INTRODUCTION: Lumbar spondylosis, a leading cause of chronic lower back pain, is often diagnosed late due to the lack of sensitive diagnostic tools. Emerging evidence links neurovascular and metabolic changes to the early pathogenesis of lumbar spondylosis. This study aims to develop a machine learning framework integrating structural magnetic resonance imaging (MRI) and metabolomic profiling to identify biomarkers predictive of early disease progression.
METHODS: Structural MRI data from the SPIDER dataset, comprising 447 sagittal T1 and T2 MRI scans from 218 participants (mean age: 45.6 ± 7.2 years; 54.1% female), was analyzed alongside serum metabolomic data from the Metabolomics Workbench. Participants were stratified into early lumbar spondylosis (n=80), advanced lumbar spondylosis (n=70), and healthy controls (n=68). Deep learning models, including convolutional neural networks (CNNs), were employed to extract imaging biomarkers such as intervertebral disc height and spinal canal narrowing. Metabolomic profiling quantified neurovascular-associated metabolites, including lactate, succinate, and adenosine. A hybrid machine learning model integrating MRI and metabolomic data was trained using an 80/20 train-test split with five-fold cross-validation. Model performance metrics included accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC).
RESULTS: The hybrid model achieved an accuracy of 86.7% (95% CI: 85.0–88.4) and an AUC of 0.91 (95% CI: 0.89–0.93) for detecting early lumbar spondylosis. Among imaging biomarkers, reduced intervertebral disc height at the L4-L5 level emerged as the most predictive feature (SHAP score: 0.31), indicating its critical role in the structural deterioration associated with early spondylosis. Elevated serum lactate was identified as the leading metabolic marker (SHAP score: 0.29), suggesting its association with early neurovascular dysregulation. Subgroup analysis revealed consistent performance in participants with comorbid conditions such as diabetes and mild obesity, achieving a sensitivity of 83.5% (95% CI: 80.2–86.8) and specificity of 88.3% (95% CI: 85.4–91.2). These results indicate the robustness of the model across diverse participant groups. When combining MRI and metabolomics data, the diagnostic performance showed a significant improvement compared to models relying on a single modality (AUC: 0.91 vs. 0.78, p < 0.001), affirming the complementary value of multimodal integration.
DISCUSSION: This study indicates the potential of integrating structural MRI and metabolomic profiling with machine learning for early lumbar spondylosis diagnosis. The multimodal approach effectively captured structural changes and metabolic disruptions, achieving higher diagnostic accuracy compared to single-modality models. Reduced intervertebral disc height emerged as the key imaging biomarker, aligning with disc degeneration pathology, while elevated serum lactate indicated early metabolic and neurovascular dysfunction. The model performed well across diverse subgroups, suggesting applicability to real-world populations. However, the relatively small dataset size may limit representation of lumbar spondylosis variations. Expanding the dataset and including broader demographic and clinical profiles could improve its utility. Investigating the causal links between metabolic markers and structural deterioration is crucial for advancing understanding. Future work should include longitudinal studies to confirm biomarker predictive capabilities and assess clinical decision-making utility. Refining the framework with larger-scale data and advanced neuroimaging could pave the way for a robust, personalized diagnostic tool for early spondylosis management.