Stance-phase kinematic descriptors differentiate clinical subgroups in women with moderate knee osteoarthritis

Total knee replacement (TKR) is the standard surgical treatment for severely symptomatic knee osteoarthritis (KOA), but treatment decisions for patients with moderate radiographic severity (Kellgren-Lawrence grades 2-3) remain variable. Radiographic grading and patient-reported outcomes do not fully capture the functional heterogeneity between treatment pathways. Stance-phase kinematics may provide objective descriptors for stratification while remaining more accessible than kinetic measurements.

Sixty-six women with moderate KOA (27 scheduled for TKR, 39 managed conservatively) underwent gait analysis. Stance-phase angles for 12 variables were extracted at three sub-phases (loading response, mid-stance, terminal stance). Group differences were assessed via repeated-measures multifactorial MANOVA. Gait speed was evaluated for confounding (Rothman's criteria) and included as a covariate in MANCOVA. Stance-phase change variables trained a Random Forest classifier with nested cross-validation and exhaustive feature selection. Associations between descriptors and clinical outcomes were examined via univariate logistic regression. Clinical construct consistency was assessed against the Osteoarthritis Initiative dataset.

The multivariate analysis identified significant Time $\times$ Treatment Group interactions for knee flexion-extension, pelvic tilt, and back flexion-extension after FDR correction. Gait speed differed between groups (TKR slower, Cohen's $d=-1.02$ ) and met Rothman criteria; after adjustment, pelvic tilt retained the lowest uncorrected $p$ -value with no significant Time $\times$ Velocity interaction. Random Forest reached a mean ROC-AUC of 0.81 ± 0.06, with pelvic tilt selected in 80% of folds. Six descriptors showed significant associations with joint pain (highest AUC: pelvic tilt 0.80, shoulder rotation 0.81); none reached comparable association with the remaining outcomes. Pelvic tilt showed inverse association with depression and positive association with pain. Adding velocity as a candidate feature did not change Random Forest performance, indicating redundancy with kinematic descriptors at the individual level.

Stance-phase kinematics differentiate two groups of women with moderate KOA who share radiographic severity but follow different treatment pathways. Pelvic tilt emerged as the most consistent descriptor across multivariate, individual-level, and pain-related analyses, and the most robust under adjustment for gait speed. Kinematics offer an accessible alternative to kinetic measurements for describing clinical heterogeneity in moderate KOA, although integration with imaging and patient-reported outcomes is necessary before clinical implementation.