Development and Validation of ‘Patient Optimizer’ (POP) Algorithms for Predicting Surgical Risk with Machine Learning

Background Pre-operative risk assessment can help clinicians prepare patients for surgery, reducing the risk of perioperative complications, length of hospital stay, readmission and mortality. Further, it can facilitate collaborative decision-making and operational planning. Objective To develop effective pre-operative risk assessment algorithms (referred to as Patient Optimizer or POP) using Machine Learning (ML) that predicts the development of post-operative complications and provides pilot data to inform the design of a larger prospective study. Methods After institutional ethics approval, we developed a baseline model that encapsulates the standard manual approach of combining patient-risk and procedure-risk. In an automated process, additional variables were included and tested with 10-fold cross-validation, and the best performing features were selected. The models were evaluated and confidence intervals calculated using bootstrapping. Clinical expertise was used to restrict the cardinality of categorical variables (e.g. pathology results) by including the most clinically relevant values. The models were created with extreme gradient-boosted trees using XGBoost [1]. We evaluated performance using the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Data was obtained from a metropolitan university teaching hospital from January 2015 to July 2020. Data collection was restricted to adult patients undergoing elective surgery. Results A total of 11,475 adult admissions were included. For predicting the risk of any postoperative complication, kidney failure and length-of-stay (LOS), POP achieved an AUROC (95%CI) of 0.755 (0.744, 0.767), 0.869 (0.846, 0.891) and 0.841 (0.833, 0.847) respectively and AUPRC of 0.651 (0.632, 0.669), 0.326 (0.293, 0.359) and 0.741 (0.729, 0.753) respectively. For 30-day readmission and in-patient mortality, POP achieved an AUROC (95%CI) of 0.61 (0.587, 0.635) and 0.866 (0.777, 0.943) respectively and AUPRC of 0.116 (0.104, 0.132) and 0.031 (0.015, 0.072) respectively. Conclusion The POP algorithms effectively predicted any post-operative complications, kidney failure and LOS in the sample population. A larger study is justified to improve the algorithm to better predict complications and length of hospital stay. A larger dataset may also improve the prediction of additional specific complications, readmissions and mortality.