Significant pain variability in persons with, or at high risk of, knee osteoarthritis: preliminary investigation based on secondary analysis of cohort data

This was a secondary analysis of prospective observational cohort data from a sample of community-dwelling symptomatic adults – the Clinical Assessment Study of the Knee (CAS(K)).

Participants were recruited from a two-stage cross-sectional postal survey of all adults ages ≥50 years registered with three general practices in North Staffordshire (irrespective of actual consulting patterns). Respondents reporting pain of any duration in or around the knee within the previous 12 months were invited to attend a research clinic at a local National Health Service Hospital Trust. The study protocol was approved by North Staffordshire Local Research Ethics Committee (project number 1430) and details have been published elsewhere [22, 23]. All participants provided written informed consent to undergo clinical and radiographic assessment. In addition, they were asked for consent to medical record review to assist in excluding pre-existing inflammatory disease. The inclusion criteria for the current analysis were as follows: age ≥50 years, registered with one of the participating general practices at the time of study, responded to both postal questionnaires, consented to further contact, and attended the research clinic. Participants were excluded if they had incomplete baseline radiographs, had not experienced knee pain within the six months prior to clinic attendance, had a pre-existing diagnosis of inflammatory arthropathy in their medical records, or had had a total knee replacement in their most affected knee. Participants who reported total knee replacement (TKR) after baseline and up to 3 years were also excluded. Patients reporting TKR after 3 years were censored at the 3 year time point.

All data were planned and gathered prospectively. At baseline, participants underwent a standardized clinical interview and physical examination conducted by one of six research therapists blinded to the findings from radiography, postal questionnaires and medical records.

Participants filled in a brief self-complete questionnaire about their knee symptoms on the day of their clinic attendance.

Plain knee radiographs were obtained on the day of clinic attendance. Three views were taken of each knee: a weight-bearing semi-flexed posteroanterior (PA) view, according to the protocol developed by Buckland-Wright et al. [24], and lateral and skyline views, both in a supine position with the knee flexed to 45°. The tibiofemoral joint was assessed using the PA view and the posterior compartment of the lateral view. The patellofemoral joint was assessed using the skyline and lateral views.

A single reader (a consultant rheumatologist with extensive training in assessing knee radiographs for knee OA), blinded to all other information on participants, scored all films. Films were scored for individual radiographic features, including osteophytes, joint space width, sclerosis, subluxation and chondrocalcinosis. PA and skyline views were assigned a Kellgren and Lawrence (K&L) grade based on these authors' original written descriptions [25]. The atlas developed by Burnett et al. [26] was used for the lateral view.

For PA, K&L score, skyline K&L score and lateral osteophytes, intra- and inter-reader reliability were assessed in a subsample of 50 participants (100 knees) and found to be very good (κ = 0.81–0.98 and 0.49–0.76, respectively) [27].

Follow up surveys, which included 11-point numerical rating scales (NRS) on current, average and worst knee pain intensity over the past 6 months [28], were mailed to Phase 2 participants at 18 months, 36 months, 54 months and 72 months.

At baseline and at each follow-up point we classed participants as reporting ‘significant pain variability’ if their recalled worst pain intensity in the past 6 months was ≥5 out of 10 and at least 2 points higher than recalled average pain intensity in the same 6 month period. To be included in the denominator, individuals had to be ‘at risk’ during that interval (i.e. average pain intensity <9 out of 10).

This definition was chosen after referring to previous studies of osteoarthritis flares which were described as worsening usual pain [11, 13], within defined limits using equivocal pain scales from flare design trials which set a minimum threshold of 50 mm on a 100 mm visual analogue scale (VAS) [29] and a pain increase of at least 20 mm on a 100 mm VAS or an increase of at least two points on a 10 point scale, from baseline [30, 31]. Definitions used in other musculoskeletal disorders such as lower back pain [32] and non-musculoskeletal conditions such as Chronic Obstructive Pulmonary Disease (COPD) were used [33, 34] where worsening of symptoms is used in addition to requiring different or extra medication. The definitions are all reliant on change and difference in pain.

Predictors available in the CAS(K) dataset were selected for analysis on the basis of being shown in previous studies to be associated with incidence and progression of knee osteoarthritis [15, 35‐39], pain outcomes [15] or acute flare-ups [11] (Table 1).

The proportion of participants classed as experiencing significant symptom variability was reported for each time point. For each follow-up time point those experiencing symptom variability at baseline were compared between those followed up and not followed up to identify any differences.

To estimate the association between the putative predictor variables and significant pain variability, we used discrete-time survival analysis. For clinical history/examination and radiographic severity predictors we used information from only one knee per individual, the “index knee”: the single painful knee in participants with unilateral knee pain or the most painful knee in individuals with bilateral knee pain. Discrete-time hazard survival models become models for dichotomous response when the data have been expanded to person-period data with one observation for each year the person is at risk. For each follow-up time point, we constructed an indicator variable on whether the patient had experienced significant pain variability in the 6 month period or not and estimated the hazard of significant pain variability using logistic discrete-time hazards model. The outcome was right censored at 72 months, which was the last follow-up time. Individuals who were lost to follow-up or withdrew from the study before the period of significant symptom variability was recorded, were also censored. To adjust for changes in proportion reporting significant pain variability over time, we included dummy variables for each follow-up time in all models. Two sets of analyses were conducted. We first modelled the time to first period of significant pain variability, ignoring additional subsequent periods of significant pain variability reported by the participant. We then used multilevel discrete-time survival (frailty) models to take into account recurrent periods of significant pain variability within participants. In the frailty model method, the association between periods of significant pain variability is explicitly modelled as a random-effect term. The frailty model was estimated using logistic discrete-time hazards model with random effects.

The association between each predictor and outcome was estimated and those with P-value <0.20 were selected for inclusion in the multivariable models. Tests of multicollinearity were performed first by pairwise correlations (one variable excluded if correlation coefficient >0.7) and then by variance inflation factor (VIF >5 considered as evidence of collinearity). We used a manual backward elimination procedure to remove variables from the multivariable model until only factors with a P-value <0.05 were retained in the final model. An a priori decision was made to include age and gender in the final models. All analyses were performed using Stata 13 [40].

Between 23 and 32% of participants were estimated to have experienced significant pain variability at each of the five time points (Table 2). Across the entire cohort follow up period 363 (47%) participants reported no periods, 202 (27%) reported one period, 90 (12%) reported two periods, 63 (8%) reported three periods, 30 (4%) reported four periods and 13 (2%) reported five periods of significant pain variability. Table 3 presents the descriptive statistics for participants reporting at least one period of significant pain variability.

Based on the outcome of time to first period of significant symptom variability, baseline measures associated with a higher risk of symptom variability in the adjusted analysis were: younger age (OR (per year): 0.96; 95% CI 0.94, 0.97), higher BMI (per kg/m²: 1.03; 1.01, 1.06), higher WOMAC knee pain scores (per unit: 1.05; 1.03, 1.10), longer time since onset (e.g. 1–5 years vs < 1 year: 1.79; 1.16, 2.75) and morning stiffness (≤30 min vs none: 1.43; 1.10, 1.85) (Table 4).

Based on the outcome of recurrent periods of significant symptom variability, i.e. allowing for those experiencing more than one episode, baseline measures associated with a higher risk of potential symptom variability in the adjusted analysis were: younger age (0.94; 0.91, 0.98), higher BMI (1.04; 1.00,1.08), higher WOMAC knee pain scores (1.10; 1.03,1.17), longer time since onset (e.g. 1–5 years vs < 1 year: (2.23; 1.11, 4.46) and morning stiffness (≤30 min vs none: 1.67; 1.07, 2.61) (Table 5).