Screening and Validation of Potential Risk Factors for Lumbar Disc Herniation Using Explainable Machine Learning Models and Mendelian Randomization

Objectives: Lumbar disc herniation (LDH) is a leading cause of low back and leg pain worldwide. This study aimed to develop and internally validate an interpretable machine-learning (ML) model for predicting LDH risk, and to explore potential causal associations of modifiable risk factors using Mendelian randomization (MR). Material and Methods: In this study, a total of 730 participants (360 LDH cases and 370 controls) were enrolled, and demographic, occupational, lifestyle, medical history and psychosocial indicators were collected. Least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm were used for feature selection. Ten ML algorithms were trained on a 70% training set and evaluated on a held-out 30% test set. Predictive performance was assessed by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision and F1-score, and calibration was evaluated by calibration curves. Shapley Additive exPlanations (SHAP) analysis was used to interpret the optimal model. Subsequently, two-sample MR analysis was performed for the modifiable factors prioritized by ML, using intervertebral disc degeneration (IVDD) genome-wide association study data as a surrogate outcome for LDH. Results: Among the 10 ML models, CatBoost achieved the highest discrimination, with an AUC of 0.876 (95% CI 0.830–0.922) in the held-out test set, and showed acceptable calibration. SHAP analysis identified a history of lumbar spine trauma and a family history of LDH as the most important contributing factors. MR analysis suggested that genetically proxied time spent watching television, sleeplessness, type 2 diabetes and smoking initiation were associated with an increased risk of IVDD, whereas physical activity was inversely associated with IVDD risk. Conclusion: This two-stage framework integrating interpretable ML and MR prioritized a set of candidate risk factors for LDH. The predictive model was internally validated only, and the online risk calculator should be regarded as a research prototype whose clinical utility requires external, multicentre validation. Causal inferences from MR apply to IVDD as a surrogate phenotype and provide indirect, hypothesis-supporting evidence for LDH.