Prognostic factors and prediction model for facial scar improvement in laser-treated patients: A machine learning-based retrospective cohort study

The face, being central and exposed, is highly susceptible to trauma and subsequent scar formation. Laser therapy is a common and effective treatment method for facial scars. However, treatment outcomes vary substantially. Consequently, we aimed to identify key prognostic factors and develop a predictive model for laser treatment outcomes using machine learning. We retrospectively enrolled patients seeking laser treatment for facial scars at our institution (2014-2024). Based on expert consensus, literature review, and clinical experience, we defined 21 potential predictors and 2 outcomes. Predictive models were developed using 7 machine learning algorithms (including Random Forest and XGBoost), supplemented by univariate and multivariate analyses. Performance was evaluated via receiver operating characteristic curves and decision curve analysis, with Shapley Additive Explanations (SHAP) providing model interpretability. The study included 1456 patients. Univariate and multivariate analyses identified 8 significant predictors, including sun protection and duration of care. The XGBoost model demonstrated superior performance in discrimination (area under the curve = 0.859), calibration (Brier score = 0.137), and precision-recall capability (precision-recall-area under the curve = 0.721). SHAP-based summary plots enabled global and local interpretation of the model. Machine learning proves a reliable tool for predicting facial scar laser treatment outcomes. The SHAP method effectively explains the XGBoost model's mechanisms, enabling clinicians to optimize personalized treatment strategies.