Patterns of Shared Variation in Knee Ultrasound for Osteoarthritis: A Machine Learning Approach.

Objective: To identify phenotypes of knee osteoarthritis (KOA) based on demographic and clinical variables, symptoms, and ultrasound (US) features using a novel machine learning approach.

Design: Johnston County Health Study participants provided demographics, symptomatic and functional assessments, and joint radiographs, which were transformed into the clinical data block. Standardized knee US were obtained, and US features composed the second data block. The Angle-based Joint and Individual Variation Explained (AJIVE) algorithm was used to identify shared and individual modes of variation. We focused on shared structure to explore how US features and non-US clinical data vary together overall, and in the subset with radiographic KOA (rKOA).

Results: This analysis included 861 participants (mean age 55 years, mean BMI 33 kg/m²); 335 (39%) had rKOA. AJIVE identified two components of shared variation (SC1 and SC2). SC1 associated osteophytes and cartilage damage on US with higher BMI, older age, and worse symptoms and outcome scores. SC2 correlated the presence of effusion and synovitis but less cartilage damage on US with better physical function and lower BMI. A similar pattern was seen in those with rKOA.

Conclusions: We identified two shared directions of variation which may represent distinct phenotypes of KOA. The first fits with prior KOA studies linking presence of osteophytes and cartilage damage to worse symptoms and function. The second may represent an inflammatory subtype of KOA, with greater effusion and synovitis but less osteophytosis and cartilage damage. These clinically feasible phenotypes should be confirmed in future studies.