Differential associations of patterns of co-existing MRI lesions with knee pain at rest and on joint loading: a within-person knee-matched case-control study

Background: To assess the association of co-existing MRI lesions with knee pain at rest or on joint loading. Methods: We included participants from Osteoarthritis Initiative whose pain score, measured by WOMAC sub-scales, differed by ≥ 1 point at rest (in bed and with sitting/lying down) or on joint loading (with walking and on stairs) between two knees. Cartilage morphology, bone marrow lesions, meniscus extrusion, meniscus morphology, Hoffa’s synovitis and synovitis-effusion were assessed using the compartment-specic MRI Osteoarthritis Knee Score. We performed latent class analyses to identify subgroups of co-existing MRI lesions and tted a conditional logistic regression model to examine their associations with knee pain. Results: Among 130 eligible participants, we identied ve subgroups of knees according to patterns of co-existing MRI lesions: I. minimal lesions; II. mild lesions; III. moderate morphological lesions; IV. moderate multiple reactive lesions; and V. severe lesions. Compared with subgroup I, the odds ratios (ORs) and 95% condence intervals (CI) of greater pain in bed were 1.6 (0.3, 7.2), 2.2 (0.5, 9.5), 6.2 (1.3, 29.6) and 11.2 (2.1, 59.2) for subgroups II-V, respectively. A similar association was observed between aforementioned subgroups and pain with sitting/lying down. The ORs (95% CI) of greater pain with walking were 1.0 (reference), 1.7 (0.5, 6.1), 0.7 (0.2, 2.3), 5.0 (1.4, 18.6) and 7.9 (2.0, 31.5) for subgroup I-V, respectively. The corresponding analysis for pain on stairs showed similar results. Conclusions: Distinct patterns of co-existing MRI lesions have different implications for the pathogenesis


Background
Pain at rest and pain on joint loading are two different manifestations of knee pain in osteoarthritis (OA) 1,2 . Pain at rest has been a marker for more severe cases of knee OA and a criterion utilized for the recommendation of total knee replacement (TKR) 3,4 . Pain on joint loading, e.g., pain with walking, is among the most commonly reported symptoms experienced by patients with knee OA 5 , and consequently contributes to physical disability [6][7][8] . These two types of pain are differentially associated with other clinical variables and response to treatment 2,9,10 . For example, neuropathic pain is more strongly associated with pain at rest than pain on joint loading among people with end-stage hip and knee OA 9 .
Additionally, patients with higher knee pain at rest are more likely to have less favorable pain relief after TKR 2 , whereas pain on joint loading is improved signi cantly compared with pain at rest among patients treated with TKR or placebo 2,10 . Collectively, these ndings suggest that pain at rest and pain on joint loading may have different underlying mechanisms and corresponding risk factors. However, to our knowledge, despite the different clinical relevance and impact on patients, differential risk factors for these two types of osteoarthritic knee pain are poorly understood.
Knee OA is a disease of the whole joint featured by structural changes in a number of tissues including cartilage, meniscus, synovium and subchondral bone. These lesions can be identi ed and assessed in a compartment-speci c semi-quantitative way on MRI 11 . Studies on different MRI structural lesions have primarily focused on individual MRI lesions and knee OA outcomes 12 . Focusing on a single MRI feature or structural lesion may be overly simplistic as it does not give a comprehensive picture of the relation between MRI lesions and pain in this whole joint disease. For example, articular cartilage is incapable of directly generating pain because it is aneural. However, cartilage defects, in isolation from other tissues, have been reported to be associated with knee pain in OA 13,14 . Such an association may limit insights into the pathogenesis of osteoarthritic knee pain.
Studying the relation of multiple MRI lesions to knee pain in OA is methodologically challenging. First, many factors that account for an individual's pain response, such as central hypersensitivity, are either not collected or not controlled for in most observational studies 15 ; thus the validity of study ndings related to an individual's pain could be affected by residual confounding. A within-person knee-matched case-control study design has been proposed to eliminate person-level confounders and improve the validity of study ndings 16,17 . Second, to include all MRI lesions into a multivariable regression model and obtain their "independent" effects may be problematic in the absence of knowledge regarding the temporal sequence among the occurrences of MRI lesions. Some effect estimates represent the total effect, and others direct effects, according to the chronology of the occurrence 18,19 . An alternative strategy is to identify the patterns of co-existing MRI lesions and examine their relation to knee pain in OA 20 , which is attractive because model building with many highly correlated predictors can be otherwise nearly impossible. To date, there are no published studies that have utilized these approaches to examine the associations of multiple MRI lesions with subtypes of osteoarthritic knee pain. Therefore, we conducted a within-person knee-matched case-control study to determine if distinct patterns of coexisting MRI lesions have differential associations with knee pain at rest and on joint loading using data collected from the Osteoarthritis Initiative (OAI).

Study sample
The OAI is a longitudinal cohort study of participants with or at high risk of knee OA. At baseline, the OAI cohort included 4,796 subjects aged 45-79 years who were recruited from four sites, Columbus, Ohio, Providence, Rhode Island, Baltimore, Maryland and Pittsburgh, Pennsylvania. Annual assessments included the questionnaires, clinical examination and imaging.
Knee pain at rest and knee pain on joint loading We used the knee-speci c Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) pain sub-scales (0-5) to de ne knee pain at rest and on joint loading. The WOMAC pain sub-scales measure the severity of pain occurring in ve scenarios during the past seven days. For the current study, we used the items of pain with sitting/lying down and pain in bed as two separate measures of knee pain at rest, and those of pain with walking and pain on stairs as measures of knee pain on joint loading.

MRI assessments
MRI readings of structural lesions of knee OA, including cartilage morphology (CartM), bone marrow lesion (BML), synovitis-effusion (EFF), Hoffa's synovitis (HFS), meniscus morphology (MM) and meniscus extrusion (MExt), were performed using the compartment-speci c semi-quantitative MRI Osteoarthritis Knee Score (MOAKS) in the OAI 11 . Scores for BML and CartM were applied in subregions for each lesion in both the tibiofemoral joint (TFJ) and patellofemoral joint (PFJ), denoted by TFJ-BML, PFJ-BML, TFJ-CartM and PFJ-CartM, respectively. We only pooled data from the knee MRI scans that were read by Boston Imaging Core Laboratory group. We restricted our analyses to those who had MRI readings for both knees in one visit.

Statistical analysis
We identi ed a matched pair of case and control knees within one participant if his/her two knees differed by ≥1 point in at least one score of knee pain at rest or on joint loading ( Figure 1). Given that the comparison is between two knees within one person, we consider a discordance of ≥1 point as a meaningful difference which is consistent with the magnitude of acute pain are evoked by an exercise session 21 . Participants were included when their knees formed a matched pair in the earliest visit between baseline and the 48-month visit. Each participant was included only once.
We used the worst CartM score from among 12 subregions in the TFJ (medial and lateral trochlea, medial and lateral central femur, medial and lateral posterior femur; medial anterior tibia, medial central tibia, medial posterior tibia, lateral anterior tibia, lateral central tibia and lateral posterior tibia) and among 2 subregions in the PFJ (medial and lateral patella on the axial view) to represent the severity of cartilage damage in each of the tibiofemoral and patellofemoral joints, respectively. The worst BML score for size from among the aforementioned 12 subregions in the TFJ and among the aforementioned 2 subregions in the PFJ was used to represent the severity of BML in each joint, respectively. The worst MM score from among 6 subregions (medial anterior, body, and posterior; lateral anterior, body, and posterior) and the worst MExt score from among 2 areas (medial and lateral) in 2 views (coronal and sagittal) were used to represent the severity of these lesions in the whole knee.
We performed latent class analysis (LCA) using "gsem" STATA procedure that ts a generalized structural equation model with categorical latent variables to identify subgroups representing distinct patterns of co-existing MRI lesions based on the prevalence and severity of CartM and BML in the TFJ and PFJ, as well as EFF, HFS, MM and MExt in the whole knee. If the proportion of knees with a score level of an MRI lesion was less than 3%, we collapsed those knees to the adjacent lower score level to avoid unstable estimations due to sparse data. We tted the LCA models with 2-6 subgroups and chose the model according to the following criteria: 1) to have a meaningful clinical relevance; 2) the lowest Akaike information criterion or Schwarz's Bayesian information criterion to identify the best model t 22 ; 3) to have su cient numbers of knees (≥10% of the sample) in each subgroup (Table 1). We assigned each knee to the subgroup with the maximum posterior probability generated from LCA model. We then tted a conditional logistic regression model to deal with the dependence of two knees within one subject in a separate analysis of pain at rest and pain on joint loading. We performed test of homogeneity of odds ratios (ORs) using "tabodds" STATA procedure. Subjects who had a TKR or had missing values for variables of interest in either knee were excluded. All statistical analyses were performed using Stata/SE 15.1 (StataCorp, Texas, USA).

Results
Among 130 eligible participants whose MRI-lesion assessments were available, 60% were women, their mean age was 63.8 years, more than 80% of participants were White, and the mean BMI was 29.6 kg/m 2 .
The patterns of co-existing MRI lesions are depicted in Figure 2. In knees of subgroup I, there was no TFJ cartilage damage, whereas all knees of subgroup II had mild TFJ-CartM. All knees of subgroup III and IV had moderate TFJ-CartM, while the prevalence of severe TFJ-CartM was 100% in subgroup V. The prevalence and severity of MM and MExt also increased from subgroup I to subgroup V. Compared with subgroup III, subgroup IV was featured by a higher prevalence and greater severity of EFF, HFS and TFJ-BML. EFF, HFS and TFJ-BML often coexisted and their prevalence and severity were highly correlated. The prevalence and severity of PFJ-CartM and PFJ-BML was not correlated with that of TFJ-CartM. For example, although knees of subgroup V had greater TFJ-CartM than knees of subgroup IV, the prevalence and severity of PFJ-CartM and PFJ-BML was lower in subgroup V. Based on the differences in prevalence and severity of MRI lesions across subgroups, we labeled them as I: minimal lesions, II: mild lesions, III: moderate morphological lesions, IV: moderate multiple reactive lesions and V: severe lesions.

Discussion
Among the OAI participants with or at high risk of knee OA, we identi ed ve distinct patterns of coexisting MRI lesions that were differentially associated with subtypes of knee pain in OA. These ndings provide insights into potential distinct pathologic pathways.
When we compared subgroup III (moderate morphological lesions) with subgroup I (minimal lesions), the results suggested that morphologic lesions of cartilage and meniscus were not signi cantly associated with knee pain at rest or knee pain on joint loading. Aneural hyaline articular cartilage and the meniscus do not generate pain directly, or appear to have a limited role in the pathogenesis of pain. Although there is data that vascular penetration and nerve growth after a meniscal tear may be a source of pain 23 , osteoarthritic knees with a meniscal tear are not more painful than those without a tear 24 . Therefore, meniscal tears are unlikely to be directly linked to knee pain in OA 25 , underlying the large body of evidence that arthroscopic surgery targetting meniscus was no better than conservative interventions such as exercise therapy 26 .
BMLs re ect increased focal loading in the subchondral bone. Synovitis-effusion is thought to be an in ammatory reaction that could be triggered by structural damage. Consistent with the published study 20 , we found that synovitis-effusion and TFJ-BML often coexisted, especially with TFJ-CartM. Meanwhile, the severity of these lesions correlated well in subgroups. In the current study, the difference between subgroup III (moderate morphological lesions) and subgroup IV (moderate multiple reactive lesions) that have the same level of TFJ-CartM mostly lies in the prevalence and severity of EFF, HFS and TFJ-BML. Knees of subgroup IV had a higher risk of greater knee pain at rest and on joint loading than knees of subgroup III, supporting other studies that have reported that synovitis-effusion and TFJ-BML may be the major sources of knee pain in OA 17,27 . Given that pain is the key reason for seeking medical care 28 , these lesions should be considered as clinically relevant markers of symtomatic knee OA. Our ndings highlight the importance of modifying either synovitis-effusion or BMLs to reduce knee pain and consequent pain-related disability.
The role of PFJ in the surgical treatment of knee OA remains controversial [29][30][31][32] . For example, there is evidence that preoperative PFJ OA or anterior knee pain does not compromise the outcome or survival of a medial unicompartmental knee replacement 29,30 . Additionally, at the time of a TKR, patellar retention was not statistically signi cantly associated with the risk of incident postoperative anterior knee pain, compared with patellar resurfacing for patients receiving TKR 31 . Based on latent class analysis, we identi ed the co-existing pattern of BMLs and CartM that are located separately in PFJ and TFJ, allowing us to examine their differential contributions to osteoarthritic knee pain. We found that knees with greater TFJ-CartM could have lower prevalence and severity of PFJ-CartM and PFJ-BML (e.g., subgroup V versus subgroup IV, or subgroup III versus subgroup II). In contrast, subgroup V (severe lesions) had larger ORs of both knee pain at rest and on joint loading than subgroup IV (moderate multiple reactive lesions), suggesting that lesions in PFJ were not signi cantly associated with subtypes of knee pain (e.g., knee pain on stairs). Our ndings, if con rmed, are informative for the management of patella during surgical treatment from the perspective of structural pathogenic source of pain.
In contrast to knee pain at rest that represents pain occurring without mechanical stimuli, knee pain on joint loading is a symptom in response to localized stress. Interestingly, joint loading can act as either a source of pain or a remedy for pain among individuals with knee OA 33 . Animal studies showed that joint loading may attenuate structural damage in OA, depending on the amount and frequency 34,35 . Meanwhile, there is evidence from human studies indicates that exercise, such as land-based training, is effective in the management knee OA 36,37 . In the present study, when we compared the subgroup III (moderate morphological lesions) with the subgroup I (minimal lesions), the ORs of knee pain on joint loading were 0.7 (pain with walking) and 0.8 (pain on stairs) whereas the ORs of knee pain at rest were 2.2 (pain in bed) and 3.1 (pain with sitting/lying down). Such a nding, if con rmed, may add additional evidence that mechanical loading can be protective for patients with early-stage symptomatic knee OA.
Our study has several strengths. We assessed knee pain at rest and knee pain on joint loading using four items of WOMAC pain sub-scale, allowing us to separately look at knee pain in response to different pathogenic sources. Since our study design allowed the assessment of the pain difference in two knees within the same person over a short period of time, we believe that participants should be able to tell a one-point difference. Moreover, in the absence of knowledge of the temporal sequence of MRI lesions, we tted latent class analysis to identify subgroups representing distinct patterns of co-existing MRI lesions. Although our study is subjective to limitations of a case-control study such a selection bias, as an advance over previous studies 38,39 , we were able to examine the differential association of multiple MRI lesions localized in different compartments with knee pain occurring with or without joint loading. Finally, person-level confounders were eliminated using a within-person knee-matched case-control study design; thus, the validity of the study ndings was improved.
Limitations of the present study should be noted. First, although we identi ed patterns of co-existing MRI lesions, we still can't tell the temporal sequence of occurrence of these MRI lesions. For example, TFJ-BML often coexisted with synovitis-effusion in subgroups 20 , making it di cult to sort out their differential contributions to knee pain. Nonetheless, in the absence of an effective structural modifying intervention for knee OA, our ndings are informative for developing structural modifying intervention targeting the structural pathogenic sources of knee pain in OA. Second, uncertainty in subgroup membership assignment might lead to potentially biased effect estimate, albeit the maximum-probability approach. However, the lowest average posterior probability of membership was above 0.92 among all subgroups, suggesting that subgroup assignment was robust. Finally, because we adopted a within-person casecontrol study design, we were not able to examine the association of person-level factors, such as central hypersensitivity and coping strategy, with knee pain in OA.

Conclusions
In conclusion, co-existing patterns of MRI lesions were identi ed and differentially associated with knee pain at rest and on joint loading in OA. Morphological lesions were not statistically signi cantly associated with pain. Synovitis-effusion and TJF-BML were highly correlated and appeared to be the major sources of pain. PFJ lesions had a limited role in the pathogenesis of osteoarthritic knee pain. These ndings are informative for optimizing a treatment strategy targeting the pathogenic sources of osteoarthritic knee pain. Availability of data and materials: The datasets generated and/or analysed during the current study are available in the OAI repository, https://nda.nih.gov/oai.