To our knowledge this is the first study describing how specific subgroups of MRI findings in motion segments, derived using Latent Class Analysis, are associated with pain. Motion segments in Subgroups 4 and 5 had multiple and more severe MRI findings at the lower lumbar levels, and people with these findings were more likely to have had both LBP over the last year and non-trivial LBP compared to those with very few MRI findings on all of their lumbar vertebral levels. None of the other subgroups at the upper or lower lumbar levels were associated with any of the three LBP outcomes. Also, the outcome LBP past month was not associated with any of the subgroups and only moderate consistency was found between the associations with LBP past year and non-trivial LBP.
Meaning of study and comparison with other studies
In a previous study, the associations between LBP in the past month, LBP in the past year and seeking care for LBP and the single MRI variables in this dataset were analysed by Kjaer et al. [
14], who found ORs that were mostly between 1.3 and 2.6 for LBP in the past year and seeking care for LBP. Only reduction in disc signal intensity, reduction in disc height and Modic changes were associated with having LBP in the past month. Modic changes (VESC) were found to be more strongly associated with LBP in the past year (OR 4.2, 95% CI 2.2–8.2) and anterolisthesis (
n = 12) (OR 6.1, 95% CI 1.0–∞). In our study, both Subgroups 4 and 5 had a large proportion of motion segments with VESC and it is possible that VESC is the single variable that drives the association we observed. On the other hand, VESC could be reflective of a late stage of the degenerative process and therefore VESC could be a proxy for severe degeneration of the lumbar motion segments. This finding is supported by Cheung and colleagues (2009) [
21] who found a positive correlation between the sum of degenerative disc MRI findings and LBP that was of more than 2 weeks duration and sufficiently severe to require physician consultation or treatment.
The OR for LBP in the past year (OR 3.4 (95% CI 1.7–7.0)) observed for Subgroup 4 and 5 combined were only moderate in size and therefore that association was no stronger than those previously found for associations with single MRI variables. [
14] However, for LBP in the past year the association (OR 11.3 (95% CI 1.6–495)) with Subgroup 5 (Severe degeneration and VESC) was larger than for Subgroup 4 (OR 3.0 (95% CI 1.5–19.6)) (Moderate degeneration and VESC), which suggests a dose-response relationship between severity of MRI findings and risk of symptoms. However, the confidence intervals of the ORs when testing Subgroup 4 and Subgroup 5 independently were very large due to the small cells in the contingency table and therefore this finding should be interpreted with caution.
This finding is in line with that of MacGregor et al. [
22], who reported a dose-response relationship between LBP and a sum score [
23] of the severity and extend of MRI findings. In their twin study exploring the structural, psychological, and genetic influences on LBP, MacGregor et al. found that structural (MRI) findings were the strongest predictor for LBP and that those with the upper quartile of the sum scores had 3.6 times the odds of reporting pain compared with those with scores in the lowest quartile [
22].
These results indicate that spinal pathology or degeneration, as seen on MRI, could still play a role in explaining some of the pain mechanisms in the bio-psycho-social model of LBP. However, as these data also show that 60% of the people with all five lumbar levels in Subgroup 1 (No or few findings) reported pain during the past year, it is obvious that MRI findings do not reflect the only pain mechanism in LBP, which underlines the need for considerable caution when interpreting MRI findings on an individual patient level.
Although, no previous studies have used Latent Class Analysis of vertebral MRI findings to explore their relationship to LBP, others have used Latent Class Analysis in other ways within LBP research. Takatalo et al. [
24] investigated the association between LBP severity and lumbar disc degeneration on MRI by using Latent Class Analysis to characterise subgroups of the pain and functional limitation features of 554 young Finnish adults using data collected at age 18, 19 and 21 years. They found five subgroups with different clinical profiles ranging from ‘no pain at all’ to ‘pain at all three time points’. They then related these clinical profiles to single MRI findings and found that disc degeneration and sum scores of disc degeneration were more prevalent in the three most symptomatic pain subgroups compared to the two least symptomatic ones, and that disc degeneration was associated with symptom severity.
Their approach of using Latent Class Analysis to subgroup clinical data, rather than MRI data, ensures that a person only belongs to one subgroup. In contrast, when creating the subgroups in our study, we focused on the co-existence of multiple MRI findings at one segmental level and so every participating person contributed five lumbar vertebral segments to the analysis. However, this approach complicates analysis of the LBP association, as it is the person (not the five vertebral levels) that report the presence of LBP and therefore, a best approximation of allocating an individual person to a single subgroup had to be made, even though this person could have the five vertebral segments in five different subgroups. That said, the modelling approach taken by Takatalo et al. [
24] does not ‘solve’ the dilemma of each person having five vertebral segments, it simply does not model this as explicitly and all studies of the association between MRI findings and pain face this same challenge.
Strengths and weaknesses
The MRI machine used in this study is a low-field (Tesla 0.2) MRI, which generates lower spatial, contrast and temporal resolution than a high-field MRI machine, but this does not automatically lead to lower diagnostic quality. Lee and colleges (2015) found that the diagnostic capability of low-field lumbar spine MRI was very comparable with that of high-field MRI, with Kappa values being excellent for disc herniation and spinal stenosis and good for nerve root compression [
25]. Also, in a study by Bendix et al. [
26] that compared the investigation of Modic changes (VESC) using low-field MRI (0.3 T) and high-field MRI (1.5 T), they found that these two types of machines were sensitive to different findings, with Modic changes Type 1 being detected three times more often using low-field MRI but Type 2 being detected two times more often when using high-field MRI. This could affect the observed prevalence of VESC in our data samples, although it is unknown which MRI approach is more accurate or if one is just more sensitive and less specific than the other.
A strength of this study is that both datasets were collected specifically for research purposes using the same MRI machine and the same experienced radiologist. However, as data from the two datasets were collected years apart (‘General population sample’ in 2000 and ‘Clinical sample’ in 2006) there is a theoretical risk of the radiologist having changed a cut-point or scale for determining when an MRI finding. This potential bias was addressed by standardising the description of the MRIs by means of a detailed protocol with high reproducibility that was used at both data collections.
The people in the general population cohort were of the same age, ensuring that all lumbar segments have the same ‘chronological age’ and that differences in the degenerative state is reflective of the ‘biological age’ of the motion segment. Although 412 people constituted a fairly large cohort for an imaging study, we still needed to merge two of the subgroups because the prevalence was too low in Subgroup 5 to enable the planned analyses. It is likely that a more detailed model would require an even larger sample size, although the prevalence of these subgroups may vary across settings.
Perspectives
Latent Class Analysis of MRI data is a novel method and previously we found that this multivariable approach detected clinically meaningful subgroups of MRI findings that could be arranged into biologically plausible hypothetical degenerative pathways with face validity [
11]. In a subsequent reproducibility study, we found that using this method in two different MRI samples showed some differences in the content of the subgroups but the overall pattern of increasing degeneration was quite similar [
12]. In the current study, we investigated the relationship between MRI findings and the presence of pain over the previous year and found that increasing severity of degeneration was associated with LBP. These results indicate that even though LBP is a multifactorial condition, motion segments with multiple and severe MRI findings could be part of the explanation of LBP. This study was conducted on a cohort of 40-year-old people and as increasing age leads to increasing spinal degeneration, it would be interesting to further study the relationship between subgroups of MRI findings and pain in other age groups, as the relationship between severe degeneration and pain is not necessarily constant over time.