128 - Using integrated MRI signatures of muscle degeneration and spinal pathology as predictors of response to treatment for chronic low back pain

INTRODUCTION
The association between lumbar spine structural abnormalities from MRI and cLBP has been long debated. Quantification of individual muscles and pathology across individual spinal levels renders numerous separate metrics that are often analyzed in isolation. This ignores the interrelated nature of multiple metrics across multiple spine segments, challenging the use of MRI for diagnosis, risk assessment, and patient stratification.

AIMS/OBJECTIVES
We hypothesized that multivariate approaches that integrate all MRI metrics and their relationship into composited numeric summaries capturing the overall MRI pathology (signatures) within an individual would associate with progression of symptoms for cLBP. We conducted a hierarchical multiple factor analysis (hMFA) of lumbar structural anatomy from baseline MRIs of subjects who participated in a clinical trial of a neurostimulation to the multifidus1,2 to assess if diverse muscle and spinal pathology data from MRI3 could be used to predict response to treatment and inform patient selection for intervention.

METHODS
T2-weighted MRI scans from the lumbar spine from 168 subjects at baseline were quantified for two different feature domains: paraspinal muscles (Multifidus, Erector Spine, Psoas, and Quadratus Lumborum) quality and the presence of adjacent spinal pathology at each spinal level. Paraspinal muscle quality was measured using algorithmic muscle segmentation model and validated thresholding prediction model for fat fraction. Adjacent spinal pathology (disc degeneration, annular tears, modic changes, disc bulge, facet arthropathy, stenosis, etc) was graded by an expert. All MRI variables were analyzed at once using hMFA following a multi-segmental hierarchy of the lumbar spine (Fig. 1). hMFA finds combinations of the analyzed variables called dimensions or signatures that explain major sources of variance in the dataset, and provides summary scores for each individual. In a subsequent analysis, the summary scores were regressed with demographics and used to predict the Oswestry Disability Index (ODI) changes over time using a linear mixed model.

RESULTS
Quantified structural anatomy from 168 subjects with MRI at study baseline were included. Overall, we observed a complex relationship between quantitative metrics of muscle quality (fat fraction) and qualitative adjacent spine pathologies across the lumbar spine. The first hMFA dimension explained 13.41% of the total variance, while dimensions 2, 3 and 4 explained 6.22%, 5.14%, and 4.95% respectively. The MRI signature represented by the first dimension can be seen in Fig. 1, with average fat fraction in MF and ES muscles across the lumbar spine being the major source of variance. In predicting ODI, patients with negative scores in dimension 1, meaning better muscle quality, presented a better improvement at 12 and 24 months of follow up after the initiation of the study (Dim. 1: p = 0.031).

DISCUSSION
Data integration methods such as hMFA that find signatures from multi-domain and multi-level datasets may have clinical utility as multivariate biomarker for LBP patient stratification and prognosis. Our results show that an integrated MRI signature for paraspinal muscle quality associated with patient response to treatment for multifidus neurostimulation. Future work will optimize this method to improved prediction for response to treatment to inform patient selection for multifidus neurostimulation.