Advanced MRI metrics improve the prediction of baseline disease severity for individuals with degenerative cervical myelopathy

Background context: Degenerative cervical myelopathy (DCM) is the most common form of atraumatic spinal cord injury globally. Degeneration of spinal discs, bony osteophyte growth and ligament pathology results in physical compression of the spinal cord contributing to damage of white matter tracts and grey matter cellular populations. This results in an insidious neurological and functional decline in patients which can lead to paralysis. Magnetic resonance imaging (MRI) confirms the diagnosis of DCM and is a prerequisite to surgical intervention, the only known treatment for this disorder. Unfortunately, there is a weak correlation between features of current commonly acquired MRI scans ("community MRI, cMRI") and the degree of disability experienced by a patient.

Purpose: This study examines the predictive ability of current MRI sequences relative to "advanced MRI" (aMRI) metrics designed to detect evidence of spinal cord injury secondary to degenerative myelopathy. We hypothesize that the utilization of higher fidelity aMRI scans will increase the effectiveness of machine learning models predicting DCM severity and may ultimately lead to a more efficient protocol for identifying patients in need of surgical intervention.

Study design/setting: Single institution analysis of imaging registry of patients with DCM.

Patient sample: 296 patients in the cMRI group and 228 patients in the aMRI group.

Outcome measures: Physiologic measures: Accuracy of machine learning algorithms to detect severity of DCM assessed clinically based on the modified Japanese Orthopedic Association (mJOA) scale.

Methods: Patients enrolled in the Canadian Spine Outcomes Research Network registry with DCM were screened and 296 cervical spine MRIs acquired in cMRI were compared with 228 aMRI acquisitions. aMRI acquisitions consisted of diffusion tensor imaging, magnetization transfer, T₂-weighted, and T₂*-weighted images. The cMRI group consisted of only T₂-weighted MRI scans. Various machine learning models were applied to both MRI groups to assess accuracy of prediction of baseline disease severity assessed clinically using the mJOA scale for cervical myelopathy.

Results: Through the utilization of Random Forest Classifiers, disease severity was predicted with 41.8% accuracy in cMRI scans and 73.3% in the aMRI scans. Across different predictive model variations tested, the aMRI scans consistently produced higher prediction accuracies compared to the cMRI counterparts.

Conclusions: aMRI metrics perform better in machine learning models at predicting disease severity of patients with DCM. Continued work is needed to refine these models and address DCM severity class imbalance concerns, ultimately improving model confidence for clinical implementation.