Related Experiment Videos
Development and validation of explainable machine learning models to predict the risk of adjacent segment disease
Mingzheng Zhao1, Honghao Yang1, Shixuan Guo1
1Department of Orthopedic Surgery, Beijing Chao-Yang Hospital, Beijing, China.
Background:
Adjacent segment disease (ASD) is a common and serious long-term complication following lumbar fusion surgery. This study aims to develop an interpretable online prediction model for the early identification of ASD after lumbar fusion surgery by leveraging machine learning (ML) and the Shapley Additive Explanations (SHAP) algorithm.
Methods:
This retrospective study included patients diagnosed with lumbar degenerative disease who underwent posterior L4-5 lumbar fusion surgery between January 2013 and January 2020. The time to ASD diagnosis ranged from 12 to 96 months at L3-4 and from 7 to 96 months at L5-S1 after surgery. Spearman correlation analysis and the Boruta algorithm were used for feature selection to ensure consistency of variables across different models. Ten machine learning models-including decision tree (DT), random forest (RF), XGBoost, logistic regression (LR), among others-were trained and evaluated. Model performance was assessed using accuracy, sensitivity (recall), specificity, precision, F1 score, and area under the receiver operating characteristic curve (AUC). Additionally, web-based calculators were developed for both segments to facilitate individualized risk prediction.
Results:
Among 441 patients, ASD incidence was 24.0% (106/441) at L3-4 and 19.95% (88/441) at L5-S1. The most important predictors included endplate bone quality (EBQ), foraminal and spinal canal stenosis, disc degeneration, facet joint osteoarthritis, Modic changes, and alignment parameters. XGBoost demonstrated the best predictive performance for L3-4 ASD, while LR was optimal for L5-S1 ASD.
Conclusions:
Machine learning models, specifically XGBoost for proximal ASD and LR for distal ASD-showed robust predictive performance after lumbar fusion. Preoperative EBQ, disc degeneration, facet joint pathology, and spinal alignment were key risk factors. The development of a web-based calculator based on these models may enable early risk stratification and support personalized clinical decision-making for lumbar fusion patients.