Radiographic Knee Osteoarthritis Identification Using Wearable Plantar Pressure: Secondary Cross-Sectional Analysis...

Radiographic knee osteoarthritis (ROA) is associated with abnormal plantar loading and altered gait coordination. Insole-based plantar-pressure sensing offers a practical wearable approach to ROA identification, but existing methods typically use either conventional biomechanical descriptors or end-to-end temporal models. An interpretable subject-level framework that captures both local abnormalities and distributed coordination is, therefore, needed.

This study aimed to develop an interpretable subject-level framework that combines conventional biomechanical descriptors with coordination-level plantar information to identify ROA.

In secondary cross-sectional analysis, using a previously reported Pearl River Osteoarthritis Cohort plantar-pressure dataset from our group, we analyzed 92 participants (43 with ROA and 49 without ROA; mean age 63.0, SD 8.4 y). The source cohort was recruited from clinical and nearby community sources at Zhujiang Hospital of Southern Medical University, Guangzhou, China, between January 2022 and February 2023. We developed an interpretable participant-level classification framework integrating conventional biomechanical descriptors with a coordination-level Plantar Synergy Index (PSI). Participant-wise stratified split allocated 72 participants for 5-fold cross-validation, and 20 for locked independent evaluation. Feature selection used Fisher-score ranking (k=50). All point estimates are accompanied by 2000-iteration bootstrap 95% CIs, and paired model comparisons used DeLong tests. All statistical tests used α=.05. Additional analyses examined feature stability, phase-resolved PSI patterns, robustness to perturbation and design variation, and comparisons with representative end-to-end temporal baselines.

The fusion model achieved the best primary classification performance on the independent test set, with an F₁-score of 0.842 (95% CI 0.615-1.000) and accuracy of 0.850 (95% CI 0.700-1.000). It also showed balanced performance across additional metrics, with area under the receiver operating characteristic curve (AUROC) of 0.818 (95% CI 0.576-1.000), precision of 0.800 (95% CI 0.538-1.000), recall of 0.889 (95% CI 0.636-1.000), and specificity of 0.818 (95% CI 0.571-1.000). DeLong testing showed that fusion outperformed PSI-only at the AUROC level (P=.02). Stable retained descriptors were dominated by biomechanical features, with a reproducible PSI component. Phase-resolved analyses suggested distributed stance-phase coordination differences. Robustness and baseline comparisons supported the stability of the framework, but the wide CIs indicate that the findings remain preliminary.

This study introduces an interpretable participant-level representation that integrates local biomechanical loading with multiscale interregional temporal coordination, extending plantar-pressure approaches based on summary descriptors or end-to-end models. Across evaluated settings, the framework showed competitive identification and preliminary robustness. Local biomechanical descriptors constituted the main discriminative basis, whereas PSI contributed a reproducible coordination-level component whose incremental predictive value remains to be established. This broadens plantar-pressure analysis from local loading to distributed coordination. External multicenter prospective validation is required before evaluating the framework as an imaging adjunct for functional assessment and longitudinal monitoring.