggest that rising lactate, increasing vasopressor dose, and tachycardia patterns contribute most to early warnings. Novelty/Value: This paper provides a structured, reproducible comparison of multi-modal architectures for early sepsis prediction, with practical guidance on data alignment, missingness, and evaluation under clinically meaningful horizons, which can inform safer deployment and governance of predictive models in hospitals.]]>
Copyrights © 2026
