An Interpretable Machine Learning Model for 90-Day Mortality After Major Head and Neck Surgery

Introduction and aims Head and neck surgery carries substantial perioperative mortality, yet existing risk models lack head-and-neck specificity and interpretability. We aimed to develop an interpretable machine learning model for 90-day all-cause mortality after major head and neck surgery using LASSO, Boruta, and SHAP. Methods This study conducted a retrospective cohort study of adult patients (≥18 years) undergoing major head and neck surgery between 2015 and 2020 in the INSPIRE database. LASSO regression and the Boruta algorithm were applied for feature selection. Five classifiers—LASSO, XGBoost, LightGBM, CatBoost, and AdaBoost—were trained and compared under repeated stratified 5-fold cross-validation. SHAP analysis of the Boruta–XGBoost model provided individualized interpretation of nonlinear prediction drivers. Results Among 15,694 patients, 90-day mortality was 0.44% (n = 69). At the natural outcome prevalence, LASSO showed the numerically highest discrimination (AUC 0.83, 95% CI 0.77-0.89), sensitivity (0.79, 95% CI 0.70-0.88), and F1 score (0.028, 95% CI 0.025-0.032). The AUCs of CatBoost, AdaBoost, XGBoost, and LightGBM were 0.81 (95% CI 0.75-0.87), 0.80 (95% CI 0.74-0.86), 0.78 (95% CI 0.72-0.84), and 0.77 (95% CI 0.70-0.84), respectively; however, the confidence intervals overlapped substantially. Within the Boruta–XGBoost model, SHAP analysis identified preoperative CRP (mean |SHAP| = 1.050), albumin (0.524), phosphate (0.419), sodium (0.419), lymphocyte count (0.348), ASA class (0.323), preoperative MACE within 30 days (0.317), platelet count (0.271), ALP (0.246), BUN (0.235), and troponin (0.154) as dominant predictors. Elevated inflammatory and metabolic markers and depleted nutritional and haematologic reserves shifted predictions toward higher mortality. Conclusions An interpretable machine learning framework integrating LASSO, Boruta, and SHAP identified parsimonious, biologically plausible predictors of 90-day mortality after head and neck surgery. Clinical significance SHAP-based individualized risk explanations translate algorithmic outputs into modifiable preoperative targets—inflammation, nutrition, electrolyte balance, and cardiovascular reserve—supporting personalized risk stratification, informed consent, and targeted prehabilitation in this high-risk surgical population.