Development and External Validation of an Explainable Machine-Learning Model for Predicting Postoperative Pulmonary Complications in Older Adults Undergoing Degenerative Spine Surgery

Le résumé fourni par la source

Background: Postoperative pulmonary complications (PPCs) remain common in older adults undergoing degenerative spine surgery and are associated with adverse outcomes. We aimed to develop and externally validate an explainable machine-learning model using routinely available perioperative variables to estimate individual PPC risk. Methods: We conducted a two-center retrospective cohort study including consecutive patients aged ≥65 years undergoing degenerative cervical or lumbar spine surgery under general anesthesia. The development cohort included 1200 patients (Affiliated Nanhua Hospital, University of South China; 2024–2025), and the external validation cohort included 600 patients (First Affiliated Hospital, University of South China; 2025). PPCs within 7 postoperative days were defined using EPCO criteria. Twenty-five prespecified predictors were considered. Missing data were handled using multiple imputation, and continuous variables were standardized. Feature selection was performed using LASSO with 10-fold cross-validation and the 1-standard-error rule. We trained six machine-learning models (DT, RF, SVM, XGBoost, LightGBM, ANN) and logistic regression as a benchmark, with hyperparameters tuned by 5-fold cross-validation in the development cohort. Performance was assessed by AUROC, calibration (plots and Brier score), decision-curve analysis, and threshold-based metrics; SHAP was used for interpretability. A web-based calculator was implemented for clinical use. Results: PPCs occurred in 222/1200 (18.5%) patients in the development cohort and 123/600 (20.5%) in the external validation cohort. LASSO selected 11 predictors. In external validation, the random forest model achieved AUROC 0.786 (95% CI 0.740–0.829) and Brier score 0.137 (95% CI 0.119–0.154), with favorable decision-curve performance, and achieved an accuracy of 73.83%, sensitivity of 68.29%, specificity of 75.26%, PPV of 41.58%, NPV of 90.20%, and an F1 score of 51.69% at the prespecified threshold. Conclusions: We developed and externally validated multiple prediction models for PPCs in older adults undergoing degenerative spine surgery. The random forest model provided balanced performance and SHAP-based interpretability and was retained as an implementation-focused model for perioperative or immediate postoperative risk estimation. Further prospective validation, comparison with established clinical risk scores, and model updating in more diverse clinical settings are needed before routine clinical implementation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.