Associations between MRI features versus knee pain severity and progression: Data from the Vancouver Longitudinal Study of Early Knee Osteoarthritis

Objective To determine associations between features of osteoarthritis (OA) on MRI and knee pain severity and knee pain progression. Design Baseline, 3.3- and 7.5-year assessments were performed for 122 subjects with baseline knee pain (age 40–79), sample-weighted for population (with knee pain) representativeness. MRIs were scored for: osteophytes (0:absent to 3:large); cartilage (0:normal to 4:full thickness defect; 0/1 collapsed); subchondral sclerosis (0:none to 3:>50% of site), subchondral cyst (0:absent to 3:severe), bone marrow lesions (0:none to 3:≥50% of site); and meniscus (0:normal to 3:maceration/resection), in 6–8 regions each. Per feature, scores were averaged across regions. Effusion/synovitis (0:absent to 3:severe) was analyzed as ≥2 vs. <2. Linear models predicted WOMAC knee pain severity (0–100), and binary models predicted 10+ (minimum perceptible clinical improvement [MPCI]) and 20+ (minimum clinically important difference [MCID]) increases. Models were adjusted for age, sex, BMI (and follow-up time for longitudinal models). Results Pain severity was associated with osteophytes (7.17 per unit average; 95% CI = 3.19, 11.15) and subchondral sclerosis (11.03; 0.68, 21.39). MPCI-based pain increase was associated with osteophytes (odds ratio per unit average 3.20; 1.36, 7.55), subchondral sclerosis (5.69; 1.06, 30.44), meniscal damage (1.68; 1.08, 2.61) and effusion/synovitis ≥2 (2.25; 1.07, 4.71). MCID-based pain increase was associated with osteophytes (3.79; 1.41, 10.20) and cartilage defects (2.42; 1.24, 4.74). Conclusions Of the features investigated, only osteophytes were consistently associated with pain cross-sectionally and longitudinally in all models. This suggests an important role of bone in early knee osteoarthritis.


Data collection
Source data came from a longitudinal study conducted in Vancouver, Canada,[30] a population-based cohort of individuals aged 40 to 79 with knee pain "on most days of the month at any time in the past and any pain in the past 12 months." Data collection for the source study has been previously described. [31,32] Subjects were excluded at baseline if they had inflammatory arthritis or fibromyalgia, previous knee arthroplasty, knee injury or surgery within the previous 6 months, knee pain referred from hips or back or were unable to undergo MRI. The examination was performed by an experienced rheumatologist (JC). We have previously reported in this cohort that, based on MRI cartilage damage and x-ray findings, 13% had no OA, 49% had pre-radiographic OA (cartilage damage but KL<2), and 38% had radiographic OA.
[32] Based on power calculations for the original source study,[30] this cohort enrolled 255 individuals, stratified by age decade and sex in roughly equal group sizes (stratum sizes were capped at 35) to ensure adequate sample size across the age-sex spectrum. Baseline visits occurred between 2002 and 2005. A random list of households was obtained from the telephone directory listings, which included address and telephone information. Invitation letters were mailed to randomly selected households. This was followed by standardized telephone screening for preliminary eligibility, followed by in-person detailed eligibility screening. Once adequate numbers were attained in a given age-sex group, data collection stopped within that group. The first time this happened, the sample collected to that point provided the estimated population proportional distribution of age and sex in those with knee pain. This distribution was used to develop a sample weight (see below). Collection continued until sufficient numbers were reached in each group. [32] In addition to the baseline cycle, two follow-up cycles were undertaken, at weighted mean 3.3 (SD 0.6) and 7.5 (SD 0.6) years. The present study includes the N = 122 subjects who were available at the third cycle (the second follow-up cycle), although only 108 were available to provide data in the middle cycle. The remainder of the baseline sample who were not available for the third visit dropped out for a variety of reasons (Fig 1), including deceased (n = 6), other no contact (40), declined participation (36), exclusions for rheumatic conditions (4), comorbidity (6), knee replacement (16), MRI contraindication (4), moved away (11), immobility (6), or MRI completed but not readable (4).
For each feature other than effusion/synovitis, region scores were added up and the sum divided by the number of regions to form region-average scores. Effusion/synovitis (0: absent, 1: mild, 2: moderate, 3: severe) was scored for the overall knee and analyzed as !2 vs. <2. Grades 0 and 1 were collapsed because MRI effusion/synovitis score 1 may reflect physiologic joint fluid. Grades 2+ were combined because there was only a single grade 3 effusion in the baseline source data, and in fact none in the longitudinal sample. Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) knee pain scale (normalized 0-100; higher numbers worse) was administered at each time point. [34] For use in sensitivity analyses, we collected self-reported data on medications, with the following question: "Please list all current medications that you are taking below, including the dose or how much of the medication you take, and how often you take it (for example NAME: Glucosamine, DOSE: 500mg, HOW OFTEN: twice a day). Please note that the types of medications have been divided into prescribed, over the counter, herbal therapies/supplements, and vitamins and minerals." We also collected self-reported data on physiotherapy use, with the following question: "Please mark all other or alternative therapies that you have used for your knee pain at any time in the past or currently. Indicate the time period of use and how often you have used it on average." [with one category labeled "Physiotherapy"].

Statistical methods
In order to obtain results that were population-representative, a sample weight was developed for the baseline sample as the ratio of population proportion in a given age-sex cell over the baseline sample proportion in that cell. The weight was scaled to sum to the baseline sample size (255). The longitudinal subset analyzed in the present study consisted of 122 subjects followed up from the original baseline sample to the third cycle.
[30] A sample weight was developed for the longitudinal sample as the ratio of baseline sample proportion in a given age-sex cell over the longitudinal sample proportion in that cell, multiplied by the baseline sample weight, thus retaining the population representativeness in this subsample. The sample weight was scaled to sum to the follow-up sample size (122). All analyses in the present study were weighted with the longitudinal sample weight.
Spearman rank correlations were computed among the MRI features at baseline. Cross-sectional models predicting WOMAC pain severity were fit on three time cycles simultaneously We performed several sensitivity analyses. First, we adjusted all models for self-reported current regular non-steroidal anti-inflammatory drug (NSAID) or acetaminophen use, and self-reported current or recent physical therapist (PT) use. Next, we explored the associations between the individual surface cartilage scores (dichotomized per surface into full thickness lesion or not) and our outcomes variables. We also performed two sensitivity analyses exploring possible alternatives to our aggregate osteophyte score: one eliminating medial/lateral patellar surfaces, the other eliminating superior/inferior patellar surfaces. Finally, we compared the followed up subsample of 122/255 to the subsample that was not followed up (133/ 255), on baseline values of the MRI features of interest, age, sex, and BMI.

Results
Weighted characteristics of the sample are described in Table 1 Fig 2 shows the distribution of the six within-subject feature maxima at baseline. Table 2 summarizes the progression of WOMAC pain as a continuous variable as well as via binary indicators for 10+ and 20+ units increase in pain. Average WOMAC pain remained fairly constant relative to its SD (changes between cycles 1-2, 2-3 and 1-3 respectively were -4.3, 4.5 and 0.7). However, subject-level pain could vary in either direction. Between cycles 1-2, 14.1% progressed by at least 10 units, and 6.9% progressed by at least 20 units. Between cycles 1-3, 22.3% progressed by at least 10 units, and 14.7% progressed by at least 20 units. Table 3 lists the coefficients and 95% confidence intervals (CIs) from the cross-sectional linear regression models, adjusted for age, sex and BMI. Significant predictors of WOMAC pain cross-sectionally included osteophytes (the regression coefficient representing the effect per unit region-average osteophyte grade across regions was 7.17 WOMAC pain units; 95% CI = 3.19, 11.15), and subchondral sclerosis (11.03; 0.68, 21.39). Choice of the exchangeable working correlation matrix in these models was validated according to lowest QIC (compared to an unstructured matrix). In sensitivity analyses further adjusted for NSAIDs, acetaminophen and PT use, statistical significance in these models did not differ. Table 4 lists the odds ratios (ORs) and 95% CIs for the longitudinal binary logistic models, adjusted for interval-baseline age, sex and BMI, plus cycle-to-cycle years of follow-up. Significant predictors of a 10+ increase in WOMAC pain included osteophytes (the odds ratio, representing the effect per unit region-average osteophyte score, was 3.20; 95% CI = 1.36, 7.55), subchondral sclerosis ( passed the Hosmer and Lemeshow GOF test at alpha = .05. In sensitivity analyses further adjusted for NSAIDs, acetaminophen and PT use, the direction of effects and statistical significance in these models did not differ. Our further exploration of full thickness cartilage lesions on individual surfaces revealed that a full thickness lesion was only borderline (p = 0.054) associated with WOMAC pain cross-sectionally on one of the six surfaces (medial-femoral), when adjusted for age, sex and BMI. This is consistent with the continuous model shown in Table 3. Full thickness cartilage lesion on the same surface (medial-femoral) was a significant predictor of both 10+ and 20+ changes in pain. Full thickness cartilage lesions on the other five surfaces were not statistically significantly associated with pain in either cross-sectional or longitudinal adjusted models.
The results of the sensitivity analyses exploring possible alternatives to our aggregate osteophyte score (one eliminating medial/lateral patellar surfaces, the other eliminating superior/ inferior patellar surfaces) were consistent with the primary analysis, both in direction of effects and statistical significance of the osteophyte variables.
Finally, there were no statistically significant differences between the followed up subsample of 122/255 vs. the subsample that was not followed up (133/255), on baseline values of the MRI features of interest, age sex and BMI (data not shown).

Discussion
In models adjusted for age, sex, BMI and, where applicable, years of follow-up, we investigated the association between knee pain severity (cross-sectional models) and progression (longitudinal models) versus the MRI features osteophytes, cartilage, subchondral sclerosis, subchondral cysts, bone marrow lesions, meniscus and effusion/synovitis, in a population-based cohort with knee pain. Cross-sectionally, the MRI features we found to be associated with pain severity included osteophytes and subchondral sclerosis. Several previous cross-sectional studies also found associations between pain severity and one or both of these features. [11,15,16,18,20,22] However, these studies did not directly analyze the WOMAC pain scale, and therefore the specific estimates could not be compared to ours. These studies also analyzed dichotomous presence of pain rather than severity, thus their results do not translate into analysis of pain severity. In addition, as discussed in the introduction, there is some disagreement among previous studies both in how pain variables are defined, as well as models and conclusions.
While many patients experienced pain progression, other did not, and some pain levels actually improved over time. Longitudinally, the MRI features we found to be associated with pain progression were dependent on the threshold for delta WOMAC pain used to define "progression", with cut points of either 10+ (based on the minimum perceptible clinical improvement or MPCI) or 20+ (based on the minimum clinically important difference or MCID). Statistically significant predictors in at least one of the pain progression models included osteophytes, cartilage defects, subchondral sclerosis, meniscal damage and effusion. Two previous longitudinal studies were identified finding associations between pain and one or more of these features. [8,9] Neither study analyzed progression on the WOMAC pain scale, although Javaid et al [9] reported odds ratios for pain incidence similar in magnitude to our findings. They reported an OR of 4.7 (1.3, 18) for the effect of severe osteophyte vs. no osteophyte on incident knee pain. This is comparable to our OR of 3.79 (1.41, 10.20) for the effect per unit region-average osteophyte score on MCID-based WOMAC pain progression, the interpretation being that more severe osteophytes relate to a greater frequency of largescale pain progression. It is noteworthy in our study that among the MRI predictors of pain progression, only osteophytes were statistically significant in both the MPCI-and MCID-based models. Subchondral sclerosis, meniscal damage and effusion/synovitis were only statistically significant in the MPCI-based models, which may suggest that these processes alone (without additional factors present), while detectable to the patient, are not sufficient to induce large increases in pain. Conversely, cartilage is only statistically significant in the MCID-based model, and while borderline significant (p<0.10) in the MPCI-based model, there it presents less than half the added risk seen in the MCID-based model. This may be explained by the fact that cartilage is aneural and avascular, and hence minor cartilage defects alone are insufficient to directly induce pain, until they become severe/full-thickness, at which point the underlying bone is exposed to an extent sufficient to effect large increases in pain (i.e., an "on-off" switch for more severe pain). Consistent with this interpretation are the findings of Sowers et al [11] (a crosssectional analysis of 724 knee MRIs) in which a significant cartilage-pain association was only evident with full-thickness defects. (In our sensitivity analyses exploring full thickness cartilage lesions on individual surfaces, we found one surface [medial-femoral] to be statistically significantly associated with both cross-sectional and longitudinal pain outcomes.) In a narrative review, Hunter et al [23] also cited exposed bone as the primary mechanism in the cartilagepain association, describing as secondary mechanisms associations between cartilage defects and other structural changes (e.g., BML or effusion). Finally, also consistent with this interpretation is our finding that osteophytes are statistically significant in both MPCI-and MCIDbased pain progression models; the continuum of minor to severe osteophytes could allow for a graduated boney involvement in knee pain. Collectively, these findings might suggest that the single most consistent (direct) cause of pain in knee OA, among those we studied, involves bone.
While some previous studies reported unadjusted (crude) results, the majority of studies (as we did) adjusted for age, sex and BMI. [9, 12-17, 20, 21, 25, 30, 32] A small number adjusted also for drug treatment, [12] depression, [17] occupation, [20] or history of joint injury. [20] To investigate the potential impact of certain potential additional adjustments on our findings, in sensitivity analyses we further adjusted all models for current regular NSAID or acetaminophen use, and current or recent PT use. In these sensitivity analyses, the direction of effects and statistical significance of the MRI features did not differ.
The limitations of our study deserve comment. While our study is population-based, it should be noted that the target population is not the overall population, but those with baseline knee pain, aged 40-79 at baseline, who were successfully followed up over an average of 7.5 years. As such, we cannot be sure that the results of this study are applicable to a more general population that includes people without baseline knee pain, or those outside the target age range of 40+. Another important limitation of this study is loss to follow-up; only 122 of the original 255 baseline subjects were successfully followed through to the third cycle. To explore the effect this might have had, we analyzed potential differences between those followed vs. not followed on age, sex, BMI, and the six MRI scores (at baseline) analyzed in this study, and found no statistically significant difference. Related to this is the sample size itself (N = 122). There are larger datasets on which similar analyses might be undertaken, for example the Osteoarthritis Initiative, and it would be prudent in future research to perform comparative analyses on these data. Another limitation of this study is that MRIs, although scored by a highly experienced, board-certified musculoskeletal radiologist (AG), were nevertheless not scored with recent semi-quantitative methods like the BLOKS/WORMS or MOAKS. This is because such methods were not yet available at the time of our baseline readings. However, the scoring methods we employed were based in part on earlier studies, e.g., Disler et al.
[33] Despite these limitations, this is one of the few studies to report results from both cross-sectional and longitudinal models together from the analysis of a common data set, and the first such bi-modal study to report effects on pain severity cross-sectionally rather than simple presence of knee pain. This study therefore offers a more complete picture of the implications of MRI features on knee pain as a whole-both cross-sectional severity (correlation) as well as longitudinal progression (prediction).
In summary, of the MRI features we investigated, only osteophytes were consistently associated with pain cross-sectionally (severity models) as well as longitudinally (MPCI-and MCIDbased pain progression models). This suggests an important role of bone in early knee osteoarthritis.

Author Contributions
Conceptualization: ECS AG JME JAK JS AT SN JC.