An interpretable machine learning model for MRI-based morphological classification of lumbar disc herniation using functional biomechanical features
Abstract
Lumbar disc herniation (LDH) is commonly assessed using magnetic resonance imaging (MRI), although disc morphology does not necessarily correspond to pain intensity, functional impairment, neurological findings, or prognosis. This retrospective single-center exploratory study examined whether functional biomechanical indicators were associated with MRI-based morphological classification of LDH. A total of 58 patients were included, comprising 15 patients with disc bulging and 43 with disc protrusion or extrusion. Five predictors—lumbar extension, straight-leg-raising asymmetry, side-flexion asymmetry, gluteus medius asymmetry, and right Jiaji regional surface electromyography root mean square—were used to construct an interpretable L2-regularized logistic regression model. Model performance was evaluated using repeated stratified five-fold cross-validation repeated 10 times, with participant-level out-of-fold probabilities averaged across repetitions. Predictors were selected before cross-validation; therefore, the reported performance may include selection-induced optimism. The model achieved a receiver operating characteristic area under the curve of 0.705 (95% confidence interval: 0.537–0.851), a precision-recall area under the curve of 0.844, and a Brier score of 0.218. At the exploratory Youden threshold of 0.566, balanced accuracy was 0.714, sensitivity was 0.628, specificity was 0.800, positive predictive value was 0.900, and negative predictive value was 0.429. A prespecified threshold of 0.500 yielded a balanced accuracy of 0.692, sensitivity of 0.651, and specificity of 0.733. Calibration was suboptimal, with an intercept of 0.999 and a slope of 0.752. In a separate exploratory bootstrap analysis, coefficient sign stability ranged from 97.0% to 99.9%, although the intervals remained wide. These findings suggest preliminary associations between selected functional biomechanical indicators and MRI-based disc morphology. Given the small and imbalanced sample, predictor-selection optimism, limited negative predictive performance, suboptimal calibration, and absence of external validation, the model should be regarded as hypothesis-generating rather than as a standalone clinical decision-support tool.