A machine learning model for post-operative sepsis prediction in acute surgical patients: a multi-centre, prospective study

The aim of this study was to develop and validate a machine learning (ML) solution for predicting sepsis among elderly surgical patients undergoing emergency procedure. Data from over 150 variables were collected across multiple domains. ML models were cross-validated using a nested-cross validation approach. The performance of individual models was assessed based on accuracy, sensitivity, and specificity. A total of 29 medical centres participated in the study, ensuring a diverse and representative sample. The study included patients undergoing emergency procedures across various surgical specialties, encompassing abdominal, thoracic, vascular, gynaecological, and urological surgeries, performed by both general and emergency surgeons in general or trauma surgery settings. Among 2571 enrolled patients, 119 were identified as having sepsis. The Random Forest model demonstrated the highest accuracy of 96.2%, with notable sensitivity and specificity. An ensemble model further improved performance, achieving an accuracy of 96.06%. ML models, show promise in accurately predicting sepsis among elderly surgical patients in emergency settings. These findings underscore the potential of ML in enhancing risk stratification and informing clinical decision-making to improve patient outcomes and post-surgery rehabilitation. Further research and validation studies are warranted to evaluate the real-world applicability and integration of these predictive models into clinical practice.