Predicting Surgical Site Infection after Lumbar Laminectomy and Discectomy: A Cutting-edge Algorithmic Approach by Incorporating Ensembled Stacking into the Current State-of-the-art for Automated Machine Learning.

To develop an algorithmic approach for predicting surgical site infections (SSIs) in patients undergoing lumbar laminectomy and discectomy for adult degenerative spinal disease (DSD) by incorporating ensembled stacking into state-of-the-art (SOTA) automated machine learning (aML). The study utilized a comprehensive dataset from a prospective multicenter surveillance study on SSIs following lumbar laminectomy and discectomy to manage adult DSD. The Google Colab environment was adopted to load the dataset using Python programming language. Nine algorithms, including eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM), Neural Network (NN), Categorical Boosting (CatBoost), and Random Forest (RF), were adopted with hyperparameter tuning using the current SOTA for aML. Ensembling of the developed algorithmic models was carried out, followed by stacking and ensembled stacking. Five-fold stratified, shuffled cross-validation was implemented. The macro-weighted average Area Under the Receiver Operating Curve (mWA-AUROC) analysis was used to evaluate the discriminating classification ability of the developed models along with other evaluation metrics. A stacked ensemble algorithmic model, comprising a stacked XGBoost model and an ensemble of XGBoost, NN, CatBoost, LGBM, and RF algorithmic models, achieved an mWA-AUROC of 0.994, an accuracy of 98.7%, a sensitivity of 90% (95% CI: 68.30% - 98.77%) and a specificity of 98.81% (95% CI: 98.15% - 99.28%) upon predicting SSI. The top-weighted constituent model, XGBoost-20, identified operative time, smoking status, and patient age as the most significant predictors of SSI. We have made the development architecture of the algorithmic model available at GitHub for external validation. This study presented a novel algorithmic approach that integrated ensembled stacking into the current SOTA for aML to predict SSIs following lumbar laminectomy and discectomy procedures for adult DSD management. The performance of the stacked ensemble model highlighted its potential to serve as a valuable tool for clinicians, enabling more informed decision-making, optimized resource utilization, and enhanced patient outcomes in spine surgery. Future research should focus on validating the performance of the model in diverse clinical settings and exploring its integration into clinical practice.