Time-Series Deep Learning and Conformal Prediction for Improved Sepsis Diagnosis in Non-ICU Hospitalized Patients.

Purpose: Sepsis, a life-threatening condition from an uncontrolled immune response to infection, is a leading cause of in-hospital mortality. Early detection is crucial, yet traditional diagnostic methods, like SIRS and SOFA, often fail to identify sepsis in non-ICU settings where monitoring is less frequent. Recent machine learning (ML) models offer new possibilities but lack generalizability and suffer from high false alarm rates. Methods: We developed a deep learning (DL) model tailored for non-ICU environments, using MIMIC-IV data with a conformal prediction framework to handle uncertainty. The model was trained on 83,813 patients and validated with the eICU-CRD dataset to test performance across hospital settings. Results: Our model predicted sepsis at 24, 12, and 6 h before onset, achieving AUROCs of 0.96, 0.98, and 0.99, respectively. The conformal approach reduced false positives and improved specificity. External validation confirmed similar performance, with a 57% reduction in false alarms at the 6 h window, supporting practical use in low-monitoring environments. Conclusions: This DL-based model enables accurate, early sepsis prediction with minimal data, addressing key clinical challenges and potentially improving resource allocation in hospital settings by reducing unnecessary ICU admissions and enhancing timely interventions.