Unsupervised Machine Learning Identifies Shared Biomechanical and Clinical Variation in Individuals with Knee Osteoarthritis and High BMI: A Secondary Analysis of the IDEA Trial.

OBJECTIVE: High body mass index (BMI≥27 kg/m2) and aberrant gait biomechanics are primary risk factors for knee osteoarthritis (KOA), yet the specific loading patterns linked to clinical disease features remain unclear. Unsupervised machine learning (ML) was used to identify vertical ground reaction force (vGRF) waveform modes of variation and characterize the shared and individual variation between vGRF profiles and clinical features in individuals with KOA and high BMI. METHODS: Baseline gait and clinical data from the Intensive Diet and Exercise for Arthritis (IDEA) trial (n=431; 29% male; 65±6 years; 34±4 kg/m2) were analyzed. vGRF waveforms (n=2,394) were represented using three functional data analysis methods: time- and bodyweight-normalization; elastic shape analysis; and optimal transport (OT). Principal component analysis extracted the dominant vGRF modes of variation, and Data Integration Via Analysis of Subspaces (DIVAS) quantified shared and individual variation among vGRF profiles and 40 clinical features. RESULTS: A sustained loading phenotype emerged across all vGRF representations (lesser early and late stance vGRF; greater midstance vGRF). DIVAS identified shared variation between sustained loading and higher BMI, elevated inflammatory biomarkers, greater pain, worse KL grade, and slower walking speed. OT identified a temporal loading phenotype that shared variation with female sex. CONCLUSIONS: Unsupervised ML identified a sustained loading biomechanical phenotype that shared variation with greater BMI, increased inflammation, worse radiographic severity, and slower walking speed. These data highlight the complex interplay among gait biomechanics and clinical KOA features in individuals with high BMI and may help to guide future work evaluating targeted gait retraining interventions.