Development and validation of an early warning model for secondary spinal epidural abscess in pyogenic spondylitis based on interpretable machine learning: a single-center retrospective cohort study

Background Secondary spinal epidural abscess (SEA) in pyogenic spondylitis (PS) can rapidly compress the spinal cord or cauda equina, leading to irreversible neurological deficits such as paraplegia. Once formed, it often requires emergency surgical decompression and is associated with significantly poorer outcomes. Therefore, early identification of high-risk patients before abscess formation is of important clinical value. Objective To develop and internally validate a machine learning model for predicting the risk of secondary SEA in patients with PS, based on clinical indicators routinely available early after admission. Methods A total of 240 patients with PS were retrospectively enrolled at a single center between January 2016 and December 2025, of whom 60 (25.0%) developed secondary SEA. Patients were stratified and randomly divided into a training set ( n = 168) and a test set ( n = 72) at a 7:3 ratio. Fifty-seven candidate variables, including clinical, laboratory, and imaging data first available early after admission, were collected. In the training set, predictors were screened using univariate logistic regression ( p < 0.10) combined with the least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation. Four models were subsequently developed: L2-regularized logistic regression, random forest, support vector machine, and eXtreme Gradient Boosting (XGBoost). Synthetic minority oversampling technique (SMOTE) was applied only within the training folds of cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), calibration curves, Hosmer-Lemeshow test, decision curve analysis, and SHapley Additive exPlanations (SHAP) interpretability analysis. Internal validation was performed using 1,000 bootstrap resamples, and robustness was examined through three sets of sensitivity analyses. Results Six predictors were identified: intervertebral disc T2 hyperintensity or enhancement, abnormal sensory level, number of vertebrae with endplate destruction, maximum diameter of paravertebral abscess, procalcitonin (PCT), and Frankel grade E (a protective factor). The random forest model achieved the numerically highest discrimination in the test set, with an AUC of 0.871 (95% CI 0.760–0.982) and an AUPRC of 0.728. At a cutoff value of 0.312, the accuracy was 79.2%, sensitivity 83.3%, and specificity 77.8%. Calibration was good (Brier score 0.108; Hosmer-Lemeshow χ 2 = 6.890, p = 0.549), and the bootstrap optimism-corrected AUC was 0.872. Decision curve analysis demonstrated positive net benefit across a threshold probability range of 0.05–0.70. In sensitivity analyses, the AUC remained stable at 0.851–0.868. Conclusion The random forest model based on six routine indicators available early after admission achieved an AUC of 0.871 (95% CI 0.760–0.982) in the test set, together with good calibration (Brier score 0.108) and clinical net benefit across a threshold probability range of 0.05–0.70. It provides a quantitative tool for early risk stratification and surveillance planning in high-risk patients, and its performance warrants further validation in multicenter prospective studies.