Abstract
OBJECTIVES: Emerging systemic immune-inflammatory biomarkers demonstrate potential for predicting postoperative complications. This study develops machine learning models to assess the combined predictive value of Aggregate Index of Systemic Inflammation (AISI), Systemic Immune-Inflammation Index (SII), CRP-Albumin-Lymphocyte (CALLY) index and Subcutaneous Lumbar Spine Index (SLSI) for surgical site infection (SSI) following posterior lumbar spinal fusion. METHODS: This retrospective study analyzed 2,921 patients undergoing posterior lumbar spinal fusion at two tertiary hospitals in Guangxi (August 2017-January 2025). Data were partitioned into training (70%) and validation (30%) groups. Feature selection via univariate regression analysis identified predictive variables, followed by model development using ten machine learning algorithms: logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), XGBoost, neural network, K-nearest neighbors(KNN), AdaBoost, LightGBM, and CatBoost. Hyperparameters were optimized with 10-fold cross-validation. The top seven performing models (assessed by AUC, accuracy, sensitivity, specificity, precision, and F1 scores) were integrated into a dynamic nomogram. Internal validation employed ROC analysis and calibration curves, while Shapley Additive Explanations (SHAP) values interpreted feature importance in the optimal model. RESULTS: Among 2,921 screened patients, 1,272 met inclusion criteria. Consensus feature selection across the seven top-performing ML algorithms identified AISI, SII, CALLY and SLSI as independent predictors of SSI. The derived nomogram demonstrated exceptional discrimination (training groups AUC = 0.966; C-index = 0.993, 95% CI 0.984-0.995) and excellent calibration. Additionally, the SHAP method emphasized the significance of AISI, SII, CALLY and SLSI as independent predictors influencing the machine learning model's predictions. CONCLUSION: The AISI, SII, CALLY and SLSI emerged as independent predictors of SSI following posterior lumbar spinal fusion. Our machine learning-derived nomogram demonstrated high discriminative accuracy and clinical applicability through dynamic risk stratification. Leveraging the SHAP methodology enhances model interpretability, thereby empowering healthcare providers to proactively mitigate SSI occurrences and enhance overall patient outcomes.