Development and validation of a prediction model for infection in chronic nonhealing wounds

Purpose: To develop and externally validate a pragmatic and interpretable model that predicts infection risk in chronic nonhealing wounds using routine electronic medical record data. Methods: 300) was used for validation. Prespecified predictors were harmonized across sites. Feature selection used the overlap of Boruta and LASSO. Eight algorithms were compared with stratified five-fold cross-validation; the random forest (RF) was selected. Performance was assessed by AUROC with 95% CI, calibration, and decision curve analysis. Model interpretability was examined with SHAP. Results: Six consensus predictors were retained: smoking, diabetes duration, wound depth, elevated C-reactive protein, elevated procalcitonin, and hypoalbuminemia. The RF achieved AUROC 0.884 (95% CI 0.841-0.928) in the testing cohort with a calibration slope of 1.00 (95% CI 0.73-1.27) and higher net benefit than treat-all and treat-none across broad thresholds. External validation showed AUROC 0.855 (95% CI 0.807-0.904) with a calibration slope of 1.00 (95% CI 0.74-1.26) and similar decision utility. SHAP indicated hypoalbuminemia and inflammatory markers as dominant drivers, consistent with clinical reasoning. Conclusion: A six-variable RF model based on readily available data provides accurate, well-calibrated, and clinically useful prediction of infection in chronic nonhealing wounds, with transparent explanations to support bedside use. To facilitate immediate clinical application, this model has been deployed as a free, user-friendly web calculator. Prospective validation and impact evaluation across diverse settings are warranted

Brier score · Clinical decision support system · Cohort · Cohort study · Interpretability · Medical record · Random forest · Receiver operating characteristic · Retrospective cohort study · Pressure Ulcer Prevention and Management · Surgical site infection prevention · Wound Healing and Treatments