Pain trajectory defines knee osteoarthritis subgroups: a prospective observational study

Osteoarthritis subgroups/phenotypes, based on pain experience over time, inform on symptom development and delivery of treatment options and open a new avenue toward personalized medicine.


Introduction
Pain is the primary symptom and descriptor of the burden of osteoarthritis (OA), a chronic disease related to substantial disability, morbidity, and costs. 29 According to 2010 estimates, it is globally among the top contributors to disability. 7,22 The socioeconomic burden of OA includes direct expenditure on nonpharmacological and pharmacological treatments, with indirect costs from productivity loss, early retirement, and premature death. 13 For individuals, long-term outcomes include pain, functional limitations of the affected joint, and reduced quality of life. Currently available treatment options-both pharmacological for pain/symptom management and surgical, ie, joint replacement-do not provide significant improvements to all patients. 30,31 The need for successful development of treatment options for all OA patients is currently unmet. Although treatment is expected to halt or minimise OA progression, pain relief is the essential determinant of cost-effectiveness. 8 There have been considerable efforts to develop OA treatment, with many promising candidates failing to reach endpoints in phase-3 clinical studies. This challenge has been explained by heterogeneity, indicating that OA requires personalised medicine. 12,17,28 Thus, to improve drug development and reduce OA burden, proper identification of phenotypesdistinct groups of patients that share the same pathophysiology-is required. 21 These phenotypes will help in selecting patients most likely to benefit from specific treatment options.
In this study, we hypothesised that OA phenotypes could be identified by patients' pain/symptom experiences over time. We focused on the most common one: knee OA. 12 We used a phase-3 clinical trial as a typically selected clinical OA population, and a more extensive prospective cohort study for the external validation. We intended to identify pain patterns/trajectories and to explore the interplay between pain and functional limitation development over time because both are outcomes of interest. To identify if and how phenotypes respond to available pharmacological treatments, we investigated the effect of medication over time. We also studied whether pain trajectory groups are associated with surgical outcomes. Finally, we explored the baseline characteristics associated with each phenotype because these could provide evidence-based recommendations for core phenotyping in personalised medicine and trial recruitment.

Study samples
The Vitamin D Effect on Osteoarthritis (VIDEO) trial was designed to investigate the effect of vitamin D supplementation (daily 800 IU oral cholecalciferol) on knee OA progression. It was a multicentre, 3-year, double-blind, placebo-controlled randomised clinical trial approved, registered, and performed in the United Kingdom (EudraCT: ref.2004-000169-37, ISRCTN94818153, CTA No.11287/0001/001). 2 Participants were included if older than 50 years with radiological evidence of knee OA and knee pain for most days of the month. Exclusion criteria were: morning knee stiffness longer than 30 minutes, secondary or inflammatory arthritis, history of knee surgery or knee replacement in previous 6 months, osteoporotic fractures, and use of bisphosphonates, supplements containing vitamin D, and glucosamine and chondroitin less than 3 months. For further details, see the study by Arden et al. 2 Although an interventional study, for this work, the VIDEO trial was used as an observational typically selected clinical OA study sample.
The Osteoarthritis Initiative (OAI) study is a prospective observational study of knee OA sponsored by the National Institute of Health. Participants age 45 to 79 years were recruited at 4 centres across the United States. Exclusion criteria were: inflammatory arthritis, severe joint space narrowing, bilateral knee replacement or plans for it in the next 3 years, comorbidities that might interfere with participation in this study, participation in clinical trials, and others. Additional study details, as well as data, are available at the https://nda.nih.gov/oai. As of February 2019, data were available through the ninth-year visit.
Participants in both studies provided written informed consents.

Outcome measures
The primary outcome in this study was the pain subscale of the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) 5 -a total score of 5 questions scaled from zero (no pain) to 100 (extreme pain). In the VIDEO trial, pain was assessed for the index knee as previous 48 hour-recall at 6-month intervals (7 repeated measures). In the OAI study, the timeframe of pain assessment was previous 7 days at annual intervals (10 repeated measures). There were reports for the left knee and right knee irrespective of the disease. We assigned a more painful knee throughout the visits for result generalisation.
The secondary outcomes in this study were the functional limitation subscale of the WOMAC, 5 assessed and scaled like the pain subscale (0-100), and surgical outcome, ie, knee replacement. In the VIDEO trial, knee replacement was recorded at the end of the trial as binary outcome. In the OAI study, exact dates of the knee replacements were recorded throughout the follow-up.

Covariates
Baseline variables were used for descriptive purpose and to assess their impact on the pain, as well as confounding variables when the pain was related to the surgical outcome. Age, sex, smoking, alcohol use, employment status, and use of supplements, glucosamine, and chondroitin were self-reported. Body mass index (BMI) was computed based on height and weight measurements. In the VIDEO trial, depression was assessed by Beck Depression Inventory containing 21 questions summed to the total score (0-63). 4 In the OAI study, Centre for Epidemiological Studies Depression Scale with 20 items and a total score 0 to 60 was used. 20 In both questionnaires, a higher score indicates worse symptomatology. In the VIDEO trial only, physical activity (once or more times per month) and quality of life were assessed. The short version of the World Health Organization Quality of Life (WHOQoL-Bref) contains 26 questions measuring 4 domains, physical health, psychological health, social relationships, and environment, each scoring from 0 to 100. Higher scores denote a better quality of life. 10 Comorbidities, defined as none, one, or more than one, were recorded in the OAI study only. A trained orthopaedic fellow or radiologist scored the radiographs according to the Kellgren-Lawrence (KL) grades. 18 Use of currently available medications that affect pain levels, directly or indirectly, ie, analgesics, nonsteroid anti-inflammatory drugs, and steroids, further referred as analgesics, was recorded in both studies throughout follow-up and used in the primary analysis as a binary time-varying covariate. Mortality during follow-up was used for descriptive purposes and sensitivity analyses. Missing values were shown per variable; these were not imputed and were considered for analyses if the percentage was less than 10%.

Statistical analysis
First, we showed baseline characteristics of our study samples for descriptive purposes. Furthermore, we conducted our analyses in 3 steps: identification of trajectories with 2 extensions, investigating the association of pain trajectories with distal surgical outcome, and identification of baseline factors for predicting trajectories.
To identify trajectory groups (latent clusters of individuals) that follow a similar pattern of how pain develops over time, we used group-based trajectory modelling. 23 We used censored normal models with up to a fourth-order polynomial and tested a different number of trajectory groups. Statistical criteria, Bayesian information criteria, 15 and group posterior probability (.0.70) 24 aided in selecting the best model fit. 1,16 We also used the Wald test for equality of trajectory coefficient estimates to confirm that trajectories are distinctive or parallel. 14 Our trajectories were related to the index knee in the VIDEO trial, and more painful knee in the OAI study. After fitting pain trajectories, we included 2 model extensions, dual trajectories and time-varying covariate. 14 Dual trajectory modelling is analysing the developmental course of 2 different but related outcomes. 25 Here, we looked into pain and functional limitations. We modelled functional limitation trajectories in the same manner as pain trajectories. Then, in the dual trajectory model, we examined pain development over time, given the information from function limitation trajectories. This analysis provides conditional probabilities joining membership across the pain and functional limitation trajectory groups. 14 In the last section of trajectory modelling, we included analgesic use as time-varying covariates into pain trajectory model. 14 It is a binary variable because the purpose of this subanalysis was to  find whether currently available analgesics significantly reduced pain over time providing the effect estimates per trajectory groups (the strata of indication severity).
To investigate the association between pain trajectory groups and knee replacement, in the VIDEO trial, we used a logistic regression forward selection method. In the OAI study, we conducted a time-to-event analysis using the Cox proportionalhazards forward selection model.
To identify baseline factors associated with pain trajectory membership and to differentiate each trajectory group, we used a multinomial regression forward selection method. We created several models with different trajectory groups of interest as referenced ones.
As sensitivity analyses, we remodelled pain trajectories excluding mortality cases during the follow-up. Furthermore, in the OAI study, we investigated the left and right knee pain trajectories. Also, dual left and right trajectories, ie, modelling left knee pain development over time having the right knee pain trajectories.

Results
The VIDEO trial included 474 participants, whereas the OAI study had 4796. Table 1 contains baseline characteristics of the study samples. Descriptive statistics, including missing values of pain, functional limitation, and analgesic use variables at every follow-up visit used for the trajectory modelling, are included in Appendix (available as supplemental digital content at http://links.lww.com/PAIN/B88).

Trajectories
We identified 4 pain trajectories described by the first-order curves in the VIDEO trial (Fig. 1A, Table 2). The classification of individuals in pain trajectory groups measured by the posterior probability of membership was very good: ranging from 0.85 to 0.91. The Wald test confirmed that the intercepts of all trajectories were statistically, significantly different from each other. The fourth trajectory was parallel (the slope was not statistically different) to the second trajectory (x 2 5 3.11, P 5 0.08) and third trajectory (x 2 5 0.01, P 5 0.93), whereas all others differed. The second trajectory model included dual trajectories. The best model fit identified 4 first-order curves that described functional limitation development (Fig. 1B, modelling details in Appendix, available as supplemental digital content at http://links.lww.com/ PAIN/B88). Functional limitation trajectories minimally affected the classification of pain trajectory groups in the dual model ( Table 2). Joint probabilities of pain and functional limitation trajectory groups showed that 92.0% of individuals classified in the overlapping groups (Appendix, available as supplemental digital content at http://links.lww.com/PAIN/B88). Finally, in the third trajectory model, analgesic use had a significant positive effect on reducing pain in the first and second trajectory groups and minimally affected the classification of pain groups compared to the primary model ( Table 2).
In the OAI study, we identified 6 trajectories described by higher-order curves: the first and fourth trajectories described by cubic curves, the second one with quartic, and third, fifth, and sixth with quadratic curves (Fig. 1C and Table 2). Posterior probabilities were very good, ranging from 0.80 to 0.90. Intercepts of pain trajectories were different; only group 4 and 5 intercepts were not different (x 2 5 0.50, P 5 0.47). Other parameters were compared between curves of the same order. The cubic curves were not parallel: all parameters were statistically significantly different. The quadratic curves of the fifth and sixth trajectories were parallel (linear component x 2 5 2.70, P 5 0.10; quadratic component x 2 5 2.52, P 5 0.11), whereas the third trajectory was significantly different from these 2. For the second-dual trajectory model-we identified 6 functional limitation trajectories (Fig. 1D, modelling details in Appendix, available as supplemental digital content at http://links.lww.com/ PAIN/B88). Additional information on functional limitation Analgesics' use had a significant positive effect in all trajectory groups except the sixth and affected the classification of fourth and fifth pain groups.

Risk for the distal outcome
We found that no single pain trajectory group in the VIDEO trial was significantly associated with index knee replacement during the 3-year follow-up. In the OAI study, third, fourth, and fifth groups but not the sixth had significantly higher hazard ratio of  having knee replacement during the 9-year follow-up when compared to the second group (Table 3), and all groups when referenced to the first group (Appendix, available as supplemental digital content at http://links.lww.com/PAIN/B88). The highest hazard ratio was in the fourth group. Table 4 contains baseline characteristics of the trajectory groups from both studies.

Descriptors of the pain trajectories
In the VIDEO trial, we created 2 models using the first and fourth trajectory groups as references. The higher BMI and the lower physical domain of WHOQoL-Bref were associated with being in all groups compared to the first group. The higher psychological domain of WHOQoL-Bref was associated with membership of the second and third trajectory groups. The second model aimed to distinguish the third and fourth groups: however, no single analysed variable showed a significant result (Appendix, available as supplemental digital content at http:// links.lww.com/PAIN/B88).
In the OAI study, the first, fourth, and sixth groups were of interest, and we created 3 models using each as a reference. Members of any painful group, compared to those with minimal pain, were significantly more likely to be younger women with higher BMI, depression score, KL grade 2 or more, and using painkillers. The membership in the fifth group compared to the fourth group was only negatively significantly associated with KL grade 1. Finally, older age, lower BMI, lower depression score, and use of supplements were significantly associated with the fourth and fifth groups compared to the sixth group. Other variables showed limited potential in distinguishing the pain groups (Appendix, available as supplemental digital content at http://links.lww.com/PAIN/B88).

Sensitivity analyses
Remodelled pain trajectories without mortality cases were the same as in the original model (Appendix, available as supplemental digital content at http://links.lww.com/PAIN/B88). In the OAI study, we identified 6 left and 6 right knee pain trajectories, described by higher-order curves like the original/generalised pain trajectory model. In the dual trajectory model, left knee pain development slightly changed in terms of group percentage and posterior probabilities, given the additional information on right knee pain trajectories. Joint probabilities showed that 64.2 of individuals were classified in the overlapping left and right knee pain groups (Appendix, available as supplemental digital content at http://links.lww.com/PAIN/B88).

Discussion
We identified knee OA subgroups/phenotypes based on pain trajectories. The number of trajectory groups, their size, and pattern of development differed due to study inclusion criteria, sample size, and follow-up duration of the VIDEO trial and the OAI study. However, due to observable similarities, we identified 4 OA phenotypes from these: low-fluctuating, mild-increasing, moderate-treatment-sensitive, and severe-treatment-insensitive pain. We found that pain and functional limitations in OA measured by WOMAC questionnaire showed the same development over time. Importantly, we identified a phenotype with severe pain that did not benefit from analgesics and had the same chance for knee replacement as the low-fluctuating phenotype. We also identified a subgroup most likely to benefit from knee replacement. Finally, using baseline factors, we were able to distinguish painful from minimally painful groups but found little to differentiate moderate from severe pain groups.
We used 2 high-quality studies of different design, size, and follow-up duration to overcome some of their complementary drawbacks. Unlike some previous studies, ours did not use any method to additionally select/match participants or make the 2 studies more similar. 6,26 Instead, we relied on a method that selected latent classes for dealing with heterogeneity. 23 In groupbased modelling, we permitted small groups to be detected when the model fit supported it. We also managed to replicate the small-sized groups, reducing the chance of spurious classes' detection. Although studies had different intervals of the outcome assessment, it provided additional evidence of the consistency and robustness of the findings. Both studies involved OA patients at different disease stages. The baseline in each case was related to the study; it is not the disease baseline because there is still a lack of OA onset definition in the field overall. 19 However, our studies had different follow-up durations, allowing us to observe time-effects more comprehensively. However, due to different Table 3 Pain trajectories as predictors of knee replacement. All models were constructed using forward selection procedure. In the VIDEO trial, variables included in the selection were pain trajectory group, treatment, vitamin D, interaction treatment and vitamin D, age, sex, body mass index, smoking, alcohol drinking, currently working, physical activity, Beck depression scale, physical, psychological, social, and environment domains of the quality of life, Kellgren-Lawrence grade of the index knee at baseline, use of medications, and use of supplements at baseline. In the OAI study, variables included in the model were pain trajectory group, age, sex, body mass index, smoking, alcohol drinking, currently working, Centre for Epidemiological Studies Depression score, comorbidities, Kellgren-Lawrence of the examined knee at baseline, knee replacement at baseline, use of medications, and use of supplements at baseline. The estimates presented were from the final models. Number of observations used in the VIDEO trial models was 425, and in the OAI study without trajectory group 1, it was 2827 for the left knee, 3049 for the right knee, and 3348 for the generalised model, depending on the missing values of Kellgren-Lawrence grade at baseline. * The model was constructed using binary logistic regression model and included pain trajectory group (forced entry for the report), vitamin D main effect, and currently working. † The models was constructed using Cox proportional hazards model, and all 3 included pain trajectory group, age, Centre for Epidemiological Studies Depression score, Kellgren-Lawrence of the examined knee at baseline, use of medications and use of supplements at baseline; in addition to these, left knee model included also sex and comorbidities, right knee model smoking, and the generalised model knee replacement at baseline. CI, confidence interval; HR, hazard ratio; OAI, Osteoarthritis Initiative; OR, odds ratio.  Table 4 Baseline characteristics of the pain trajectory groups.    follow-up durations, we did not directly compare trajectory groups between the samples. Although different inclusion and exclusion criteria were used to some extent, samples were similar in terms of demographic and lifestyle factors. As expected, trial participants were more severe in clinical and radiographic aspects. We used a set of baseline characteristics that did not entirely overlap. Although this was a study limitation, it represented the diversity of OA covariate measures used. There were 4 phenotypes identified in the trial and replicated in the OAI study. We named phenotypes by indicating baseline pain and its pattern or responsiveness to treatments. The minimal pain trajectory group from the OAI study represented people with minimal-to-neglected knee pain, thus not assumed an OA phenotype. These were sampled in the cohort but not the trial. The first trajectory group selected in the VIDEO trial and the second trajectory group in the OAI study presented the lowfluctuating phenotype. In the short term, it showed pain improvement but fluctuated in the long term, albeit staying quite low. This phenotype included slightly more than a third of both samples and reflected reports from previous studies. 3,6,26,27 The second trajectory group in the trial and the third in the cohort represented the second mild-increasing phenotype (also reported before). 6,26 This phenotype is the only more common one in the trial than in the cohort. The remarkable observation was related to the third moderate-treatment-sensitive phenotype: third trajectory group in the trial and the fourth and fifth groups in the cohort. In a 3-year window, this phenotype presented moderate-increasing pain in both studies. However, in longer term, this phenotype divided into 2 subgroups: in one, patients benefited from knee replacement (fourth trajectory), whereas in another, patients continued to experience moderate pain despite significant analgesic effects (fifth trajectory). This phenotype included 15% to 20% of the samples. Similar observations were found in studies using the CAS-K 26 and CHECK 3 cohorts, but not in the 5-trajectory model previously identified in the OAI study. 6,26 However, previous studies did not examine treatment effects on pain trajectories. Finally, the fourth phenotype included 3% of both our samples. It was also shown earlier 3,6,26 and here additionally described by severe-treatment-insensitive pain (fourth group in the trial and sixth group in the cohort).
Furthermore, we found that functional limitations followed identical development to pain, indicating that people experienced these 2 outcomes very similarly or could not distinguish between them. Some studies examined functional limitation in knee OA, 11,32 but none of these did not look into interplay between pain and functional limitation over time. Although the first 3 phenotypes were responsive to currently available analgesics, the fourth was not. OA phenotypes were not significantly associated with knee replacement during a 3-year follow-up because only a few replacements took place during the period. However, in the longer term-as the number of replacements increased-mildincreasing and moderate-treatment-sensitive phenotypes but not severe-treatment-insensitive had significant odds of having knee replacement when compared to the mild-fluctuating one. The severe-treatment-insensitive phenotype in the OAI study included the youngest women with the highest BMI and depression score, with more comorbidities, using analgesics, but without pain relief. The left and right knee trajectories and their dual model showed that most people develop the same pain pattern irrespective of laterality, probably due to central pain processing. The overlapping group percentage was likely lower due to individuals' unilateral pathologies.
The baseline clinical and lifestyle factors in our study were modest in differentiating the phenotypes. Overall, the variable with a consistently positive relationship to pain seemed to be BMI, indicating metabolic differences between phenotypes. The age effect was transposed between study samples. As previously discussed, the exclusion criteria in the OAI study likely led to a healthier population being selected. 6 More generally, age effect can be a random observation unrelated to pain phenotypes, instead indicating time-dependent exposure and a molecular process to be detected. Interestingly, 75% of the fourth phenotype were women, indicating a sex-specific mechanism of severe-treatment-insensitive pain. Finally, the 9-year follow-up is the longest regular/annual follow-up so far in knee OA, and it has given us better insights into long-term pain, the slowprogressing character of OA, and its relation to OA end-stage.
To conclude, our approach provided robust results regarding pain experience for OA patient phenotyping with clinical, research, and trial-design relevance. Pain should remain the primary outcome under investigation because functional limitations do not add information. Besides pain duration, we should also consider pain intensity. The cutoff for inclusion in clinical trials should be pain intensity above 20%, and for the sensitivity analyses above 50% of the scale. This range is also the indicator for delivery of currently available treatments. Patients experiencing pain above 50% of the scale need novel pharmacological treatments and careful consideration of safety issues due to comorbidities. Due to the reproducibility between study designs, it creates a template for reanalysing available longitudinal data pools with further characterisation. To improve phenotype differentiation beyond this report, we suggest using molecular and genetic tools 9 that should provide inside into dysregulated molecular pathways to target. Then, pain with additional tools will lead to an optimal set of criteria for selecting patients for treatment options and future OA clinical trials.