Applied Model-Based Clustering of Functional Data - Distinguishing Between Phenotypes of Early Knee Osteoarthritis

Abstract

Context: OA is ranked as the 11th highest contributor to global disability and its prevalence is increasing. By gaining a better understanding of OA heterogeneity, we can potentially contribute to the design of clinical trials, prevention strategies, and treatments. To the best of our knowledge, MBCFD has not been attempted to derive knee OA phenotypes. MBCFD treats the data as curves that can potentially allow us to see complex trends not detected by traditional distance-based clustering algorithms. Objective: The study aimed to improve OA heterogeneity understanding by testing an MBCFD method’s ability to derive clinically-relevant and statistically-significant phenotypes and to assess its performance vis-a-vis a method widely used in the scientific literature. Methods: This work is based on the CRISP-IDM method. We identified widely-used algorithms in the literature to derive knee OA phenotypes, as well as their characteristics and derived phenotypes. We selected an appropriate MBCFD algorithm and, through iterative data exploration steps with domain experts, we identified clinically-relevant phenotypes with MBCFD as well as computed the statistical significance between the groups. Subsequently, we compared the performance of the MBCFD method to HCA. Results: MBCFD was able to detect clinically-relevant and statistically-significant knee OA phenotypes for the univariate case. However, for the multivariate case, the phenotypes were clinically relevant but no statistical significance was found between the groups. In addition, MBCFD outperforms HCA in the univariate case but not in the multivariate case.