Prediction of postoperative infections by strategic data imputation and explainable machine learning

Abstract Infections following surgery drive patient morbidity and mortality, making early detection essential. Traditional predictive models utilize preoperative patient characteristics and features of the surgical procedure. This study enhances these baseline determinants by incorporating postoperative responses. We have designed and tested a procedure-agnostic workflow to forecast infection and generated baseline predictions from preoperative and intraoperative phenotypes extracted from electronic health records. Laboratory values were then strategically imputed to build a prediction model that incorporates both their static and kinetic properties, achieving high recall for infection. The integration of kinetics revealed unknown informative combinations of routine markers among hepatic, renal, and bone marrow functions. Applying explainable machine learning ensures clinical applicability by elucidating how different organ systems respond to and predict infectious complications. By enabling earlier detection of postoperative infections, this work enhances patient outcomes and optimizes resource utilization in surgical care.