Journal: NPJ Digital Medicine

Article Title: Continuous assessment of daily-living gait using self-supervised learning of wrist-worn accelerometer data

doi: 10.1038/s41746-026-02528-2

Figure Lengend Snippet: Curves represent mean values (±1 standard deviation across 10 folds) obtained from nested cross-validation. Three different feature sets were compared: supervised-based accelerometry features (blue) extracted from standardized walking tests (8-ft and 32-ft walks) measuring gait speed, step length, cadence, and stride regularity; physical activity features (red) and gait quality features (green) extracted from daily-living accelerometer data. Classification performance, indicated by the area under the curve (AUC), improved from supervised-based features (AUC = 0.67 ± 0.09) to physical activity features (AUC = 0.69 ± 0.06), and was highest when using all available wrist multi-day gait metrics (AUC = 0.80 ± 0.06). A Friedman test showed a significant effect of feature set on AUC ( p < 0.001). Post hoc Wilcoxon tests revealed the gait quality model outperformed both supervised ( p < 0.01) and physical activity models ( p < 0.01), with no significant difference between the latter two. Note that these results are based on the application of both stages of the ElderNet pipeline: gait detection, followed by the estimation of wrist-derived gait features.

Article Snippet: Dataset 2: Rush Memory and Aging Project (MAP), which includes 819 older adults (mean age 83.4 ± 7.3 years) who wore a wrist accelerometer (Axivity AX3, 50 Hz sampling rate or GENEActiv, 40 Hz sampling rate) for up to 10 consecutive days.

Techniques: Standard Deviation, Biomarker Discovery, Activity Assay, Derivative Assay