ABSTRACT

Sepsis is a life-threatening dysregulated host response to infection, and each hour of treatment delay increases mortality risk. While electronic health record (EHR)–only models can flag risk based on vitals, labs, and clinical notes, radiographic evidence of evolving infection (e.g., chest X-ray opacities) is often available but under-utilized in automated screening. We present a deployment-minded multimodal fusion approach that integrates medical imaging (CXR) with time-ordered EHR streams using a CNN–RNN hybrid. A convolutional encoder (DenseNet-121) extracts visual embeddings from time-stamped chest radiographs, while stacked LSTMs summarize multivariate vitals/labs; cross-modal attention fuses the two under irregular sampling and missing-image conditions. We train with Sepsis-3 labels (SOFA-based onset) and address class imbalance via focal loss and cost-sensitive sampling; calibration uses temperature scaling. In a retrospective cohort assembled by time-aligning MIMIC-III/IV EHR with MIMIC-CXR studies, the fused model improves AUROC (EHR-only 0.83 ? fusion 0.87), AUPRC (0.39 ? 0.47), and lead time to 6-hour-ahead onset (median +48 minutes) compared with strong EHR baselines; gains are largest in pneumonia-associated sepsis and immunocompromised subgroups. An ablation shows attention-based mid-level fusion is more robust to missing radiographs than early/late strategies. We discuss integration via SMART-on-FHIR, drift monitoring, and bias audits. Limitations include retrospective single-system training and incomplete imaging availability. The results suggest that EHR+imaging fusion can enhance early sepsis prediction while remaining compatible with clinical workflows.