Development and internal validation of an explainable machine learning model for predicting textbook outcome after free flap reconstruction in oral cancer.

OBJECTIVE: To develop and internally validate an explainable machine learning model for predicting textbook outcome (TO) after free flap reconstruction in oral cancer surgery. METHODS: This single-center retrospective study included 752 patients with oral cancer who underwent radical resection with immediate free flap reconstruction between August 2017 and August 2025. TO was defined as no 31-day mortality, no 31-day unplanned readmission, postoperative length of stay ≤ 21 days, no major complications, negative surgical margins, and adequate unilateral neck dissection (≥16 lymph nodes retrieved). Patients were randomly divided into a training cohort (n = 526) and a validation cohort (n = 226). Feature selection was performed using false discovery rate-based collinearity filtering and random forest recursive feature elimination. Seven machine learning models were developed and compared. SHapley Additive exPlanations (SHAP) were used to improve interpretability. RESULTS: The overall TO achievement rate was 38.2%. Among all candidate models, XGBoost showed the best performance in the validation cohort, with an area under the receiver operating characteristic curve of 0.843 (95% CI, 0.790-0.890), an accuracy of 0.764, a sensitivity of 0.733, and a Brier score of 0.152. Albumin, age, operative time, tumor T stage, and neutrophil-to-lymphocyte ratio were the most important predictors. SHAP analysis identified clinically relevant threshold effects at age > 50 years, albumin > 45 g/L, and operative time > 400 min. CONCLUSIONS: The XGBoost model demonstrated good performance for predicting TO after free flap reconstruction in oral cancer and may support perioperative risk stratification and individualized management.