How many repeated observations are enough for digital gait summaries? A multi-dataset framework for sampling...

Repeated digital gait and mobility assessments are increasingly summarized into indicators at the participant level, yet the number of repeated observations needed for different summary families to support stable interpretation remains unclear. We developed and evaluated a multi-dataset framework to estimate sampling sufficiency for repeated digital gait and mobility summaries in older adults.

Four publicly available datasets were analyzed. Three datasets with stride observations spanning optical motion capture, foot-mounted inertial sensing, and cross-modality inertial/video recordings served as the development tier based on stride observations, and one dataset based on Timed Up and Go (TUG) served as a supportive transfer dataset with repeated phase- and trial-level observations. Five summary families at the participant level were compared: Mean, Top10Mean, Bottom10Mean, Best2, and Worst2. For each combination of dataset, feature, and summary family, repeated subsampling within each participant was used to estimate the minimum stable observation count relative to each participant's reference based on all available data, using a prespecified median absolute percentage error criterion (MdAPE). Downstream analyses of associations with anchors were restricted to summaries that met the stability criterion. We then performed a focused comparison between Mean and Bottom10Mean in the primary dataset with stride observations and representative inferential-stability analyses under repeated subsampling.

Across datasets with stride observations, minimum stable observation counts depended strongly on summary family and sensing context. Mean generally required the fewest observations, whereas numerical-tail and extreme summaries defined by feature orientation required substantially denser repeated sampling and did not consistently reach threshold across all features. After stability qualification, selected noncentral summaries retained statistically supported associations with functional, frailty-related, and fall-related anchors. In Dataset 2, the lower-tail summary Bottom10Mean retained information beyond the corresponding mean for a subset of feature-anchor combinations. In the dataset based on TUG, all evaluated summary families reached the stability criterion with lower observation burden than in the datasets with stride observations, providing supportive evidence that the same analytical logic was also supported in a different setting with repeated observations. Repeated subsampling further showed that insufficient observation support primarily attenuated downstream effect estimates and increased uncertainty rather than causing frequent directional reversals.

Repeated digital gait and mobility summaries should not be assumed to be interchangeable. Their analytical and potential translational interpretation should depend on whether sufficient repeated observations have been collected for the intended summary family. Beyond gait analysis, the framework may help guide protocol design, data quality control, and summary selection in repeatedly sampled clinical data.