Development and internal validation of prediction models for 30-day composite postoperative adverse outcomes in patients with maxillofacial space infections using preoperative inflammatory and nutritional indices: a retrospective cohort study

Abstract Objective To develop and compare models for predicting 30-day postoperative adverse outcomes in patients with maxillofacial space infections using preoperative clinical characteristics, infection extent, and inflammatory and nutritional indices. Methods In this single-center retrospective cohort study, we included patients undergoing incision and drainage and/or debridement between January 2018 and June 2025. The primary endpoint was a 30-day composite postoperative adverse outcome comprising repeat drainage/debridement, airway support, sepsis progression, severe deep infection progression, ICU admission, or death. Five hundred patients were randomly divided in a 7:3 ratio into training and test sets for model development and internal hold-out validation. Remaining missing values in nonessential candidate predictors were imputed using random forests, and features were selected by LASSO. LR, RF, SVM, NN, GBM, and XGBoost models were evaluated for discrimination, calibration, clinical net benefit, and SHAP-based interpretability. Results Of 500 patients, 166 experienced adverse outcomes; 350 and 150 were assigned to the training and test sets. LASSO selected 10 variables. In the exploratory post-LASSO multivariable analysis, age, sepsis at admission, deep neck involvement, mediastinal extension, neutrophil-to-lymphocyte ratio, prognostic nutritional index, and systemic immune-inflammation index were associated with postoperative adverse outcomes. SVM achieved AUCs of 0.837 (95% CI, 0.793–0.882) and 0.771 (95% CI, 0.681–0.853) in the training and test sets. Test-set sensitivity, specificity, and accuracy were 0.700, 0.940, and 0.860. LR achieved a slightly higher test-set AUC than SVM (0.773 vs. 0.771), whereas SVM had the lowest Brier scores (0.062 and 0.088 in the training and test sets, respectively) and generally showed greater net benefit across threshold probabilities of approximately 5%–55% in the test set. SHAP ranked neutrophil-to-lymphocyte ratio, prognostic nutritional index, C-reactive protein, and systemic immune-inflammation index as the four most important features. Conclusion The SVM model showed moderate discrimination and low overall prediction error but imperfect calibration in the test set, with discrimination similar to that of logistic regression. The models should be regarded as candidate prediction models pending external validation, recalibration where necessary, and clinical impact assessment.