Article Text
PDF
Abstract
Objectives Disability prevention strategies are more achievable before osteoarthritis disease drives impairment. It is critical to identify high-risk groups, for strategy implementation and trial eligibility. An established measure, gait speed is associated with disability and mortality. We sought to develop and validate risk stratification trees for incident slow gait in persons at high risk for knee osteoarthritis, feasible in community and clinical settings.
Methods Osteoarthritis Initiative (derivation cohort) and Multicenter Osteoarthritis Study (validation cohort) participants at high risk for knee osteoarthritis were included. Outcome was incident slow gait over up to 10-year follow-up. Derivation cohort classification and regression tree analysis identified predictors from easily assessed variables and developed risk stratification models, then applied to the validation cohort. Logistic regression compared risk group predictive values; area under the receiver operating characteristic curves (AUCs) summarised discrimination ability.
Results 1870 (derivation) and 1279 (validation) persons were included. The most parsimonious tree identified three risk groups, from stratification based on age and WOMAC Function. A 7-risk-group tree also included education, strenuous sport/recreational activity, obesity and depressive symptoms; outcome occurred in 11%, varying 0%–29 % (derivation) and 2%–23 % (validation) depending on risk group. AUCs were comparable in the two cohorts (7-risk-group tree, 0.75, 95% CI 0.72 to 0.78 (derivation); 0.72, 95% CI 0.68 to 0.76 (validation)).
Conclusions In persons at high risk for knee osteoarthritis, easily acquired data can be used to identify those at high risk of incident functional impairment. Outcome risk varied greatly depending on tree-based risk group membership. These trees can inform individual awareness of risk for impaired function and define eligibility for prevention trials.
- osteoarthritis
- knee osteoarthritis
- disability
- functional impairment
- prevention
Statistics from Altmetric.com
Key messages
This is the first report to develop and validate risk stratification trees for incident functional impairment in persons with pre-osteoarthritis, defined as at high risk but not yet with radiographic disease.
Our derivation and validation cohorts are from carefully designed, prospective, longitudinal cohort studies in which participants were comprehensively characterised and with multiyear follow-up.
Forty potential predictors were considered to develop the trees; all predictors considered and in the final trees are easily assessed.
The frequency of incident functional impairment varied according to tree-based risk group membership. Area under the receiver operating characteristic curves were comparable in the derivation and validation cohorts.
The identified trees are feasible for clinical settings to stratify risk and motivate prevention efforts at a stage when such efforts are most likely to be realisable and effective.
Introduction
Knee osteoarthritis (OA) is a major cause of disability. Estimating that OA accounts for 2.4% of all years lived with disability (YLD), WHO ranked OA 10th among contributors to 1990–2013 global YLDs.1–3 Managing disability in knee OA is challenging, especially since beneficial approaches—for example, physical activity and exercise—are difficult for persons with this disease. In these individuals, pain, deformity, deconditioning and reduced aerobic capacity limit activity and exercise, and adaptations to avoid pain may be entrenched. Managing knee OA disability is costly, in part due to morbidity, loss of mobility and effect on work. To reduce individual and societal burdens of knee OA, early prevention strategies may be more effective and less costly than managing established disability.
Effective prevention approaches could delay or reduce the ultimate severity of knee OA disability. Prevention trials are impeded by uncertainties regarding the population to study. First, should the target be persons with knee OA or with pre-OA, defined here as at high risk for knee OA but not yet with radiographic disease? Risk factors for functional impairment, a precursor of disability in knee OA, have been identified.4–33 In persons with knee OA, OA disease and disease-exacerbated factors—for example, pain,4 5 7–9 28–30 buckling,21 33 decreased confidence,15 33 malalignment6 and proprioceptive inaccuracy9—have been associated with functional decline and may modify effects of factors like body weight. These disease-related factors complicate efforts to prevent decline, especially since interventions targeting them are inadequate. Further, these factors likely make lifestyle and behavioural modifications more difficult to achieve. In contrast, persons with pre-OA are at a stage when modification is more likely realisable and effective. Notably, their pathway to disability does not only go through knee OA; focus on pre-OA enables capturing individuals with chronic knee pain whether or not they develop OA.34
Second, prevention trials optimally target persons at high risk for the outcome.35 However, to our knowledge, a method to stratify risk of functional impairment has not been reported either for pre-OA or existing knee OA. Risk stratification methods are critical to identify high-risk groups for trial eligibility and dissemination of prevention strategies.
An established measure of functional impairment, slow gait speed is associated with disability, increased morbidity and excess mortality in older individuals.36–40 Our objective was to develop and validate a practical, user-friendly method of risk stratification for incident slow gait speed in persons with pre-OA, applicable to community and clinical settings. A cohort study of persons with or at high risk for knee OA, the Osteoarthritis Initiative (OAI), offered a unique opportunity to follow individuals with pre-OA; we leveraged the OAI to study this group by extending their follow-up to 10 years. The Multicenter Osteoarthritis Study (MOST), which also includes a large cohort at high risk for knee OA, provided the best current opportunity to validate this method.
Methods
Derivation and validation cohorts
The OAI (4796 persons, 45–79 years) and MOST (3026 persons, 50–79 years) provided our derivation and validation cohorts. OAI and MOST are prospective, observational, longitudinal cohort studies of individuals with or at high risk to develop knee OA (see online supplementary table 1 for study details).41 We additionally required baseline absence of OA in both knees (Kellgren and Lawrence (KL) radiographic grade <2). OAI and MOST used the same radiographic acquisition protocol and centralised reading site.42 43 Persons with slow gait speed (<1 m/s) at baseline were excluded. The Institutional Review Board at each site approved the study.
Supplemental material
Predictors
In the derivation cohort, 40 baseline variables were considered (table 1), including age, sex, race, ethnicity, education, health insurance, marital status and living alone. Physical activity variables included sitting; walking; light, moderate and strenuous sport/recreation; and muscle strength/endurance using Physical Activity Scale for the Elderly44 subscales. Knee pain frequency in both knees was considered. Frequent medication use was for knee symptoms most days of 1 month in the past 12 months. The Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) Pain, WOMAC Stiffness, WOMAC Function,45 KOOS Pain and KOOS Symptoms46 were included, worse of the two knees. Individual Knee Injury and Osteoarthritis Outcome Score (KOOS) Quality of Life items were ≥weekly aware of problems with knees; and ≥moderate, for modified lifestyle to avoid damaging activities to knees, how much troubled with lack of confidence in knees and general difficulty with knees. Whether a participant had limited activities due to knee symptoms in the past 1 month was included. OA or OA symptoms in other joints included observation of hard bumps on joints closest to fingertips, back pain (any, past 1 month) and hip, ankle and foot pain (most days of 1 month during the past 12 months). Comorbidity variables included overweight, obesity, depressive symptoms (Center for Epidemiologic Studies Depression Scale,47 score ≥16), a questionnaire version of the Charlson Index48 (score ≥2) and falls (any, past year). Body mass index (BMI; weight (kg)/height (m2)) was overweight if ≥25≤BMI<30 and obese if BMI ≥30. Other variables included previous knee injury (ever so badly that it was difficult to walk for ≥1 week), previous knee surgery (ever any surgery to either knee), family history of knee replacement and smoking (current). All variables were self-report except BMI.49 50 MOST generally employed similar methods (differences noted in table 1 footnotes).
Characteristics of the derivation and validation cohorts*
Outcome
The outcome was incident slow gait speed (<1 m/s)51–53 at any follow-up, excluding persons with slow gait speed at baseline. Gait speed was measured using a timed 20 m walk in the OAI at baseline and 12, 24, 36, 48, 72, 96 and 120 months, and in MOST at baseline and 30, 60 and 84 month follow-up visits.
Statistical analysis
Classification and regression tree (CART) methods were used in the derivation cohort to identify the best predictor set and develop risk stratification models.54–56 CART, in contrast to logistic regression models, can generate classification/decision trees, following a priori decisions to maximise predictive accuracy based on cross-validation and classify persons into risk groups. Briefly, CART segregates different values of the predictors (classification) through a decision tree composed of progressive binary splits based on recursive partitioning analysis. Every value of each predictor is considered as a potential split, and the optimal split is selected based on an impurity criterion (the reduction in the residual sum of squares due to a binary split of the data at that tree node). When missing values are encountered in considering a split, probability and impurity measures are calculated from surrogates. CART includes all records with outcomes; for any missing predictors, ‘surrogate splitters’ are substituted, back-up rules that mimic primary splitting rules. Each parent node produces two child nodes, which in turn can become parent nodes, with tree building and pruning until the statistical criterion indicates tree fit without overfitting. Terminal nodes, created if no further split was made, are mutually exclusive and exhaustive sample subgroups. Nodes were constrained to a minimum of 60 persons in parent and 30 in child nodes. To avoid overfitting, tree models were evaluated for predictive ability using 10-fold cross-validation. Outcome rates for each terminal node were used to create risk stratification groups in the derivation cohort. The predictive value of the risk stratification models was assessed by ORs and 95% CIs using logistic regression to compare risk group pairs.
Derived trees were then prospectively applied to the validation cohort to independently test their ability to identify participants in different risk groups. Incidence outcome rates for these risk groups and ORs and 95% CIs comparing risk groups were calculated.
The discrimination ability of the prediction models was compared using area under receiver operating characteristic curves (AUCs). Salford Predictive Modeler’s CART V.8.0 was used for CART and SAS V.9.4 for logistic regression and AUC analyses.
Patient and public involvement
We did not involve patients or the public in our work.
Results
The number of persons who had KL 0/0 (KL 0 in both knees), 0/1 or 1/1, and without slow gait speed at baseline was 1870 in the derivation cohort and 1279 in the validation cohort (online supplementary figures 1A–B). Incident slow gait speed occurred at a study visit within 10-year follow-up in 206/1870 (11.0%) persons in the derivation cohort and within 7-year follow-up in 143/1279 (11.2%) in the validation cohort. Baseline characteristics of each cohort are shown in table 1. In derivation cohort participants without and with the outcome, mean (SD) baseline gait speed (m/s) was 1.40 (0.18) and 1.19 (0.14), and change in gait speed (baseline minus final observation) was −0.05 (0.16) and −0.27 (0.18). In validation cohort participants without and with the outcome, baseline gait speed was 1.31 (0.16) and 1.16 (0.12), and change in gait speed was 0.004 (0.14) and −0.25 (0.16).
Supplemental material
In the derivation cohort, the most parsimonious model was a 3-risk-group tree including age and WOMAC Function, shown with outcome frequency for each risk group in figure 1. Moreover, 7-risk-group, 9-risk-group, 11-risk-group and larger risk-group trees were within 1 SE of the measured performance of the most parsimonious tree, but trees with 11 or more risk groups were complex and difficult to interpret. The 7-risk-group tree included age, WOMAC Function, education, strenuous activity, obesity and high depressive symptoms as discriminators; risk groups are shown in figure 2. Table 2 summarises characteristics of these seven risk groups in both cohorts. The 9-risk-group tree additionally included overweight or obesity.
Predictors of incident slow gait speed and risk stratification, derivation cohort (OAI). The figure depicts the optimal (most parsimonious) tree and frequency of outcome for each terminal node (risk group): age >66.5 years (High Risk-1), age ≤66.5 and WOMAC Function >23.7 (High Risk-2), age ≤66.5 and WOMAC Function ≤23.7 (low risk). OAI, Osteoarthritis Initiative.
Predictors of incident slow gait speed and risk stratification, derivation cohort (OAI). The figure depicts the 7-risk-group tree and frequency of outcome for each risk group: age >66.5 (High Risk-3), age ≤66.5, WOMAC Function >23.7 (High Risk-4), age ≤66.5, WOMAC Function ≤23.7, education >some college (Low Risk-2), age ≤66.5, WOMAC Function ≤23.7, education ≤some college, strenuous activities >2 days (Low Risk-1), age ≤66.5, WOMAC Function ≤23.7, education ≤some college, strenuous activities ≤2 days, obese (High Risk-1), age ≤66.5, WOMAC Function ≤23.7, education ≤some college, strenuous activities ≤2 days, non-obese, with high depressive symptoms (High Risk-2), age ≤66.5, WOMAC Function ≤23.7, education ≤some college, strenuous activities ≤2 days, non-obese, without high depressive symptoms (Low Risk-3).
Characteristics of the derivation and validation cohorts in risk groups from the 7-risk-group model*
For risk groups identified by the 3-risk-group tree (figure 1), ORs comparing risk groups were significant for High Risk-1 versus Low Risk groups (OR 5.24, 95% CI 3.80 to 7.24) and for High Risk-2 versus Low Risk groups (OR 7.13, 95% CI 4.35 to 11.69) but not for High Risk-2 versus High Risk-1 groups (OR 1.36, 95% CI 0.84 to 2.21). For risk groups identified by the 7-risk-group tree, comparisons are shown in online supplementary table 2.
The trees generated by analysis of the derivation cohort were then tested for their ability to risk stratify persons in the validation cohort; figures 3 and 4 depict the 3-risk-group tree and 7-risk-group tree, respectively.
Predictors of incident slow gait speed and risk stratification, validation cohort (MOST). The figure depicts the findings when the most parsimonious tree generated by analysis of the derivation cohort was tested for its ability to risk stratify persons in the validation cohort (MOST). MOST, Multicenter Osteoarthritis Study.
Predictors of incident slow gait speed and risk stratification, validation cohort (MOST). The figure depicts the findings when the 7-risk-group tree generated by analysis of the derivation cohort was tested for its ability to risk stratify persons in the validation cohort (MOST). MOST, Multicenter Osteoarthritis Study.
In the validation cohort, for risk groups identified by the 3-risk-group tree, ORs comparing risk groups were significant for High Risk-1 versus Low Risk groups (OR 4.28, 95% CI 2.94 to 6.24) and for High Risk-2 versus Low Risk groups (OR 2.29, 95% CI 1.22 to 4.27), but not for High Risk-2 versus High Risk-1 groups (OR 0.53, 95% CI 0.29 to 1.00). For risk groups identified by the 7-risk-group tree, comparisons are shown in online supplementary table 3. Distribution of incident outcomes, and Kaplan-Meier estimates of probability of outcome-free follow-up, overall and by 3-risk-group trees is shown in online supplementary table 4A-4D for both cohorts.
AUCs to summarise discrimination ability were comparable in the derivation and validation cohorts: for the 3 -risk-group tree, AUC 0.70, 95% CI 0.67 to 0.74 and AUC 0.67, 95% CI 0.62 to 0.71; for the 7-risk-group tree, AUC 0.75, 95% CI 0.72 to 0.78 and AUC 0.72, 95% CI 0.68 to 0.76; and for the 9-risk-group tree, AUC 0.77, 95% CI 0.74 to 0.80 and AUC 0.73, 95% CI 0.69 to 0.77. The AUCs for the 7-risk-group tree were better than for the 3-risk-group tree, for both cohorts (each p<0.0001). The AUCs for the 9-risk-group tree were better than for the 7-risk-group tree, for both the derivation and validation cohorts: p<0.0001 and p=0.02.
Discussion
In analyses considering 40 baseline variables, a 3-risk-group tree (including age and WOMAC Function) was identified as the most parsimonious model. A 7-risk-group tree (including age, WOMAC Function, education, strenuous activity, obesity and high depressive symptoms) and a 9-risk-group tree performed comparably with the parsimonious tree; AUCs were best for the 9-risk-group tree but were not substantially different from the 7-risk-group tree. Overall, incident slow gait speed occurred in 11%, but the risk varied greatly, between 0% and 29% in the derivation cohort and between 2% and 23% in the validation cohort, depending on risk group membership using the 7-risk-group tree. In both cohorts, ORs comparing risk groups from the 3-risk-group tree were significantly different from 1.0 for High Risk-1 versus Low Risk and for High Risk-2 versus Low Risk. AUCs to summarise discrimination ability were comparable in the two cohorts. These findings suggest that in persons at high risk for knee OA, risk of functional decline can be estimated using easily acquired data.
Methods to stratify risk of functional decline have not been reported for persons with knee OA or pre-OA. Previous longitudinal studies of functional outcome have evaluated persons with knee OA, frequent knee pain, or a pool of persons with or at high risk to develop knee OA. These studies have identified risk factors including age, female sex, socioeconomic status, BMI, pain, comorbidity, depressive symptoms, knee buckling, low knee confidence, falls, laxity, malalignment, disease severity, proprioceptive inaccuracy, sleep disturbance and community mobility barriers, while greater physical activity, less sedentary time, aerobic exercise, strength, self-efficacy and social support were associated with a reduced risk.4–33 These studies have not separately examined pre-OA. While risk stratification in persons with knee OA is important, focusing only on this stage bypasses a compelling stage for prevention, before disease consequences become dominant. Persons with pre-OA are at a stage when lifestyle and behavioural modification to prevent decline are more likely achievable and effective. Further, our findings are relevant to the large pool at high risk for knee OA, including those with chronic knee pain, whether or not they develop knee OA.
The findings in MOST provided some, although not perfect, validation. Outcome frequency was consistently higher in MOST high-risk than low-risk groups, but was sometimes lower in MOST than in the OAI. Possible reasons include fewer follow-up visits, shorter follow-up duration and higher baseline prevalence of no frequent pain in both knees in MOST (table 1). High-risk groups with the greatest difference in outcome frequency between OAI and MOST differed in at least two ways (table 2). First, baseline prevalence of no frequent pain in both knees was higher in these MOST versus OAI groups. Second, while only persons with bilateral KL <2 were included, frequency of KL 0 in both knees was higher in the MOST groups; OAI high-risk groups included more individuals with KL 1 in one or both knees.
Outcome frequency in low risk groups was similar between the two cohorts, reinforcing the concept of a resilient phenotype. Risk of incident slow gait speed was low in persons <66.5 years, with a WOMAC Function score ≤23.7, if above the education threshold. Being younger and with a better WOMAC Function score and below the education threshold could be overcome by one of two routes—any strenuous activity ≥2 days/week, or not being obese and not having high depressive symptoms (figures 2 and 4). The findings demonstrate the importance of validation, and that comparable performance cannot be assumed even when studies are similarly designed. Comparability of AUCs in the two cohorts provides further evidence of validation and generalisability.
This study has limitations. We used easily assessed variables, deliberately to maximise application of these trees. However, other variables may influence risk discrimination. The validation cohort had a shorter follow-up duration. Both OAI and MOST were designed to study community-dwelling individuals at high risk for knee OA and recruited from population lists but did not use random sampling; to accrue a comparably sized random sample would require this data collection in a very large population study. We included persons without radiographic knee OA. These findings should be validated in a high-risk population without self-reported knee OA. These results may not be generalisable to a non-US population. As an objective performance measure associated with disability and survival,36–40 gait speed was a logical choice to measure outcome. A threshold is more interpretable than change. However, an inherent issue is that individuals closer to a threshold may be more likely to cross it, and predictors may be weighted towards variables associated with being closer to it. Notably, change also has limitations, for example, with interpretability, how to incorporate from where a person starts and what magnitude of change is meaningful at different starting points.
This prognostic stratification could be applied in community and clinical settings to promote awareness of risk and motivate efforts to prevent poor outcome. This would involve identifying persons at high risk for knee OA, and then among them, those at high risk for functional impairment. The approach to identify the former, carefully developed and very similar in the OAI and MOST, is translatable into a short paper or electronic form; this with a tree would yield an easily completed, simple and inexpensive tool. The current findings suggest that the 3-risk-group and 7-risk-group trees are reasonable alternatives. If simplicity is required, the smaller tree may suffice. The 7-risk-group tree is slightly more burdensome but had better AUCs; in theory, the modifiable factors (strenuous activity, high depressive symptoms, obesity) in the 7-risk-group tree could serve to motivate. These findings have impact at two levels: first, as a tool to enhance awareness of risk of impaired function at an early stage, which may motivate steps to prevent decline; and second, to help define eligibility for functional decline prevention trials. For a sense of magnitude, among 1000 with pre-OA, 299 would be classified high risk (High Risk-1 or High Risk-2 using the 3-risk-group tree), of whom 72 (24.1%) would be expected to experience incident slow gait speed over the coming 7–10 years. Of the other 701 persons not classified high risk, 38 (5.4%) would be expected to experience this outcome. There are several potential interventions to prevent disability in pre-OA; an abundant literature suggests the most cost-effective and scalable may include physical activity promotion. Awareness of risk at the stage of our sample, not yet afflicted by knee OA, would be information at a point when these individuals are well enough to act and to perceive such action as a preservation of wellness.
In conclusion, in persons at high risk for knee OA, easily acquired data can be used to identify those at high risk of incident slow gait speed. Outcome risk varied greatly depending on risk group identified using the trees. These trees can inform an individual’s awareness, at an early stage, of risk for impaired function, and define eligibility for prevention trials.
References
Footnotes
Handling editor Josef S Smolen
Correction notice This article has been corrected since it published Online First. The author affiliations have been updated.
Contributors Every author met the ICMJE criteria for authorship.
Funding This work was supported by NIH/NIAMS R01 AR065473, R01AR066601, P30AR072579 grants. The Osteoarthritis Initiative is a public-private partnership comprised of five NIH contracts (N01-AR-2-2258; N01-AR-2-2259; N01-AR-2-2260; N01-AR-2-2261;N01-AR-2-2262) funded by the NIH and conducted by the OAI Study Investigators. Private sector funding for the OAI is managed by the Foundation for the NIH.The Multicenter Osteoarthritis Study is supported by NIH grants U01-AG-18820,U01-AG-18832, U01-AG-18947, and U01-AG-19079.
Competing interests None declared.
Patient consent for publication Not required.
Ethics approval The Institutional Review Board at each site approved the study.
Provenance and peer review Not commissioned; externally peer reviewed.
Data availability statement Data are available on reasonable request.