Derivation and validation of heart rate variability based Machine learning prognostic models for patients with suspected sepsis

Objective: To develop and compare heart rate variability (HRV) based machine learning algorithms for septic shock and sepsis prediction against established metrics validated on the Medical Information Mart for Intensive Care III (MIMIC-III) database for early detection of septic patient deterioration. Methods: We developed HRV-based prediction models in a prospective cohort study in an emergency department (ED). Adult patients with a clinical diagnosis of infection and suspected sepsis were recruited in our ED between April 2016 and September 2017. Different machine learning algorithms were trained to continuously predict septic shock onset 120 min ahead on the included patients using HRV obtained in five-minute intervals. We further applied transfer learning among patients with Sepsis-3 in the MIMIC- III database. Results: We recruited and included 38 and 755 patients with a total recording time of 360 and 19,321 patient hours from our ED and the MIMIC-III database, respectively. The extreme gradient boosting (XGBoost) model outperformed other machine learning models and conventional models including the delta SOFA score, National Early Warning Score (NEWS), and Modified Early Warning Score (MEWS) in predicting septic shock (AUROC: 0.94; 95% CI: 0.91–0.96) during internal validation. After transfer learning, the XGBoost model had an AUROC of 0.91 (95% CI: 0.90–0.91) to predict septic shock in patients with Sepsis-3 in the MIMIC-III database. Conclusions: HRV-based XGBoost models provide a promising non-invasive tool for early septic shock prediction and patient deterioration due to infection, potentially enhancing patient management strategies.