Introduction
Lumbar spine degeneration progresses at different rates and over decades, making it difficult to study in humans. Furthermore, multiple degenerative trajectories likely exist and are challenging to delineate using cross-sectional imaging data. The aim of this study was to use an innovative data-driven approach discover trajectories of lumbar spine degeneration from clinical MRI using baseline data from a large cohort of patients with chronic low back pain (cLBP). Identifying such trajectories could help establish interventional phenotypes and, thereby, improve patient selection for clinical trials and stratification for large-scale association studies.
Methods
This study utilized baseline data from 450 adult participants in the prospective comeBACK study.1 A subset of 337 participants whose clinical MRI has been graded for lumbar spine degenerative features was available for inclusion. Grading included disc degeneration, endplate pathologies, facet arthropathy, and sacroiliac joint pathology.2
To infer progression patterns from these cross-sectional imaging features (biomarkers), we used pySuStaIn3 to train an event-based model (EBM). The EBM estimates the probabilistic ordering of biomarker transitions from “normal” to “abnormal” states, identifying a most likely sequence of changes. Models were trained for 1-4 clusters (i.e. progression pattern subtypes). The optimal number of clusters was identified using information criterion and event position diagrams. All patients were subtyped and staged using the final trained model.
Differences in pain and physical function between the resulting subtypes were tested using a generalized linear mixed-effects model after propensity score matching (PSM) for the covariates age, gender, BMI, and smoking. PSM used logistic regression and 1:1 nearest neighbour matching without replacement. Associations between inferred stage and pain and physical function in each subtype were tested using linear regression controlling for the same covariates. Pain and physical function were examined using PROMIS Physical Function, PROMIS Pain Interference, PEG-3, Fear-Avoidance Beliefs Questionnaire, and presence of chronic widespread pain.
Results
The final model revealed two distinct progression patterns (Figure 1). A “disc-first” subtype (n=207, 61%) was characterised by a caudal to cranial segmental progression of disc degeneration prior to facet arthritis. A “facet-first” subtype (n=130, 39%) was characterised by caudal to cranial progression of facet arthritis without disc degeneration, followed by disc degeneration initiated in the mid-lumbar spine. The patients in this facet-first subtype were a decade older, on average. In both subtypes, the later stages were significantly associated with older age (p<0.01 each) and higher BMI (p≤0.01 each). There were no differences in pain or physical function between subtypes or across stages within subtypes.
Discussion
This unsupervised data-driven model of lumbar spine degeneration progression identified two distinct trajectories: a disc-first subtype which exhibited early disc degeneration in the lower lumbar spine that progresses cranially and prior to facet arthritis, and a facet-first subtype which was characterised by healthier discs, especially at lower lumbar segments. These two subtypes align with dichotomous degeneration phenotypes proposed previously.4–6
Future studies will include validation with 24-month follow-up MRI data and replication in population-based cohorts with longer follow-up. If validated, these findings may improve patient phenotyping at scale.