Trajectories of Structural Disease Progression in Knee Osteoarthritis

Abstract

Objective

Knee osteoarthritis (OA) is a heterogeneous disease, with most patients experiencing slow disease progression and some with rapid deterioration. We aimed to identify groups of persons with symptomatic knee OA experiencing rapid structural progression.

Methods

We selected participants from the Osteoarthritis Initiative with baseline (BL) Kellgren-Lawrence (KL) grade 1 – 3, knee pain, and with joint space width (JSW) on fixed-flexion knee radiographs assessed at BL and ≥1 follow-up over 8 years. We used latent class growth analysis to identify subgroups of JSW progression, jointly modeling time to knee replacement (KR) to account for potential informative dropouts. After identifying trajectories, we used logistic regression to assess the association between BL characteristics and JSW trajectory group.

Results

We used data from 1578 participants. BL radiographic severity was KL1 in 17%, KL2 in 50%, and KL3 in 33%. We identified 3 distinct JSW trajectories: 86% stable, 6% with stable JSW followed by late progression, and 8% with early progression; incorporating information about KR resulted in 47% of KRs initially classified as stable being re-classified to one of the progressing trajectories. Prior knee surgery was associated with being in the late progressing vs. the stable trajectory while obesity was associated with being in the early progression vs. stable trajectory.

Conclusion

In addition to a subgroup of individuals experiencing early structural progression, 8-year longitudinal data allowed the identification of a late progressing trajectory. Incorporating information about KR was important to properly identify longitudinal structural trajectories in knee OA.

Introduction

Knee osteoarthritis (OA) affects 14 million individuals in the United States and over 300 million adults worldwide.^1,2 OA of the lower extremities is associated with decreased quality of life, leading to health expenditures in the United States alone totaling over $27 billion annually.^3,4 While traditionally considered a disease of aging, recent data suggest that over 8 million knee OA patients in the US are under age 65 and 2 million are under age 45.¹

Despite the clinical and economic impact of knee OA, no disease-modifying agents are currently approved. While traditionally viewed as a disease characterized by inevitable, but often slow, progression, recent work suggests a more nuanced model of OA natural history. Felson et al. proposed that knee OA progression fits a pattern of inertia: knees that are stable tend to remain stable while knees that have begun progressing are likely to continue to progress.⁵ Identifying patients likely to experience rapid disease worsening is a top research priority; this would allow better recruitment strategies for clinical trials, optimizing the execution of trials focused on DMOADs. Identifying risk factors for rapid progression could shed light on targets for prevention of structural deterioration.⁶

Latent class growth analysis (LCGA) is a statistical approach to identify clusters of subjects with similar progression trajectories in longitudinal data. A recent study used this methodology to assess trajectories of progression in MRI-based cartilage thickness over two years in a cohort of subjects with knee OA.⁷ The authors found approximately 12% of subjects were in a progression trajectory. Another study assessed trajectories of progression in x-ray-based joint space width (JSW) and included knee OA free subjects who had risk factors putting them at high risk for developing OA.⁸ This study found a stable trajectory and a progressing trajectory, with 29% of subjects experiencing progression losing approximately 60% of baseline JSW. This method has not yet been used to assess long-term JSW changes in subjects with existing symptomatic knee OA.

One challenge in assessing long-term progression of disease in longitudinal studies is subject drop out before the scheduled end of follow-up, e.g., due to death or loss-to-follow-up. A difficulty in assessing longitudinal JSW is knee replacement (KR) surgery, which precludes further assessment of JSW. Subjects undergoing KR progress, on average, more quickly than subjects not having KR.^9–11 By not accounting for dropout due to KR in estimating disease trajectories, we may fail to accurately identify a trajectory of rapid progressors or to appropriately categorize KR recipients.

Approaches to define structural deterioration trajectories in knee OA that account for KR may facilitate the identification of rapid knee OA progressors and risk factors indicative of such rapid decline. We aimed to use LCGA to identify groups of persons with symptomatic knee OA experiencing rapid structural progression, taking into consideration implications of informative dropout due to KR.

Methods

Sample

We used data from the Osteoarthritis Initiative (OAI), a multicenter, longitudinal, observational study of knee OA.^12,13 Men and women ages 45–79 were enrolled at four clinical centers between 2004 and 2006. Questionnaires were administered annually for 9 years. Centrally read radiographs are available annually through year 4 and at years 6 and 8. To identify a cohort with radiographic, symptomatic knee OA, we selected participants with baseline (BL) Kellgren-Lawrence (KL) grade 1 – 3, with joint space width (JSW) assessed at BL and at least one follow-up, and with knee pain (OAI definition of symptomatic knee OA: pain, aching, or stiffness in or around the knee in the past 12 months).¹⁴ We selected one knee per participant; choosing the knee with higher WOMAC pain, or randomly selecting a knee if the pain was equal.

Outcome

JSW is used to measure structural disease progression in knee OA. It was assessed in the OAI with a fixed-flexion knee radiograph.¹⁵ We used serial measures of fixed JSW at x=0.25 in the medial tibiofemoral compartment. Fixed location-specific JSW has been shown to be more responsive and have higher reproducibility compared to minimum JSW.^16,17 We excluded poor quality x-rays based on tibial rim distance, beam angle, and poor image quality (details in Appendix section 1).

Covariates

We considered potential predictors of trajectory group membership based on evidence from prior studies:^7,8,18–20 sex, age, alignment, history of knee injury, history of knee surgery, presence of hand OA, family history of KR, baseline BMI, and baseline pain, function, and symptom duration. Alignment was quantified by femur-tibia angle (FTA); an angle of approximately −4.7° represents neutral alignment, more negative values represent varus alignment, higher values represent valgus.²¹ Hand OA was present if a subject had three or more Heberden’s nodes (across both hands) at the baseline clinical examination.¹⁸ Baseline BMI was grouped into normal weight or overweight (BMI < 30), and obese (BMI ≥ 30). Pain and function were assessed with the WOMAC pain and function subscales (0–100, 100 worst). Participants with frequent knee symptoms (most days for at least one month) were asked about duration with the question “How many years ago did this pain, aching, or stiffness start?” with response options: 1 year or less, 2 to 5 years, more than 5 years.

Knee Replacement

We identified knees that underwent knee replacement (KR) (details in appendix 1d). Follow-up time for subjects with KR was the time between baseline visit and KR date. For subjects not undergoing KR, follow-up time was the time between baseline visit and last study visit.

Statistical Analyses

We used latent class growth analysis (LCGA) to identify distinct subgroups of JSW progression. LCGA allows for the modeling of distinct subgroups based on longitudinal trajectory.^22–24 We considered between one and five trajectories, following the steps outlined by Lennon et al. to consider the number, shape, and random effects (RE) structure of the trajectories.²⁵ We chose the initial working RE structure after examining the residuals of a model with no REs; random intercept, random slope, or a quadratic term were considered if the residual profile was flat, straight line, or a curve, respectively.²⁶ Using the initial working RE structure, we determined the number of trajectories based on fit considering the Bayesian Information Criterion (BIC), the posterior group membership probabilities, and the number of subjects in each group. The BIC is a measure of model fit that measures improvements in model likelihood accounting for the number of parameters estimated. Each subject is assigned a posterior group membership probability (PGMP) for each trajectory, representing the probability that a subject belongs to that specific trajectory. A subject is assigned to the trajectory with the highest PGMP. Literature suggests that the average PGMP should be ≥ 0.7 for each trajectory.²⁷ Finally, for model stability, we sought trajectories with at least 50 subjects per group (approximately 3% of the cohort).

While LCGA can handle missing data under the missing at random assumption (i.e., when missing data is assumed to be related to observed but not unobserved outcomes and covariates), these models on their own do not accommodate informative dropout.²⁸ In the case of KR, we might expect that subjects progressing rapidly are more likely to undergo KR.²⁹ We used an extension of the LCGA framework to jointly model longitudinal JSW and time to KR. This approach can be used to correct for potential biases due to informative dropout.^30–32 We re-estimated the trajectories of structural progression, using the RE structure determined in the earlier steps, to jointly model longitudinal JSW with time to KR, using joint latent class mixed models.³⁰ We evaluated models with 1–5 trajectories in similar fashion to the non-joint models.

The LCGA was conducted using the R package lcmm.³³ The gridsearch function was used with 100 starting values and 20 iterations in order to ensure that final solutions converged to the maximum likelihood and not a local maxima.³⁴

We assessed the association between trajectory group membership and baseline covariates using multinomial logistic regression, advancing covariates with p<0.05 in univariate analysis into a multivariable model. We considered models without (model 1) and with (model 2) adjusting for baseline disease severity. Due to high correlation between pain and function, only pain was considered for multivariable modeling.

We assessed the association between trajectory group membership and longitudinal pain using linear mixed-effects models, with repeated measurements of pain as the outcome and trajectory group as the predictor. The model adjusted for time, quadratic time, KR, and the interaction between time, trajectory group, and KR. We used pattern-mixture modeling to account for potential informative dropout due to KR: the final pain estimates within each trajectory are a weighted average of no KR and KR groups, with weights equal to the proportion of subjects in each group.³⁵

Sensitivity Analysis

Our primary analysis included subjects with KL grade 1–3, to capture subjects with both early and established OA. Since pharmacologic trials in OA largely focus on subjects with KL grade 2–3, we re-ran the modeling steps excluding those subjects with KL grade 1.

Results

Sample

Our final analytic sample included 1578 subjects. The average age was 62 years, 62% female, and 46% obese (Table 1). The average baseline WOMAC pain was 22 (0 – 100, 100 worst) (SD 19.5); 17% of knees were KL 1, 50% KL 2, and 33% KL 3. The median number of JSW measurements post-baseline was 5; 32% of subjects had all 6 follow-up measurements while 8% had only 1 follow-up measurement. Subjects included vs. excluded from the analytic sample were similar on baseline characteristics with the exception of baseline KL (Appendix Table 1). Per the quantitative image assessment protocol in the OAI, only a subset of knees in the incidence subcohort without radiographic OA had images assessed; thus, approximately 50% of KL1 knees were included compared to 80% of KL 2 and 3 knees.³⁶

Table 1.

Cohort Characteristics for the Analytic Sample. Presented in cells: mean (SD); median for continuous variables, n (%) for categorical variables.

Characteristic	Statistic
WOMAC Pain Score	22.4 (19.5) 15.0
WOMAC Disability Score	20.0 (18.4) 15.6
Duration of frequent knee symptoms
No frequent knee symptoms	474 (30.2%)
1 year or less	191 (12.2%)
2 to 5 years	462 (29.4%)
More than 5 years	442 (28.2%)
Baseline JSW (medial JSW at x=0.250)	5.4 (1.3) 5.4
Baseline JSW (medial minimum JSW)	4.0 (1.2) 4.0
Baseline KL grade
1	265 (17%)
2	792 (50%)
3	521 (33%)
Age	61.8 (8.9) 62.0
Sex
Male	601 (38%)
Female	977 (62%)
Obese (BMI>=30)
No	856 (54%)
Yes	722 (46%)
Alignment1	−5.5 (2.5) −5.7
History of knee injury2
No	950 (61%)
Yes	613 (39%)
Previous knee surgery3
No	1210 (77%)
Yes	365 (23%)
Hand osteoarthritis
No	1075 (68%)
Yes	502 (32%)
Family history of knee replacement
No	1327 (85%)
Yes	228 (15%)

WOMAC: Western Ontario and McMaster Universities Osteoarthritis Index. Scaled 0–100, 100 worst.

JSW: Joint Space Width

KL: Kellgren-Lawrence

BMI: Body Mass Index

¹Femur-tibia angle; An angle of approximately −4.7° represents neutral alignment; more negative values represent varus; higher values represent valgus.

²History of knee injury was assessed with the question, have you ever injured your knee badly enough to limit ability to walk for at least two days.

³History of knee surgery was assessed with the question: did you ever have surgery or arthroscopy on your knee.

Latent Class Growth Analysis

Model selection is described in detail in the Appendix Section 2. We chose an initial working RE structure of random intercepts and slopes with a quadratic term based on review of the residual plots (Appendix Section 2a). We then examined models with between 1 and 5 groups and found the 3-group solution to be the best fit (Appendix 2b). The 3-group solution with random intercepts and slopes with a quadratic term was selected as the initial model; 88.5% of subjects were in a stable trajectory, 4.1% in a late progressing trajectory, with stable JSW over the first 3–4 years of follow-up followed by progression over years 4–8, and 7.4% were in an early progression trajectory, with progression from baseline to year 4 (Appendix Figure 5). In the initial model, 8% of the subjects in stable trajectory underwent KR over the course of follow-up, compared to 6% of the late progressors and 38% of the early progressors. After jointly modeling time to KR and latent trajectory, the 3-group solution remained the best fit (Figure 1, Appendix 2d). The majority (85.7%, n=1353) of the cohort was in the stable trajectory, with approximately 0.5mm of joint space loss over 8 years of follow-up (i.e., ~0.0625mm/year); 5.9% of the cohort (n=93) was in a late progression trajectory, with slow progression over the first 3 years of follow-up and progression of almost 2mm of JSW over years 4–8 (i.e., almost 0.5mm/year in the latter 4 years); 8.4% of the cohort (n=132) was an early progression trajectory, losing approximately 2mm of joint space width over the first 4 years of follow-up (i.e., ~0.5mm/year in the 1^st 4 years) and then stabilizing. The average posterior group membership probabilities were 0.93, 0.77, and 0.86 in the stable, late progression, and early progression trajectories, respectively.

Figure 1.

Time is along the X-axis and Joint Space Width is along the Y-axis. The line represents the estimated JSW for that trajectory while the band displays the 95% confidence interval. Trajectory 1 in blue includes 85.7% of the cohort and demonstrates a slow stable decline. Trajectory 2 in red includes 5.9% of the cohort and demonstrates late progression, with a sharp decline after year 3. Trajectory 3 in green includes 8.4% of the cohort and demonstrates an early sharp decline, leveling off at year 6. The filled circles represent the observed JSW in each trajectory group at each timepoint.

Effect of Incorporating KR

While the number and shape of trajectories was largely unchanged when we moved from the initial JSW model to the final joint JSW and KR model, 32% of KRs were re-classified, most from the stable to a progressing trajectory (Table 2). Among the 109 subjects with KR initially classified in the stable trajectory, 47% were re-classified: 28 into the late progressing trajectory and 23 into the early progressing trajectory.

Table 2.

Re-classification of Subjects after Incorporating information on Total Knee Replacement.

		Final Trajectory
Original Trajectory	Knee Replacement	Stable n=1353	Late Progression n=93	Early Progression n=132
Stable (n=1397)	No	1287	1	0
	Yes	58	28	23
Late Progressor (n=65)	No	1	60	0
	Yes	0	4	0
Early Progressor (n=116)	No	7	0	65
	Yes	0	0	44

Association Between Baseline Characteristics and JSW Progression Trajectory

The early progression trajectory had lower mean baseline JSW and higher mean WOMAC pain and function compared to the stable and late progressing trajectories (Table 3). Subjects in the early progression trajectory were more likely to have baseline KL 3, with 65% of early progressors with baseline KL 3 compared to 29% in the stable and 40% in the late progressing trajectories.

Table 3.

Associations between Baseline Characteristics and Trajectory Group Membership. Presented in cells: mean (SD); median for continuous variables, n (%) for categorical variables.

	Association with Trajectory Group
Characteristic	Stable n=1353	Late Progression n=93	Early Progression n=132	P-value
WOMAC Pain Score	21.7 (19.4) 15.0	21.7 (17.3) 20.0	30.1 (20.4) 30.0	<.0001
WOMAC Function Score	19.2 (18.3) 14.7	19.6 (16.2) 17.6	29.3 (19.3) 28.0	<.0001
Duration of frequent knee symptoms				0.1877
No frequent knee symptoms	422 (31%)	25 (27%)	27 (20%)
1 year or less	159 (12%)	13 (14%)	19 (14%)
2 to 5 years	386 (29%)	32 (34%)	44 (33%)
More than 5 years	377 (28%)	23 (25%)	42 (32%)
Baseline JSW (medial JSW at x=0.250)	5.5 (1.2) 5.4	5.5 (1.4) 5.4	4.6 (1.5) 4.5	<.0001
Baseline JSW (medial minimum)	4.1 (1.2) 4.1	4.0 (1.2) 4.0	3.1 (1.5) 2.8
Baseline KL				<.0001
1	242 (18%)	13 (14%)	10 (8%)
2	713 (53%)	43 (46%)	36 (27%)
3	398 (29%)	37 (40%)	86 (65%)
Age	61.9 (9.0) 62.0	61.1 (9.1) 60.0	61.5 (7.7) 61.0	0.5897
Sex				0.7956
Male	511 (38%)	38 (41%)	52 (39%)
Female	842 (62%)	55 (59%)	80 (61%)
Obese (BMI>=30)				0.0016
No	753 (56%)	50 (54%)	52 (39%)
Yes	599 (44%)	43 (46%)	80 (61%)
Alignment1	−5.3 (2.5) −5.5	−5.9 (2.3) −5.8	−7.3 (2.2) −7.5	<.0001
History of knee injury2				0.3592
No	821 (61%)	57 (62%)	72 (55%)
Yes	519 (39%)	35 (38%)	59 (45%)
Previous knee surgery3				0.0374
No	1050 (78%)	62 (67%)	98 (74%)
Yes	300 (22%)	31 (33%)	34 (26%)
Hand osteoarthritis				0.7394
No	921 (68%)	61 (66%)	93 (70%)
Yes	431 (32%)	32 (34%)	39 (30%)
Family history of knee replacement				0.3330
No	1141 (86%)	79 (87%)	107 (81%)
Yes	191 (14%)	12 (13%)	25 (19%)

WOMAC: Western Ontario and McMaster Universities Osteoarthritis Index. Scaled 0–100, 100 worst.

JSW: Joint Space Width

KL: Kellgren-Lawrence

BMI: Body Mass Index

¹Femur-tibia angle; An angle of approximately −4.7° represents neutral alignment; more negative values represent varus; higher values represent valgus.

²History of knee injury was assessed with the question, have you ever injured your knee badly enough to limit ability to walk for at least two days.

³History of knee surgery was assessed with the question: did you ever have surgery or arthroscopy on your knee.

Obesity, history of knee surgery, and alignment were significantly associated with trajectory group membership (Table 3). In a multivariable logistic regression unadjusted for baseline disease severity (Table 4, model 1), obesity was associated with a 2.0 times increased odds of being in the early progressing vs. stable trajectory (95% confidence interval (CI): 1.4 – 3.0), while there was not a significant association between obesity and the late progressive vs. stable trajectory (OR 1.1, 95% CI: (0.7 – 1.6)) (Table 4). History of knee surgery was associated with a 1.7 times increased odds of being in the late progression vs. stable trajectory (95% CI: 1.1 – 2.7) and was not significantly associated with being in the early progression vs. stable trajectory (OR 0.9 (95% CI: 0.6 – 1.4)). Each 1 degree decrease in FTA (i.e., more varus) was associated with a 1.5 (95% CI: 1.3 – 1.6) times increased odds of being in the early progressing vs. stable trajectory (e.g., compared to a participant with an FTA of 0, a participant with an FTA of −1 has 1.5 times the odds of being in the early progressing vs. stable trajectory); while FTA was not significantly associated with being in the late progressing vs. stable trajectory (OR 1.1 (95% CI: 0.99 – 1.2)). Results were slightly attenuated after adjusted for baseline pain, JSW, and KL (Table 4, model 2). Subjects with baseline KL 3 had a 2.7 times increased odds of being in the early progression vs. stable trajectory (95% CI: 1.3, 5.8).

Table 4.

Multivariable Logistic Regression

	Model 1		Model 2
Characteristic	OR (95% CI) early progressor vs. stable	OR (95% CI) late progressor vs. stable	OR (95% CI) early progressor vs. stable	OR (95% CI) late progressor vs. stable
Obese (BMI>=30)
No	REF	REF	REF	REF
Yes	2.0 (1.4, 3.0)	1.1 (0.7, 1.6)	1.6 (1.1, 2.4)	1.1 (0.7, 1.7)
Alignment1	1.5 (1.3, 1.6)	1.1 (0.99, 1.2)	1.3 (1.2, 1.4)	1.1 (1.01, 1.2)
Previous knee surgery2
No	REF	REF	REF	REF
Yes	0.9 (0.6, 1.4)	1.7 (1.1, 2.7)	0.7 (0.5, 1.1)	1.5 (0.96, 2.4)
WOMAC Pain Score	--	--	1.02 (1.01, 1.03)	1.0 (0.99, 1.01)
Baseline JSW (medial JSW at x=0.250)	--	--	0.9 (0.7, 1.1)	1.2 (0.98, 1.4)
Baseline KL
1	--	--	REF	REF
2	--	--	1.1 (0.5, 2.3)	1.1 (0.6, 2.1)
3	--	--	2.7 (1.3, 5.8)	1.7 (0.8, 3.4)

WOMAC: Western Ontario and McMaster Universities Osteoarthritis Index. Scaled 0–100, 100 worst. OR for 1 unit increase.

JSW: Joint Space Width

KL: Kellgren-Lawrence

BMI: Body Mass Index

REF=Reference group

¹Femur-tibia angle; OR per 1 unit decrease.

²History of knee surgery was assessed with the question: did you ever have surgery or arthroscopy on your knee.

Association Between Symptoms and JSW Progression Trajectory

Median baseline WOMAC pain was 15 in the stable trajectory, 20 in the late progressing trajectory, and 30 in the early progressing trajectory (Table 3). Median WOMAC function scores were 14.7, 17.6, and 28.0, respectively. The stable trajectory had slightly decreasing pain over 9 years; estimates from the linear mixed effects model suggest a decrease of approximately 6 points over the first 6 years, with slight increasing pain after year 6 (Figure 2). Subjects in the late progressing trajectory increased pain by 11 points over 9 years, with stable pain over the first 4 years followed by increasing pain over years 4 – 9 (11.8 points increase). Those in the early progressing trajectory had relatively stable moderate pain, increasing by approximately 3 points over 9 years.

Figure 2.

Time is along the X-axis and WOMAC Pain (0–100, 100 worst) is along the Y-axis. The line represents the mean WOMAC Pain from the linear mixed effects model for each JSW trajectory group while the band displays the 95% confidence interval. The blue solid line represents the stable JSW trajectory (Trajectory 1). The red dashed line represents the late progressing JSW trajectory (Trajectory 2). The green dashed and dotted line represents the early progressing JSW trajectory (Trajectory 3).

Sensitivity Analysis: KL 2 and 3

Results were similar when we restricted the analysis to the 1,313 subjects with baseline KL 2 or 3 (Appendix Section 3). Eighty-four percent of subjects were in the stable trajectory, 6.7% in the late progressing trajectory, and 9% in the early progressing trajectory. Associations with baseline characteristics were similar to the primary analysis.

Discussion

We used a large longitudinal cohort of subjects with knee OA to investigate patterns of structural progression over 8 years. We identified three trajectories of progression: a stable trajectory losing on average only 0.5mm of JSW over 8 years (~0.0625mm/year); late progression (~0.5mm over years 0–4 followed by 2mm of JSN over years 4–8); and early progression (~2mm of JSN over the first 4 years, followed by 0.2mm over years 4–8). Subjects in the early progression trajectory were more likely to have advanced OA at baseline, with less JSW, higher KL, and worse pain and function. Incorporating information about potential not-at-random data loss due to KR resulted in 32% of subjects with KR being reclassified.

Our analysis is the first to use data from a prevalent OA OAI cohort over 8 years to estimate JSW trajectories. Previous studies using trajectory modeling primarily focused on shorter follow-up. Deveza et al. assessed trajectories of MRI-based cartilage thickness using growth mixture modeling on a subset of subjects from the OAI.⁷ This study is the first to evaluate trajectories of MRI-based cartilage thickness; however, this measure was available only on a selected subset of OAI participants (project 9B, FNIH OA biomarkers consortium) and only at baseline, one, and two years. The authors found that 88% of subjects were in the stable trajectory, 10% were in a moderate progression trajectory and the remaining 2% in a rapid progression trajectory. Bartlett et al. evaluated patterns of JSW over two years in subjects with painful knee OA and found that approximately 9% of the cohort belonged to a progressing trajectory.

We incorporated data from over 1500 OAI participants with radiographic knee OA and pain over 8 years of follow-up. This long follow-up allowed us to uncover a trajectory of ‘late progressors’ – subject that look similar to the stable trajectory over the first few years and then progress. Previous work with shorter follow-up has not identified these late progressors.

Halilaj et al. used data from the OAI incidence cohort over 8 years, which included subjects without knee OA at baseline but at high risk for developing the disease. Using a mixed-effects mixture model, the authors found two trajectories of structural progression, with 71% of subjects in the stable trajectory and 29% in the progressor trajectory. Subjects in the progressor trajectory lost approximately 60% of their baseline JSW over 8 years of follow-up, with approximately 0.3mm of JSN annually. In our analysis, subjects in both the early and late progression trajectories lost approximately 45% of baseline JSW. Focusing on change from baseline rather than the absolute value allowed the authors to focus specifically on patterns of change. We chose to model the absolute value and found that baseline JSW value was associated with trajectory group membership. This fits with the “inertia” hypothesis in knee OA - knees that have begun progressing are likely to experience further worsening.⁵ Baseline differences in disease severity could be explained by the so-called “horse-racing effect” – knees that have already started progressing are likely to be “out in front” (i.e., have less joint space) at baseline because they were in a worsening trajectory before the start of the study.³⁷

There is no universal definition of knee OA progression. In the context of clinical trials of anti-nerve growth factor (NGF) compounds, rapidly progressing OA (RPOA) has been defined as loss of joint space of ≥2 mm in less than one year without gross joint failure.³⁸ The FNIH OA Biomarkers Consortium, on the other hand, defined radiographic progression as a decrease in JSW from baseline to 24, 36 or 48 months of ≥0.7mm, based on study-specific smallest detectable change.³⁹ The annual progression observed in this study varied slightly by year, and averaged approximately 0.5mm/year during the periods of progression (BL – year 4 for early progressors; years 4 –8 for late progressors), greater than the radiographic progression defined by FNIH, and much lower than the RPOA type 1 rate currently being used to define adverse joint safety events in the NGF programs. Thus, a rate of 0.5mm/year could potentially serve as a benchmark for meaningful radiographic progression.

Previous work has shown that dropouts due to KR are likely to be informative.¹¹ While incorporating KR did not uncover a new trajectory of progressors, it led to re-classifying approximately 32% of subjects undergoing KR, most from the stable to a progressing trajectory. Only 10% of subjects underwent KR over the 8 years of follow-up, so it is possible that the effect of KR on estimating the size and shapes of trajectory groups would have been more striking if more subjects had undergone the procedure. We did not incorporate missing data due to other reasons (e.g., non-response, death) as these are unlikely to be associated with unobserved JSN, and thus were thought to be not informative.

In addition to differences in baseline disease severity, we found that obesity, prior knee surgery, and alignment were associated with being in a progression trajectory. While obesity and knee injury are consistently noted to be risk factors for incident knee OA, previous work has pointed to the potential for different risk factors for OA incidence vs. progression.^29,40 Work in both in the OAI and other cohorts has been unable to consistently identify predictors of structural disease progression.²⁰ Deveza et al. identified subjects progressing in MRI-based quantitative cartilage with adequate discrimination (AUC 0.79), but only included a subset of patients with frequent knee pain.⁷ In addition to obesity, the authors found family history of KR, baseline WOMAC pain, baseline medial JSW, and pain duration were significant predictors of progression. In our study, pain duration and family history of KR were not significantly associated with JSW trajectory. Halilaj et al., focusing on trajectories of change in JSW from baseline, were unable to predict progression with high discrimination.⁸ The progression of OA is a complex process involving inflammatory, mechanical, genetic and metabolic factors.^41,42 It is possible that knees that baseline predictors for those knees that have already begun to progress (i.e., early progressors) are different from those that have not. Methods to identify patterns of change across multiple domains may allow for better understanding of disease heterogeneity and thus better predictors of progression.

Cross-sectional evidence from the MOST and Framingham studies show increasing pain with increasing JSN grade, osteophyte grade, and KL grade.⁴³ Previous work on pain trajectories in the OAI showed that while increasing KL is associated with increasing odds of being in a worse pain trajectory, the trajectories themselves show relatively stable pain over time, even among participants with baseline KL 4.⁴⁴ Pain in knee OA has been shown to lead to avoidance of activities as people learn to accommodate.⁴⁵ Perhaps those with early progression have had more opportunity to accommodate to their pain over time, while those with late progression are experiencing more acute changes in pain.

This study has several important limitations. Decisions regarding the number of trajectories and trajectory shapes are somewhat subjective. We set a threshold of 50 subjects (approximately 3% of the cohort) as a meaningful trajectory size; other studies have included smaller group sizes. KR is an elective surgery; the decision depends on many factors, including race, education, and age, among others.⁴⁶ The OAI cohort may not be generalizable to the entire population of patients with OA; the OAI exclusion criteria, including the inability to undergo MRI due to size limits, comorbid conditions that could interfere with follow-up, and plans for bilateral knee replacement, may have led to the selection of a relatively healthy OA cohort..⁴⁷

In summary, we found three distinct patterns of JSW progression in a cohort of subjects with knee OA: the majority experienced stable JSW over time, while the progression patterns included an early progression trajectory and an as yet undescribed late progression trajectory. Osteoarthritis is a slowly progressing disease, and we found that incorporating long-term follow-up of 8 years and informative missingness due to KR were important in defining disease trajectories.

Significance and Innovations.

• We assessed trajectories of longitudinal structural progression in knee osteoarthritis, jointly modeling joint space width and time to knee replacement. This is the first analysis to investigate how informative dropout due to knee replacement affects estimates of disease progression in knee osteoarthritis.

• The long follow-up period of 8 years allowed us to uncover a trajectory of ‘late progressors’ – subject that look similar to the stable trajectory over the first few years of follow-up and then progress. Previous work focusing only on a few years of follow-up has not identified these late progressors.

• The finding of approximately 0.5mm/year of joint space narrowing during periods of disease progression could potentially serve as a benchmark for meaningful radiographic progression in future studies.