To develop and validate an explainable machine learning model for predicting bacterial biofilm formation in clinical wounds. We conducted a multicenter retrospective cohort study at two tertiary hospitals in China. From 550 eligible patients, a training cohort (n = 385) and a testing cohort (n = 165) were created, with an independent cohort (n = 300) for external validation. Predictors available at the index encounter were prespecified and imputed. Feature selection combined Boruta and LASSO, yielding six variables: Debridement performed, Chronic wound, Thermal therapy, Negative pressure wound therapy, Diabetes mellitus, and Silver dressing used. Eight algorithms were tuned with stratified fivefold cross-validation. Discrimination, calibration, and decision curve analysis were assessed. SHAP was used for model interpretation. Random forest achieved the best performance with AUC 0.929 in training, 0.861 in testing, and 0.837 in external validation, with good calibration and consistent net benefit on decision curves. SHAP ranked Debridement performed as most influential. Debridement performed, Thermal therapy, Negative pressure wound therapy, and Silver dressing used showed predominantly negative SHAP values, indicating inverse associations with biofilm formation. Chronic wound and Diabetes mellitus showed positive SHAP values. An interpretable random forest model with six routinely collected predictors accurately estimated biofilm formation in clinical wounds. The model’s strong external performance and biological plausibility suggest potential utility for early risk stratification and tailored wound management, warranting prospective validation in diverse clinical environments.
Jiang et al. (Sat,) studied this question.