Surgical MAchine learning model for predicting Risk of postoperative infecTions in general elective surgery (SMART):...

To apply a machine learning (ML) model that we developed and internally validated for predicting postoperative infection likelihood after elective general abdominal surgery, SMART, among 2716 patients.

The United Kingdom Health Data Research (UKHDR) Hub for Acute Care (PIONEER) supplied retrospective pseudonymised data for model training. These data contained demographic information, vital signs, microbiological investigations, comorbidities, surgical information, and infection diagnosis for elective general surgical patients (n = 2,716). Predictors were selected using an integrated approach of ML methods (feature elimination) and expert input. Recursive feature elimination with cross-validation was run on these predictors using Python(v3.8.2). Twelve algorithms were used, and an ensemble model with the three highest performing models was developed.

Nineteen predictors were selected to build the model, including demographics (e.g. age), comorbidities, microbiology data (e.g. multidrug-resistant infections), and laboratory investigation (CRP). The gradient-boosting classifier was found to be the best-performing model. The ensemble model showed high performance during training with 85.3% sensitivity, 74.6% specificity, and AUC=0.89, and during internal validation, with 96.9% sensitivity, 74.1% specificity, and AUC=0.86.

The ML model showed high performance in predicting postoperative infections in elective surgery. It used modifiable predicators that aided its clinical application. Identifying patients at higher risk of postoperative infections before surgery can promote early interventions and reduce antimicrobial resistance risk. External validation and testing are necessary for successful clinical implementation.