Explainable Machine Learning for Perioperative Risk Stratification of Radiographic Adjacent Segment Degeneration After Short-Segment Lumbar Fusion

Study Design. Retrospective Cohort Study. Objectives. To develop and externally validate an explainable machine-learning framework for perioperative risk stratification of radiographic adjacent segment degeneration (ASDeg) after short-segment lumbar fusion. Summary of Background Data. Radiographic ASDeg is frequently observed after lumbar fusion and may represent an early structural phenotype preceding symptomatic adjacent segment disease (ASDis) in some patients. However, existing risk assessment approaches are limited by heterogeneous risk factors, insufficient model interpretability, and limited external validation. Machine-learning methods may improve perioperative risk stratification by integrating clinical, radiographic, surgical, and functional variables. Methods. Clinical data were retrospectively collected from two hospitals. The internal cohort included 570 patients who underwent posterior short-segment lumbar fusion for lumbar degenerative disease, and an independent cohort of 150 patients from another institution was used for external validation. The internal cohort was randomly divided into training and internal test sets at a 7:3 ratio using stratified sampling according to ASDeg status. Feature selection was performed exclusively in the training set using least absolute shrinkage and selection operator regression (LASSO), random forest-recursive feature elimination (RE-RFE), and Boruta. Five algorithms were developed and compared: logistic regression, random forest (RF), extreme gradient boosting (XGBoost), Light Gradient Boosting Machine(LightGBM), and multilayer perceptron (MLP). Model performance was evaluated using discrimination, calibration, precision-recall (P-R) analysis, and decision-curve analysis(DCA). Shapley Additive Explanations (SHAP) were used for model interpretation. Results. Radiographic ASDeg occurred in 212 of 570 patients in the internal cohort. Five perioperative variables were retained for model construction: preoperative intervertebral space height (ISH), postoperative pelvic incidence-lumbar lordosis (PI-LL) mismatch, frailty, Coflex implantation, and preoperative Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC)-defined lower-extremity dysfunction. Among the candidate algorithms, the RF model showed the highest discriminative performance, with an AUROC of 0.782 in the internal test set and 0.749 in the external validation cohort. SHAP analysis identified preoperative ISH as the strongest contributor to model output. Conclusions. This externally validated RF-based model provides a structured and interpretable framework for postoperative radiographic ASDeg risk stratification after short-segment lumbar fusion. By integrating clinically accessible perioperative variables, the model may support individualized imaging follow-up and provide a preliminary basis for future studies using symptomatic adjacent segment disease or revision surgery as clinically oriented endpoints.