Machine learning approach to predict surgical site infection in head and neck squamous cell carcinoma patients after free flap reconstruction.

Abstract

Head and neck squamous cell carcinoma (HNSCC) patients undergoing free flap reconstruction face a high risk of surgical site infection (SSI). Logistic regression (LR) models for SSI prediction are limited by linear assumptions, while machine learning (ML) approaches like random forest (RF) may offer superior performance by handling complex clinical data. This study aimed to identify SSI risk factors and compare the predictive performance of LR and RF models. This retrospective study included 442 HNSCC patients. Two predictive models were constructed based on LR and RF methods, respectively. The predictive performance of two models was assessed based on area under the receiver operator characteristic curves, calibration curve, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) was applied in the RF model to analyze the impact of features on prediction results. The RF model outperformed LR, achieving higher accuracy, sensitivity, specificity, and AUC. Calibration curves indicated superior alignment of RF predictions with observed outcomes. DCA revealed higher net benefits for RF across a wide probability threshold range. SHAP analysis identified PNI, operation time, and NLR as top predictors. Novel systemic markers (PNI, NLR) and clinical factors are critical for risk stratification.