Precision medicine-based machine learning analyses to explore optimal exercise therapies for individuals with knee osteoarthritis

Abstract

Purpose: Exercised-based interventions, including physical therapy (PT) are core treatments for knee osteoarthritis (OA). Patients vary considerably in their level of improvement following different exercise-based interventions, but very little is known about drivers of this variability. Precision medicine-based machine learning methods can be used to identify optimal treatments for each patient, based on individual characteristics. In these analyses, we used a precision medicine approach to discover factors underlying differential improvement in the Physical Therapy vs Internet-Based Exercise Training for Knee OA (PATH-IN) trial, which compared the effects of standard PT with internet-based exercise training (IBET), both relative to a usual care / wait list control group (WT). Methods: The PATH-IN study enrolled 350 participants with symptomatic knee OA (physician diagnosis and pain, aching or stiffness on most days of the week). Outcomes were assessed at baseline, 4-month follow-up (primary outcome time point) and 12-month follow-up. The primary outcome for both PATH-IN and these exploratory analyses was the Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) total score. Some participants were missing WOMAC scores at follow-up, resulting in sample sizes of 303 and 302 participants for analyses of 4- and 12-month outcomes, respectively. Our precision medicine approach involved estimating an “optimal treatment regime” (from among the 3 study arms of PT, IBET, and WT) for each patient. To address limitations of existing machine learning methods, particularly the ability to identify factors that drive the optimal treatment, we developed a new algorithm, Tree-based Interpretable Learning. This algorithm divides study participants into multiple disjoint subgroups where each subgroup receives a single optimal treatment among the three options. The algorithm judiciously splits the data set into two groups in a way that maximizes the average outcome after assigning estimated treatments, by random forest (which is an ensemble of multiple decision trees), to the first group and assigning one of three treatments to the other group. The algorithm then repeats splitting the group that received estimated treatments by a random forest in the previous step until the splitting no longer sufficiently improves the cross-validated estimate of the average outcome. In each loop, the improvement is evaluated by testing the difference of the cross-validated estimate of the average outcome of the estimated treatment regime and the estimate of the average outcome of the zero-order model (ZOM), which assigns a single best treatment to all patients. This methodology is meaningful in that it produces an interpretable treatment rule by utilizing random forests, which yields a low-bias estimate yet reduces the “black-box” aspect of machine learning algorithms. We conducted separate analyses for 4- and 12-month outcomes. Participant demographic, clinical, and psychosocial characteristics included as potential predictors of the optimal treatment regime are shown in Table 1. Results: For the analysis for WOMAC score at the 4-month follow-up, body mass index (BMI), Brief Fear of Movement score (BF; higher scores indicate more fear of movement), and Self Efficacy for exercise scale (SE; higher scores indicate greater self-efficacy) were identified as characteristics that effectively divided participants, creating five subgroups (Figure 1). For the analysis for WOMAC score at the 12-month follow-up, age, BMI, and BF score were identified as important features resulting in seven disjoint subgroups (Figure 2). Improvement over the average outcome was seen for both the 4- and 12-month follow-up visits (p=0.0062 and p=0.0042, respectively). Conclusions: In this study, optimal treatment regimes were identified for 4- and 12-month WOMAC outcomes, yielding significantly improved estimates compared with the average outcome. This indicates that the proposed, tailored regimes would be more beneficial to participants than assigning a single best treatment (based on overall results from the clinical trial) to all patients. Importantly, multiple subgroups of patients were assigned to IBET as the optimal treatment regime; two key characteristics of these subgroups were lower BMI and low fear of movement. Patients with these characteristics may benefit from a low-cost, self-directed program and may not require PT, which can be more costly. In contrast, results suggest that patients with higher BMI and greater fear of movement may benefit most from PT. There was a subgroup of patients for whom WT was the optimal treatment regime at 4 months, indicating that neither PT nor IBET were particularly effective. This subgroup was characterized by very high BMI; based on prior research, these patients may experience the most benefit initially with a weight loss intervention. In summary, these analyses identified easy-to-assess patient characteristics that can help tailor referrals to different exercise-based interventions for knee OA. Additional studies are needed to confirm these findings in other samples and for other exercise-based therapies.View Large Image Figure ViewerDownload Hi-res image Download (PPT)View Large Image Figure ViewerDownload Hi-res image Download (PPT)