Pain trajectory groups in persons with, or at high risk of, knee osteoarthritis: findings from the Knee Clinical Assessment Study and the Osteoarthritis Initiative

Summary

Objective

The authors aimed to characterize distinct trajectories of knee pain in adults who had, or were at high risk of, knee osteoarthritis using data from two population-based cohorts.

Method

Latent class growth analysis was applied to measures of knee pain severity on activity obtained at 18-month intervals for up to 6 years between 2002 and 2009 from symptomatic participants aged over 50 years in the Knee Clinical Assessment Study (CAS-K) in the United Kingdom. The optimum latent class growth model from CAS-K was then tested for reproducibility in a matched sample of participants from the Osteoarthritis Initiative (OAI) in the United States.

Results

A 5-class linear model produced interpretable trajectories in CAS-K with reasonable goodness of fit and which were labelled “Mild, non-progressive” (N = 201, 35%), “Progressive” (N = 162, 28%), “Moderate” (N = 124, 22%) “Improving” (N = 68, 12%), and “Severe, non-improving” (N = 15, 3%). We were able to reproduce “Mild, non-progressive”, “Moderate”, and “Severe, non-improving” classes in the matched sample of participants from the OAI, however, absence of a “Progressive” class and instability of the “Improving” classes in the OAI was observed.

Conclusions

Our findings strengthen the grounds for moving beyond a simple stereotype of osteoarthritis as “slowly progressive”. Mild, non-progressive or improving symptom trajectories, although difficult to reproduce, can nevertheless represent a genuinely favourable prognosis for a sizeable minority.

Introduction

Knee osteoarthritis can be considered from two perspectives: structural damage, and pain and functional limitation¹. The two perspectives are linked but do not match up one-to-one in individual patients, either cross-sectionally or over time¹. Regardless of this, the popular image of knee osteoarthritis is of a slowly but inexorably progressive, “degenerative” condition ending in the need for total joint arthroplasty. Indeed, when viewed as a single population, the ‘average prognosis’ of symptomatic knee osteoarthritis is characterised by small annual losses in joint space narrowing² and a generally unfavourable long-term course of self-reported pain and disability³. Yet recent longitudinal studies have reported high inter-individual differences in the course of osteoarthritis structural progression, symptoms, and functional limitation^4–8 raising the possibility of an alternative view of the condition: that it comprises an uncertain number of groups with distinct long-term trajectories, not all of which may be progressive. This is plausible given a range of potential underlying phenotypes including clinically documented aggressive forms of osteoarthritis^9,10, a realisation that structural disease progression may be averted through slow but successful repair, at least in early phases^11,12, and evidence of complex processes of psychological and behavioural adjustment to chronic pain^13,14.

Six separate studies of prevalent cases of arthralgia or osteoarthritis of the hip^15,16, knee^17,18,20 or mixed hip and knee¹⁹ have used either two-step cluster analysis¹⁹ or latent class growth modelling^15–18,20 to fit trajectory groups to repeated measures of pain^15,16,20, disability^18–20 or joint space width¹⁷, over periods of observation ranging from two^15,17 to six^19,20 years. Diverse findings are normally to be expected from such methods, but this has been particularly evident due to the heterogeneous design of these studies. While being mindful of Nagin & Odgers'²¹ caution against the “quixotic quest to identify the true number of groups”, attempting to replicate findings in separate populations is a cornerstone in other forms of prognosis research²² and in epidemiologic research more generally²³. For the purposes of this paper we focused on the clinical syndrome of OA, in which use-related pain is the cardinal feature. Using samples from two cohorts in the United Kingdom and the United States, we have sought not only to derive developmental trajectories of pain severity in persons with, or at high risk of, knee osteoarthritis but also to investigate the extent to which they could be reproduced in a separate population with careful matching and similar data collection procedures.

Methods

Study population

Self-report, clinical assessment, and plain radiography data from two prospective cohort studies – the Knee Clinical Assessment Study (CAS-K) and the Osteoarthritis Initiative (OAI) – were used. The design and methods of the CAS-K and OAI have been described in detail elsewhere^24–26.

Briefly, CAS-K is a prospective cohort study of knee pain in adults over the age of 50 years. Between 2002 and 2003 individuals reporting knee pain in the previous 12 months were identified via a postal survey sent to all adults aged 50 years and over registered with three general practices in North Staffordshire, United Kingdom. Of 2226 respondents reporting knee pain and consenting to further contact 819 attended a research clinic at baseline for more detailed assessment, including plain radiography of the knees. Participants were followed up by postal survey at 18, 36, 54, and 72 months (n = 776, 707, 602 and 512 respondents respectively). In addition, 95% of eligible participants at baseline consented to medical record review and all participants were offered repeat clinical and radiographic assessments at 36 and 72 months. Full ethical approval was gained prior to participant recruitment and participant consent was gained for each study stage.

The OAI is a publicly and privately funded prospective longitudinal cohort study of community-based persons aged 45–79 years with, or at high risk of, developing clinically significant knee OA at baseline, and approved by the Institutional Review Board at the OAI Coordinating Center, the University of California, San Francisco. Participants were recruited between 2004 and 2006 via telephone screen of 17,457 adults following focussed mailings, advertisements, and community presentations (see Web Appendix 1 for study exclusion criteria). 4796 eligible participants attended an enrolment visit and were invited to attend yearly follow-up research clinics for a period of 6-years (n = 4496, 4324, 4270, 4256, 3935, 3823 respectively, with data collected at clinic or by telephone if clinic attendance was not possible). For the current analysis, a sample of symptomatic OAI participants was taken to match those in CAS-K (further details in Statistical analysis). The data for the OAI cohort are available in the public domain and can be downloaded from the OAI database, which is available for public access at http://www.oai.ucsf.edu/. Specific datasets downloaded for the current study are listed in Web Appendix 1, available at http://www.journals.elsevier.com/osteoarthritis-and-cartilage/

Measurements

Severity of knee pain on activity was measured by self-report on each occasion and in both cohorts using the five-item Pain subscale of the Western Ontario and McMaster Universities Arthritis Index (WOMAC²⁷). We focussed on knee pain severity as our outcome of interest as it is of importance to patients and we have previously argued that pain is inherently easier to attribute to the knee than functional limitation^28,29, although both WOMAC Pain and Function scores may follow very similar trajectories over time²⁰. The WOMAC was used to measure knee pain severity as it has been shown to be a valid and reliable measure of knee pain severity on activity and is extensively used in other studies³⁰. It asks about severity of knee pain on activities “in the last 48 h” with response options “none”, “mild”, “moderate”, “severe” or “extreme” (total score 0–20). In the OAI the timeframe was amended to “in the last 7 days”. The questions were completed separately for each knee in the OAI and we took the maximum sub-scale score for the left and right knee to make it comparable to the person-level WOMAC score in CAS-K.

The following variables, recorded at baseline, were used for matching: age, sex, WOMAC Pain, WOMAC Function (0–68), body mass index (BMI) derived from participants' measured height and weight, and Kellgren–Lawrence grade (0–4) for tibiofemoral osteoarthritis in the most affected knee obtained from weight-bearing, bilateral, posteroanterior projection views on plain radiography using standardised protocols (the most affected knee was respondent-nominated in CAS-K and was researcher-defined in the OAI as the knee with the highest WOMAC pain score. For both cohorts, the most affected knee was selected at random when both knees were of equal severity).

In CAS-K a self-reported measure of locomotor disability was used to explore the construct validity of the trajectories extracted (it was collected at baseline and 6-year follow-up and combines three questions on walking and two questions on stair climbing from the SF-36 Physical Functioning subscale^31,32. It is scaled −4.906 (less disability) to 5.035 (more disability); converted to 0 to 9.941 for analysis). In addition, the following variables were used in CAS-K to describe the baseline characteristics of participants in each of the trajectory groups: area-level index of multiple deprivation³³, social class^34,35, employment status, going on to full time education after school, general health (measured on a likert scale from “Excellent” to “Poor”³¹), anxiety and depression (as measured by the Hospital Anxiety and Depression Scale – scaled 0–21 for anxiety and pain respectively³⁶), widespread pain³⁷, pain/aching/stiffness in the knee in the last month, knee pain severity at the present time (scaled 0–10), time of onset, severity of tibiofemoral and patellofemoral joint radiographic OA³⁸ and a prognostic score based on original work by Von Korff and Miglioretti (2005)³⁹ and previously applied to the CAS-K cohort⁴⁰.

In both CAS-K and OAI the receipt of total knee arthroplasty during the follow-up period was identified through a combination of medical record review, study radiographs taken at follow-up visits, and self-reports from each follow-up questionnaire. Self-reported treatment data were also collected in both cohorts to indicate treatments used at baseline and in the month preceding each follow-up visit (CAS-K: knee-specific treatments used for any duration; OAI: arthritis or joint pain treatments used for more than half the days in the last month). Any health care consultations for knee pain were also recorded in CAS-K over the 6-year follow-up period with general practitioner (GP) consultations measured by medical record review and physiotherapist/hospital specialist consultations by self-report using a time frame at each data collection point of “since your baseline assessment”.

Statistical analysis

Our analysis was conducted in four steps: identification of distinct, interpretable trajectories of knee pain severity and assessment of construct validity of the trajectories within the CAS-K cohort; selection of a sample of OAI participants matched on key baseline characteristics to CAS-K participants; attempted reproduction of the number and pattern of trajectories observed in CAS-K within the matched OAI sample; comparison of the baseline characteristics and description of health care use between different trajectories within the same cohort and between similar trajectories in CAS-K and OAI. The four steps were completed after excluding any WOMAC Pain scores collected post total knee arthroplasty in either knee.

Latent class growth modelling to identify trajectories of knee pain severity in CAS–K participants

Latent class growth analysis is a semi-parametric statistical technique that aims to model individual trajectories over time using polynomial curves after accounting for unobserved heterogeneity within the population^41,42. A search for the optimum latent class growth model was conducted in CAS-K by fitting a one-class linear model to the data and then successively increasing the number of classes until model fit no longer improved. Model fit was based on statistical criteria (see Web Appendix 2) and on a judgement of model interpretability as recommended by Jung et al.⁴¹. We defined model interpretability in this context to be a model that showed at least one group with a significant change in their knee pain trajectory over time. We used this definition to ensure that our latent class growth model provided more understanding of our data than could be gained from a random effects model fitted to the sample as a whole. We were mindful, however, to ensure that statistical fit was not greatly reduced for this criteria to be satisfied. A series of sensitivity analyses were conducted on CAS-K and OAI to evaluate the assumption of linear trajectories, and to examine the impact of excluding the baseline measurement, and varying the number of non-missing measurement points required for the participant to be included in the analysis (Web Appendix 3).

To address construct validity, change in locomotor disability from baseline to 6-year follow-up was described for each trajectory group from the optimum model in CAS-K. It was hypothesised that trajectory group membership would be related to change in locomotor disability although specific hypotheses around the nature of this relationship were not stated until the number and form of the trajectory groups were known.

Selection of a matched sample of OAI participants

Propensity score matching⁴³ was used to obtain a sample of OAI participants matched to CAS-K participants on the following baseline characteristics: age, gender, BMI, WOMAC Pain and Function scores and Kellgren–Lawrence score in the tibiofemoral joint of the patients' most problematic (index) knee. This approach was taken so that model reproducibility could be tested after controlling for (as far as possible) any baseline imbalance between the cohorts (for further details see Web Appendix 4).

Knee pain severity trajectories in OAI participants

A latent class growth model with the optimum number of classes and polynomial form (as derived from the CAS-K data) was applied to the matched OAI data. Regression coefficients from the matched OAI model were compared to those from CAS-K after regression coefficients had been converted into per-year rates of change in WOMAC pain for ease of comparison. Goodness-of-fit indices (as described in Web Appendix 2) were compared for models with $\tilde{k}$ classes, $\tilde{k}+1$ classes, and $\tilde{k}-1$ classes (where $\tilde{k}$ is the optimum number of classes in CAS-K) to see if a model with $\tilde{k}$ classes was indeed optimum for the OAI data.

Descriptive characteristics and health care use

Descriptive statistics were used to compare the baseline characteristics of participants in the derived trajectory groups within CAS-K and multinomial logistic regression was used to test whether such differences were statistically significant. The rates of health care use in each cohort were also described using numbers and percentages.

Results

Study population

A total of 570 CAS-K participants (mean (SD) age: 64 (8.0) years, 54% female) were eligible for inclusion after excluding 16 cases with an existing diagnosis of inflammatory arthritis at baseline, 213 with WOMAC Pain data missing at baseline or available at fewer than two follow-up points, and 20 missing data on the matching variables. Participants excluded from the analysis were older, had more severe knee pain, more functional difficulty and greater evidence of tibiofemoral osteoarthritis at baseline than those included in the analysis (a finding that could have occurred as inclusion in the analysis was influenced by excluding participants' WOMAC data post total knee arthroplasty) (Web Table 1). We have previously shown that attendees at the clinical assessment are largely representative of the population of older adults with knee pain²⁶.

Identification of pain trajectories in CAS-K

A linear 1-class latent class growth model showed an average increase in WOMAC Pain score of 0.08 points per year over the 6-year period (95% confidence interval: 0.02, 0.14). A 4-class linear model improved model fit, but with classes differentiated only by baseline score and not by slope (intercepts 2.4, 6.0, 10.4 and 15.3 respectively). The 5-class linear model produced interpretable trajectories and goodness-of-fit statistics that were not greatly inferior to the 4-class model, and was considered optimal (Table I). The trajectory groups were labelled on the basis of their intercepts and slopes as: “Mild, non-progressive” (N = 201, 35%), “Progressive” (N = 162, 28%), “Moderate” (N = 124, 22%) “Improving” (N = 68, 12%), “Severe, non-improving” (N = 15, 3%) (Fig. 1).

Table I.

Goodness-of-fit statistics for linear WOMAC pain models

Classes	AIC	BIC	ABIC	Entropy	VLMR LRT	Adjusted LMR LRT	PB LRT	Class N	Average posterior probability
CAS-K
1	14,682	14,713	14,690					570	1.0
2	13,691	13,734	13,703	0.83	P < 0.001	P < 0.001	P < 0.001	346, 224	0.96, 0.94
3	13,361	13,417	13,376	0.83	P < 0.001	P < 0.001	P < 0.001	215, 120, 235	0.94, 0.94, 0.90
4	13,263	13,333	13,282	0.86	P < 0.001	P < 0.001	P < 0.001	226, 14, 211, 119	0.90, 0.93, 0.94, 0.93
5	13,235	13,318	13,258	0.79	P = 0.206	P = 0.220	P < 0.001	201, 68, 15, 162, 124	0.95, 0.70, 0.95, 0.80, 0.91
6	13,223	13,318	13,249	0.75	P = 0.155	P = 0.165	P = 0.013	46, 126, 14, 81, 109, 194	0.69, 0.72, 0.93, 0.70, 0.91, 0.93
OAI
4	18,059	18,137	18,080	0.86	P = 0.249	P = 0.260	P < 0.001	251, 44, 118, 157	0.95, 0.92, 0.91, 0.87
5	18,004	18,095	18,029	0.83	P = 0.207	P = 0.214	P < 0.001	235, 26, 106, 139, 64	0.94, 0.89, 0.85, 0.83, 0.85
6	17,961	18,065	17,989	0.82	P = 0.037	P = 0.041	P < 0.001	80, 27, 232, 63, 30, 138	0.81, 0.90, 0.94, 0.84, 0.81, 0.81

Abbreviations: AIC, Akaike Information Criteria; ABIC, Sample-size adjusted BIC; BIC, Bayesian Information Criteria; LMR LRT, Lo-Mendell-Rubin likelihood ratio test; PBLRT, parametric bootstrapped likelihood ratio test; VLMR LRT, Vuong-Lo-Mendell-Rubin likelihood ratio test.

Fig. 1.

WOMAC Pain Scores by Trajectory Group Membership for (A) CAS-K and (B) Matched OAI Sample (N = 570). Abbreviations: PYRC = Per-year rate of change in WOMAC points; 95% confidence interval in brackets.

In support of construct validity, participants in the “Progressive” group showed locomotor disability change scores that were significantly different from those in the “Mild, non-progressive” group i.e., a greater rate of deterioration in the “Progressive” group when compared to the “Mild, non-progressive” group (Table II). In contrast, despite change in locomotor disability being greater in the “Improving” group than in the reference group, this difference was not statistically significant, so evidence of construct validity was lacking for this comparison.

Table II.

Change in locomotor disability by trajectory groups in CAS-K

	Change in disability
Mild, non-progressive (n = 141)	0
Progressive (n = 119)	−0.60 (−1.13, −0.07)
Moderate (n = 78)	−0.40 (−1.00, 0.20)
Improving (n = 44)	0.50 (−0.23, 1.24)
Severe, non-improving (n = 10)	0.34 (−1.06, 1.73)

Change in disability = Change in locomotor disability (baseline – 6-year follow-up) i.e., a negative score on the outcome represents a worsening of locomotor disability over the 6-year follow-up time-period. Figures are regression coefficients and 95% confidence intervals from an analysis of variance model with change in loco-motor disability as the dependent variable.

Reproducibility of pain trajectories in OAI

A matched sample of 570 OAI participants was drawn from a total pool of 3315 eligible participants. The characteristics of the matched OAI data and CAS-K were similar (Table III) and the median and interquartile range of the propensity score difference was zero.

Table III.

Baseline characteristics of participants in the CAS-K analysis cohort (United Kingdom (2002–2003)) and the matched OAI sample (United States (2004–2006))

	CAS-K (N = 570)	OAI (N = 570)
Age (years): mean (SD)∗	64 (8.0)	64 (8.5)
Female gender∗	309 (54)	296 (52)
BMI: mean (SD)∗	29.5 (5.0)	29.4 (4.9)
Time since problem† onset
No knee pain for most days in a month		143 (25)
<1 year	57 (10)	63 (11)
1–5 years	201 (35)	160 (28)
5+ years	312 (55)	204 (36)
WOMAC function27: median (IQR)∗	17.0 (6.0, 30.0)	18.0 (5.0, 30.0)
Kellgren–Lawrence score in tibiofemoral joint∗
0	308 (54)	298 (52)
1	54 (9)	51 (9.0)
2	82 (14)	96 (17)
3	70 (12)	63 (11)
4	56 (10)	62 (11)
Baseline exclusion criteria in OAI
Bilateral total knee arthroplasty	0 (0)	0 (0)
Bilateral severe joint space narrowing	22 (4)	0 (0)
Unilateral total knee arthroplasty and contralateral severe joint space narrowing	0 (0)	0 (0)
Aged 80 years and over	21 (4)	0 (0)
Excessive weight for magnetic resonance imaging scan‡	10 (2)	0 (0)
WOMAC pain27: median (IQR)
Baseline∗	5.0 (2.0, 9.0)	5.0 (2.0, 9.0)
12-months	N/A	3.0 (1.0, 7.0)
18-months	5.0 (2.0, 9.0)	N/A
24-months	N/A	3.0 (1.0, 7.0)
36-months	6.0 (2.0, 10.0)	3.0 (0.0, 7.0)
48-months	N/A	3.0 (0.0, 7.0)
54-months	5.0 (2.0, 9.0)	N/A
60-months	N/A	3.0 (0.0, 7.0)
72-months	6.0 (3.0, 10.0)	3.0 (0.0, 7.0)

Abbreviations: IQR, Inter-quartile range; N/A, Not applicable; SD, standard deviation; WOMAC, Western Ontario & McMaster Universities Osteoarthritis Index. Numbers are N (%) unless otherwise stated.

^∗Variable used in matching.

^†For CAS-K, onset relates to the current knee problem, for OAI, onset relates to knee pain, aching or stiffness experienced on most days or more in a month for either knee.

^‡Males >130 kg, females >114 kg.

A 5-class linear latent class growth model was applied to the OAI data and fitted the matched OAI data well (all posterior probabilities >0.7 and entropy = 0.83), with fit indices similar to those derived from a 4- or 6-class model (Table I). The model supported the existence of “Mild, non-progressive” and “Severe, non-improving” classes (presenting with similar prevalence to CAS-K). However, the “Mild, non-progressive” class in the OAI includes some participants with relatively high baseline WOMAC pain scores that are then reduced and maintained at a lower level for all subsequent time points (Fig. 1). A “Progressive” subgroup did not emerge in the OAI data and three subgroups of “Moderate” trajectories were apparent in the OAI data, compared to only one in CAS-K.

Participant baseline characteristics and health care utilisation

The “Severe, non-improving” group of participants are characterised by having the highest percentage of females, lowest social class, highest levels of anxiety and depression, poorer general health, worst pain and function, and more evidence of severe radiographic OA (Table IV). However this group does not have longer duration of knee problems. All baseline characteristics showed differences that were statistically significant between the trajectory groups (P < 0.05).

Table IV.

Baseline characteristics and health care use over the 6-year follow-up period for CAS-K (N = 570)

Characteristic	Mild, non-progressive N = 201	Progressive N = 162	Moderate N = 124	Improving N = 68	Severe, non-improving N = 15
Baseline characteristics
Age∗	63 (8.1)	65 (7.4)	66 (8.4)	64 (8.5)	64 (6.7)
Female gender	108 (54)	94 (58)	70 (56)	25 (37)	12 (80)
BMI∗	27.7 (4.2)	29.5 (4.5)	31.7 (5.6)	30.1 (5.0)	31.6 (5.5)
Index of multiple deprivation33,∗	14,497 (8138)	12,112 (7430)	11,063 (7414)	12,514 (7601)	9083 (4870)
Manual occupation34,35	81 (45)	91 (61)	74 (69)	36 (62)	11 (79)
Currently employed	65 (34)	37 (23)	14 (12)	14 (21)	0 (0)
Went on to full time education after school	49 (25)	13 (8)	11 (9)	9 (14)	0 (0)
Overall general health
Excellent	16 (8)	0 (0)	1 (1)	2 (3)	0 (0)
Very good	56 (28)	23 (14)	9 (7)	10 (15)	0 (0)
Good	97 (49)	94 (59)	37 (30)	30 (44)	3 (20)
Fair	28 (14)	37 (23)	60 (49)	19 (28)	10 (67)
Poor	2 (1)	6(4)	16 (13)	7 (10)	2 (13)
HADS – anxiety36,∗	5.4 (3.4)	6.8 (4.0)	8.1 (4.1)	6.7 (5.0)	9.3 (2.0)
HADS – depression36,†	2.0 (1.0, 4.0)	4.0 (2.0, 6.0)	6.0 (4.0, 9.0)	4.0 (2.0, 7.0)	7.0 (6.0, 11.0)
Widespread pain37	21 (10)	36 (22)	49 (40)	15 (22)	8 (53)
Locomotor disability score31,32,∗	2.2 (1.8)	4.0 (2.2)	6.3 (2.6)	4.3 (2.9)	7.9 (1.2)
Pain, aching or stiffness in the knee in the last month
No days	32 (16)	6 (4)	1 (1)	1 (1)	0 (0)
Few days	79 (39)	41 (25)	8 (6)	12 (18)	0 (0)
Some days	58 (29)	51 (31)	24 (19)	21 (31)	1 (7)
Most days	28 (14)	46 (28)	54 (44)	27 (40)	5 (33)
All days	4 (2)	18 (11)	37 (30)	7 (10)	9 (60)
Knee pain severity at the present time†	1.0 (0.0, 2.0)	3.0 (1.0, 4.0)	5.0 (4.0, 7.0)	3.0 (1.0, 5.0)	8.0 (7.0, 8.0)
WOMAC function27,†	5.0 (1.0, 10.0)	17.5 (10.0, 23.0)	35.0 (29.9, 41.0)	25.0 (18.5, 34.0)	50.0 (44.0, 55.0)
Onset of knee problem
<1 year	33 (16)	12 (7)	2 (2)	9 (13)	1 (7)
1-year to 5-years	70 (35)	67 (41)	36 (29)	21 (31)	7 (47)
5-years or more	98 (49)	83 (51)	86 (69)	38 (56)	7 (47)
Severity of tibiofemoral joint radiographic OA38
No definite radiographic OA	135 (67)	85 (52)	61 (49)	43 (63)	5 (33)
Mild radiographic OA	42 (21)	29 (18)	21 (17)	11 (16)	4 (27)
Moderate/severe radiographic OA	24 (12)	48 (30)	42 (34)	14 (21)	6 (40)
Severity of patellofemoral joint radiographic OA38
No definite radiographic OA	103 (51)	51 (31)	32 (26)	21 (31)	2 (13)
Mild radiographic OA	71 (35)	73 (45)	57 (46)	34 (50)	6 (40)
Moderate/severe radiographic OA	27 (13)	38 (23)	35 (28)	13 (19)	7 (47)
Chronic pain risk score39,†	5.0 (3.5, 8.0)	10.0 (7.0, 13.0)	17.0 (11.0, 21.0)	11.0 (8.0, 14.0)	22.5 (21.0, 23.0)
Health care use
Total knee arthroplasty received during 6-yr follow-up period	2 (1)	4 (2)	8 (6)	1 (1)	3 (20)
Average‡ treatment use prevalence over the five study time points (%)
Paracetamol	15%	44%	42%	28%	36%
NSAIDS§	16%	30%	36%	24%	31%
Other analgesia‖	10%	36%	55%	20%	73%
Creams, gels or rubs	15%	40%	49%	28%	73%
Exercise for knee pain	8%	16%	25%	13%	36%
Dieting to lose weight	14%	28%	35%	30%	46%
Treatment or consultation received at any point during the 6-yr follow-up period
Knee injection¶	6 (4)	10 (9)	16 (20)	1 (2)	5 (45)
Physiotherapist	28 (20)	45 (38)	43 (49)	14 (31)	8 (62)
Hospital specialist	17 (12)	31 (26)	47 (52)	13 (28)	7 (58)
GP#	65 (34)	100 (63)	80 (71)	29 (45)	12 (80)

Abbreviations: HADS, Hospital Anxiety and Depression Scale. Figures are numbers and percentages unless otherwise stated. For statistical testing, categorisation of the following variables was used due to small sample sizes: General health: Excellent/Very Good/Good vs Fair/Poor, Knee pain in the last month: Most days or more vs some days or less, and Onset of knee problem: <1 year vs >1 year.

^∗Mean (standard deviation).

^†Median (interquartile range).

^‡Average treatment prevalence is a weighted mean with $wi=Ni/{\sum }_{\text{i}=1}^{5}Ni\text{and}AP={\sum }_{\text{i}=1}^{5}wi\times p_i$ where w = weight, i = time point, N = number of participants without missing data, AP = average prevalence and p = prevalence of treatment use.

^§Includes ibuprofen, aspirin, diclofenac, naproxen and COX-II.

^‖Includes co-proxamol, co-codamol and dihydrocodeine.

^¶Knee injection data was collected at each visit using a time frame of “since your baseline assessment” so an average prevalence could not be calculated.

^#Denominator includes only participants giving consent to medical record review N = 543.

Health care utilisation was higher for the trajectory class representing greater pain severity i.e., the “Severe, non-improving” compared to the “Mild, non-progressive” group in CAS-K and the OAI (Tables IV and V).

Table V.

Health care use over the 6-year follow-up period for the matched OAI sample (N = 570)

Characteristic	Mild, non-progressive N = 235	Moderate (A) (intercept = 4.8) N = 139	Moderate (B) (intercept = 6.9) N = 106	Moderate (C) (intercept = 9.9) N = 64	Severe, non-improving N = 26
Age∗	64 (8.5)	64 (8.3)	64 (8.6)	64 (9.0)	59 (6.9)
Female gender N (%)	111 (47)	65 (47)	61 (58)	39 (61)	20 (77)
Knee function
Baseline WOMAC function27,†	4.0 (0.0, 13.0)	20.0 (12.0, 26.0)	25.3 (19.0, 34.0)	35.5 (29.4, 41.0)	41.7 (37.0, 48.9)
Health care use
Total knee arthroplasty received during 6-yr follow-up period	1 (0)	6 (4)	12 (11)	5 (8)	1 (4)
Average‡ treatment use prevalence over the seven study time points (%)
Paracetamol (Tylenol, acetaminophen)	6%	12%	19%	25%	29%
NSAIDS	18%	30%	31%	35%	53%
COXIBS (e.g., Bextra, Celebrex)	2%	5%	7%	5%	6%
Strong prescription pain medication (e.g., Narcotics)	2%	4%	7%	12%	14%

Figures are numbers and percentages unless otherwise stated.

^∗Mean (standard deviation).

^†Median (interquartile range).

^‡Average treatment prevalence is a weighted mean with $wi=Ni/{\sum }_{\text{i}=1}^{7}Ni\text{and}AP={\sum }_{\text{i}=1}^{7}wi\times p_i$ where w = weight, i = time point, N = number of participants without missing data, AP = average prevalence and p = prevalence of treatment use.

Sensitivity analyses

The choice of the optimum number of classes, the size of the goodness-of-fit statistics and the resulting class allocation in CAS-K did not differ greatly depending on whether a linear, quadratic or cubic model was used to model the data (Web Tables 2 and 3) and all models converged to a global solution. For the CAS-K data, trajectory plots were insensitive to the number of measurement points with complete WOMAC Pain data. However when the baseline measurement was excluded from the analysis only the “Mild, non-progressive”, “Moderate” and “Severe, non-improving” groups were replicated (Web Figs. 1–3).

The sensitivity analyses of the OAI data revealed some model instability for the 5-class linear model when criteria for the number of non-missing WOMAC scores required for analysis were varied, however all models supported the presence of “Mild, non-progressive”, “Severe, non-improving” classes and the absence of a “Progressive” class (Web Figs. 1 and 2). The model was relatively stable when baseline data were omitted from the analysis (Web Fig. 3).

Discussion

Osteoarthritis is a common, often disabling, chronic condition of importance to the health of the public and health services worldwide^44,45. Discovering reproducible pain trajectory groups is of potential conceptual and practical significance for translation into clinical medicine to inform more personalised advice on prognosis, efficient targeting of treatment, and intensity of monitoring. Our study findings corroborate some previous observations on distinctive pain and functional limitation trajectories in knee osteoarthritis while providing several new insights. Our focus has been on the clinical syndrome of pain and functional limitation which is associated with but not coterminous with radiographically defined structural knee OA.

The dominant picture that emerges from data in both cohorts in this study and in previous studies of osteoarthritis^{15,16,18–20,46} and other common musculoskeletal pains in adulthood⁴⁷ is that of trajectories of pain severity and self-reported functional limitation which, in the context of existing clinical care and self-management, are often non-progressive over several years. Although not directly observable in the present data, fluctuations and exacerbations are likely to be super-imposed on these underlying trajectories^5,46,48,49. In the open population a large group is characterised by mild, non-progressive pain equivalent to the fiftieth centile in normative data from adults of the same age in the general population⁵⁰. Given the favourable prognosis, preventing unnecessary costs and harms associated with over-diagnosis and over-management are likely to be important considerations. At the other extreme, a small number of individuals in both UK and US cohorts were assigned to a trajectory group we labelled “severe, non-improving” and that was characterised by high health care utilisation and persistently severe pain exceeding levels typically seen in cases recommended for joint arthroplasty⁵¹. The demographic characteristics and absence of joint arthroplasty in the OAI sample members of this trajectory group, however, argue against a simple interpretation of severe end-stage osteoarthritis and further characterisation is needed.

In CAS-K, when all data time-points were analysed one class was extracted that showed an improvement in pain severity over time with this pattern accounting for 11–20%. Symptom improvement in osteoarthritis is plausible^14,52 and has been observed in some previous studies^19,53,54. The extraction of an ‘improver’ group was, however, highly sensitive in the OAI data to the number of completed measurement points required for inclusion in the analysis. Differences between the two cohorts in the timing of enrolment in relation to symptom fluctuation may be responsible but it serves to highlight the wider challenge of beginning observations in prevalent cases with time-varying symptoms.

A concern is that symptom improvement may nevertheless be accompanied by significant functional deterioration, due for example to progressive avoidance of activities associated with actual or anticipated pain^53,55,56. Our data do not support this concern. We found no evidence over the 6-year period among CAS-K improvers of either worsening self-reported locomotor disability or deteriorating objective impairment as judged by repeated measurements of knee flexion range of movement (data not shown).

Conversely, irrespective of the number of measurement points with WOMAC pain data, a ‘progressive’ group was found in every analysis of CAS-K (except when baseline was removed from the analysis) and in no analyses of OAI, a finding unlikely to be solely explained by selective loss to follow-up in the OAI as follow-up rates are generally high. In general, there appeared to be a more favourable prognosis in the US-based OAI cohort than in the UK-based CAS-K cohort, consistent with previous cross-sectional comparisons of clinical severity between OAI and other cohorts⁵⁷.

A number of limitations in the current study should be highlighted. A maximum of seven measurements with minimum interval of 12 months would be expected to limit the ability to detect rapid resolution or deterioration and to selectively promote the identification of trajectories that do not fluctuate greatly over time. Six years is an insufficiently long period of observation to completely rule out progression and our findings must be interpreted in the context of that time-frame. The small sample size in the “Severe, non-improving” group, although replicated in all of our sensitivity analyses, limits the ability we have to explore which combination of participant characteristics best predict trajectory group membership so our observations about between-group differences are interpreted without adjustment for any of the additional participant characteristics observed.

We also note that some of our comparisons are limited by the different data collection methods used in each cohort, with a particular limitation being that we have assumed that the person-specific WOMAC knee pain scores in CAS-K are equivalent to the maximum of the side-specific WOMAC measures in the OAI, and that our findings on health care use should be viewed in the context of differing health care systems in the UK and US (e.g., access to treatment and services). In addition, we have not explored the more detailed issues around healthcare use such as participant's willingness to accept, use and adhere to the treatments offered that could affect participant's symptom trajectories over time. We also note that as the OAI and CAS-K cohorts are not inception cohorts for this analysis, the baseline data reflects the broad range of OA radiographic and clinical severities that could present to health care services at any one point in time. We have also limited our comparisons of trajectory shape between cohorts to be descriptive, rather than statistical, as it would be influenced by the sample sizes obtained in the trajectory groups compared. Additionally, as matching was used to test for model reproducibility our results only apply to participants with baseline characteristics similar to those in CAS-K rather than to participants in cohorts where differing baseline inclusion criteria apply.

Finally our study was about clinical OA and not radiographic severity. The concern may be that those with no definite radiographic OA at baseline dominate those who have a favourable clinical course. However Table IV illustrates that stratification by radiographic grade at baseline does not neatly separate those who subsequently have favourable and unfavourable pain trajectories over time.

Overall, however, our findings strengthen the empirical basis for moving beyond the general characterisation of osteoarthritis as “slowly progressive”. Indeed the observation of mild, non-progressive or improving symptom trajectories, which appear to be more than an artefact or simply the outcome of activity avoidance or high levels of treatment, directly challenges this negative stereotype. For a sizeable minority of individuals, including those with definite radiographic evidence of osteoarthritis which is now regarded as being relatively ‘late’ in the disease process, pain has a favourable prognosis in the medium-to long-term.

Author contributions

The authors contributed to the manuscript as follows: conception and design: GP, ET, EN, DvdW and PC; analysis and interpretation of the data: EN, GP, ET, DvdW and PC; drafting of the article: GP, EN, ET, DvdW and PC. Final approval of the article: All.

Role of the funders

This work was supported by the Medical Research Council (Grant G9900220), Arthritis Research UK (Grant 18175), National Institute for Health Research (Grant NF-SI-0509-10183 to P.R.C.), and by Support for Science funding secured by North Staffordshire Primary Care Research Consortium for National Health Service support costs. The OAI is a public–private partnership comprised of five contracts (N01-AR-2-2258; N01-AR-2-2259; N01-AR-2-2260; N01-AR-2-2261; N01-AR-2-2262) funded by the National Institutes of Health, a branch of the Department of Health and Human Services, and conducted by the OAI Study Investigators. Private funding partners include Merck Research Laboratories; Novartis Pharmaceuticals Corporation, GlaxoSmithKline; and Pfizer, Inc. Private sector funding for the OAI is managed by the Foundation for the National Institutes of Health.

This manuscript was prepared using an OAI public use data set and does not necessarily reflect the opinions or views of the OAI investigators, the NIH, or the private funding partners. The funders did not contribute to data collection, analysis or interpretation of the data, manuscript preparation or submission.

The findings in this manuscript were presented as an oral paper to the Osteoarthritis Research Society International (OARSI) World Congress on Osteoarthritis on 19 April 2013 in Philadelphia.

Conflict of interest

The authors have no conflicts of interest to declare.