DATA-DRIVEN DISCOVERY OF KNEE OSTEOARTHRITIS SUBGROUPS VIA CLUSTER ANALYSIS OF MRI BIOMARKERS

INTRODUCTION

Identifying structural morphotypes in knee OA, subgroups defined by anatomical and morphological attributes, may facilitate personalized treatment by aligning specific patterns of joint damage with treatment mechanism of action. Cluster analysis is a type of unsupervised machine learning used to uncover subgroups and may provide insight into structural morphotypes in knee OA.

OBJECTIVE

To use cluster analysis to investigate possible subgroups defined by imaging features in a cohort of persons with knee OA.

METHODS

We used data from the PROGRESS OA study, the second phase of the FNIH OA Biomarkers Consortium project, which includes data from the placebo arms of several completed RCTs testing various therapeutic interventions for symptomatic knee OA. MRIs were obtained at baseline and read according to the MRI OA Knee Score (MOAKS) by an experienced radiologist. We included MOAKS assessments of BML size, osteophytes, cartilage, Hoffa-synovitis, effusion-synovitis, and meniscus in the clustering algorithms. Raw ordinal MOAKS scores were used in this analysis. We used Partitioning Around Medoids (PAM) for clustering. PAM is similar to K-means, but instead of defining cluster center as the centroid (mean), the medoid is used, making the method more robust to outliers and appropriate for non-Gaussian data. We undertook several approaches to clustering to perform dimension reduction and incorporate correlations between MOAKS scores A: PAM on Gower’s distance; B: PAM on the dissimilarity matrix from Spearman correlation; C: PAM after non-metric multidimensional scaling (NMDS) using Gower distance for dimension reduction. These approaches aimed to uncover patterns orthogonal to disease severity. The number of clusters was selected based on silhouette width and the gap statistic. Silhouette scores 0.25 to 0.5 indicate weak to reasonable fit.

RESULTS

356 participants from four RCTs were included, 138 (39%) with KLG 2 radiographs and 218 (61%) with KLG 3. The cohort was 57% female with average age 62 (SD 8). The number of clusters ranged from 2 to 3 depending on method. There was modest to high overlap between clustering solutions from different methods, suggesting some stability of clustering solutions. Average silhouette scores were 0.19, 0.13, 0.40 for methods A, B, and C, suggesting poor to modest fit. This could suggest weak structure, overlapping clusters, or need for additional dimension reduction. Methods A and C had one cluster dominated (>95% KLG 3) by KLG 3 knees (Figure 1). Investigation of MOAKS assessments by cluster for each of three clustering solutions is shown in Table 1. For example, method C suggested 3 clusters. Clusters 1 and 2 are both approximately 55-60% KLG 2. Cluster 1 has more lateral cartilage damage, and higher BML and osteophyte scores, while cluster 2 has more medial cartilage damage and medial meniscal damage. Cluster 3 is 96% KLG 3 and has extensive medial cartilage damage, with 84% with widespread full-thickness damage in the MFTJ.

CONCLUSION

While knees can be separated into clusters based on tissue damage assessed by the MOAKS system, both level of disease severity and compartment involvement (medial vs. lateral) play important roles. Silhouette scores suggest the potential for overlapping clusters or the need for additional data reduction. The advanced disease stage common in DMOAD trial populations may limit the ability to identify meaningful structural morphotypes.