Temporal Consistency and Averaging Sensitivity of Released Smart-Insole Gait Summaries in Participants with and without Knee Osteoarthritis

Repeated smart-insole outputs may support longitudinal gait research, but consistency of processed summaries does not establish device accuracy or a clinical monitoring schedule. We analyzed 10 released observations per participant from AI Hub dataset 71966, retaining 178 participants (89 provider-labeled normal and 89 knee osteoarthritis) after exclusion of a previously reported opposing-label duplicate pair. Analyses combined absolute-agreement intraclass correlations (ICCs), participant-bootstrap intervals, residual variability, nonoverlapping 2-, 3-, and 5-observation averages, and exploratory adjusted coefficient-of-variation (CV) models. Single-observation ICCs were 0.827–0.844 for midfoot/rearfoot pressure, 0.663–0.727 for stride length, speed, and forefoot pressure, 0.398 for distance, and approximately 0.014 for balance. Five-observation empirical ICCs were lower than Spearman–Brown predictions for every non-balance endpoint; gait speed yielded 0.887 versus 0.927 and left stride length 0.870 versus 0.908. Left-balance residual SD was 2.388 in the released scale despite low median CV. Linear-programming median regression and solution-choice sensitivity replaced warning-prone bootstrap fitting; the maximum change of a nominal interval endpoint from its reference was 0.0996 CV percentage points. Calendar dates, wear time, and criterion measurements were unavailable. These results characterize release-specific temporal consistency and qualify model-based averaging projections; they do not validate calendar-duration prescriptions, clinical change thresholds, or the underlying device.