Interpretable Evolutionary Learning for Personalized Risk Prediction of Postoperative Infection after Liver Resection

Early diagnosis and treatment of postoperative infection (POI) are crucial for optimizing clinical outcomes. However, current POI prediction models remain unsatisfactory due to their limited interpretability and lack of personalized risk insights, restricting precise antibiotic use and personalized postoperative care strategies. This study aims to develop an interpretable model for predicting the individual POI risk within 30 days after liver resection. This retrospective study included 785 liver resection patients with 125 preoperative and early postoperative clinical factors. The dataset was divided into training (n = 551) and test (n = 234) sets in a 7:3 ratio. An interpretable evolutionary learning-driven method, EL-PIRLR, was proposed to develop a POI risk prediction model after liver resection. EL-PIRLR performs optimal feature selection in conjunction with an XGBoost classifier. EL-PIRLR identified 27 risk factors for modeling and achieved an accuracy of 87.50% and an AUC of 0.889 on the test set. Postoperative fever patterns, surgical difficulty, and the Charlson Comorbidity index emerged as crucial predictors of the POI risk. Using an uncertainty plot analysis, defining risk scores between 0.1 and 0.9 as the uncertainty region, EL-PIRLR achieved 92.05% accuracy for confidently predicted cases, covering 84.62% of the dataset. This demonstrates high reliability with a 98.70% negative predictive value in identifying non-infection patients. EL-PIRLR outperformed several state-of-the-art feature selection techniques and classification models. Precise POI prediction significantly impacts infection management and postoperative care. EL-PIRLR enhances the precision of hepatectomy POI risk prediction, facilitating personalized risk evaluations and aiding clinical decision-making.