Oral Presentation 51st International Society for the Study of the Lumbar Spine Annual Meeting 2025

INTRODUCTION: Low back pain (LBP) is a heterogeneous disease with biological, psychological, and social components, and the interaction of these components is poorly understood.¹ In previous work, we screened hundreds of variables related to musculoskeletal health and identified that income and education, socioeconomic status (SES) metrics, were the strongest predictors of LBP severity and chronicity.² However, we could not evaluate anatomical factors like intervertebral disc degeneration, which is frequently implicated as a driver of LBP and has not been investigated in the context of SES. We hypothesize that low SES accelerates disc degeneration and LBP. Here, we utilize deep learning to automate the analysis of dual x-ray absorptiometry (DEXA) scans in the UK Biobank (UKB) imaging dataset (10,440 images) to enable a large-scale assessment of disc degeneration and SES.

METHODS: Study Population The UKB is a biomedical database that includes health information and lateral spine DEXA imaging for 50,000 people. Deep Learning Model Development Lumbar vertebral bodies (T12 to L4) were manually segmented from DEXA scans to develop a training dataset for machine learning (N=611 images). A computer vision model was developed that receives a DEXA scan as input and outputs a quadrilateral that corresponds to the corners of 5 lumbar vertebral bodies (Figure 1). Statistical Analysis To determine our preliminary model accuracy, we used the Intersection Over Union (IoU) metric, calculating an aggregate IoU for each spinal level (IoU = 1, perfect overlap; IoU = 0, no overlap). Following segmentation, we calculated disc height index³ (DHI: mean disc height divided by mean anterior-posterior disc width). We analyzed data using an ordinal regression model to determine the relationship between income/ neighborhood level multiple deprivation index (MDI) and LBP, as well as a mixed effects model to estimate the relationship between income/MDI and DHI, while controlling for age, sex, and body mass index (BMI) in both models.

RESULTS: Our model predicted vertebral body quadrilaterals in training and unseen test data (train IoU = 0.96, test IoU =.91) and was used to infer data for 10,440 participants. Confirming previous studies, there were significant relationships (p<0.05) between age, sex, BMI and pain and DHI (Table 1). We also found a significant relationship between spinal level and DHI. There were significant relationships between income and pain as well as MDI and pain (Figure 2b), but no relationship (p>0.05) between income or MDI and DHI (Figure 2a).

DISCUSSION: In this work, we sought to determine whether there is an anatomical driver of LBP in lower SES populations by evaluating the relationship between SES and lumbar disc degeneration measured by DHI. First, we developed a deep learning platform to measure DHI in lateral DEXA scans and inferred vertebral body location on 10,440 participants. We confirm the reliability of analysis by demonstrating well-known relationships between DHI and age, sex, and BMI⁴ as well as between pain and SES². Surprisingly, no relationship existed between SES and DHI. Thus, lumbar disc degeneration may not play a significant role in the development of LBP in the context of SES.