Using hierarchical unsupervised learning to integrate and reduce multi-level and multi-paraspinal muscle MRI data in relation to low back pain

Abstract

Purpose: The paraspinal muscles (PSM) are a key feature potentially related to low back pain (LBP), and their structure and composition can be quantified using MRI. Most commonly, quantifying PSM measures across individual muscles and individual spinal levels renders numerous separate metrics that are analyzed in isolation. However, comprehensive multivariate approaches would be more appropriate for analyzing the PSM within an individual. To establish and test these methods, we hypothesized that multivariate summaries of PSM MRI measures would associate with the presence of LBP symptoms (i.e., pain intensity).

Methods: We applied hierarchical multiple factor analysis (hMFA), an unsupervised integrative method, to clinical PSM MRI data from unique cohort datasets including a longitudinal cohort of astronauts with pre- and post-spaceflight data and a cohort of chronic LBP subjects and asymptomatic controls. Three specific use cases were investigated: (1) predicting longitudinal changes in pain using combinations of baseline PSM measures; (2) integrating baseline and post-spaceflight MRI to assess longitudinal change in PSM and how it relates to pain; and (3) integrating PSM quality and adjacent spinal pathology between LBP patients and controls.

Results: Overall, we found distinct complex relationships with pain intensity between particular muscles and spinal levels. Subjects with high asymmetry between left and right lean muscle composition and differences between spinal segments PSM quality and structure are more likely to increase in pain reported outcome after prolonged time in microgravity. Moreover, changes in PSM quality and structure between pre and post-spaceflight relate to increase in pain after prolonged microgravity. Finally, we show how unsupervised hMFA recapitulates previous research on the association of CEP damage and LBP diagnostic.

Conclusion: Our analysis considers the spine as a multi-segmental unit as opposed to a series of discrete and isolated spine segments. Integrative and multivariate approaches can be used to distill large and complex imaging datasets thereby improving the clinical utility of MRI-based biomarkers, and providing metrics for further analytical goals, including phenotyping.

Keywords: Hierarchical unsupervised learning; Integrative analysis; Low back pain; Lumbar spine; MRI; Multiple factor analysis; Paraspinal muscles.