Interpretable Machine Learning Models Based on Shapley Additive Explanations for Predicting the Risk of Cerebrospinal Fluid Leakage in Lumbar Fusion Surgery

Abstract

Study Design. Retrospective study. Objectives. The objective of this investigation was to formulate and internally verify a customized machine learning (ML) framework for forecasting cerebrospinal fluid leakage (CSFL) in lumbar fusion surgery. This was accomplished by integrating imaging parameters and using the SHapley Additive exPlanation (SHAP) technique to elucidate the interpretability of the model. Summary of Background Data. Given the increasing incidence and surgical volume of spinal degeneration worldwide, accurate predictions of postoperative complications are urgently needed. SHAP-based interpretable ML models have not been used for CSFL risk factor analysis in lumbar fusion surgery. Methods. Clinical and imaging data were retrospectively collected from 3505 patients who underwent lumbar fusion surgery. Six distinct machine learning models were formulated: extreme gradient boosting (XGBoost), decision tree (DT), random forest (RF), support vector machine (SVM), Gaussian naive Bayes (GaussianNB), and K-nearest neighbors (KNN) models. Evaluation of model performance on the test dataset was performed using performance metrics, and the analysis was executed through the SHAP framework. Results. CSFL was detected in 95 (2.71%) of 3505 patients. Notably, the XGBoost model exhibited outstanding accuracy in forecasting CSFLs, with high precision (0.9815), recall (0.6667), accuracy (0.8182), F1 score (0.7347), and AUC (0.7343). In addition, through SHAP analysis, significant predictors of CSFL were identified, including ligamentum flavum thickness, zygapophysial joint degeneration grade, central spinal stenosis grade, decompression segment count, decompression mode, intervertebral height difference, Cobb angle, intervertebral height index difference, operation mode, lumbar segment lordosis angle difference, Meyerding grade of lumbar spondylolisthesis, and revision surgery. Conclusions. The combination of the XGBoost model with the SHAP is an effective tool for predicting the risk of CSFL during lumbar fusion surgery. Its implementation could aid clinicians in making informed decisions, potentially enhancing patient outcomes and lowering healthcare expenses. This study advocates for the adoption of this approach in clinical settings to enhance the evaluation of CSFL risk among patients undergoing lumbar fusion.