We developed and validated the Watch Walk method, providing researchers with an online platform to upload wrist motion sensor data and download daily gait performance metrics at no cost.
The Watch Walk method was validated using data from 101 participants aged 19 to 91 in Sydney and Hong Kong. Participants wore a wrist motion sensor on their dominant wrist while being video-recorded in free-living environments and on an electronic walkway. We synchronized motion sensor signals with video recordings and ground-truth measurements from the walkway. The signals were processed into 4-second segments, extracting 54 features to train a two-stage machine learning classifier for identifying walks and corresponding hand positions. Walking speed was estimated using participants’ sex, body height, and six motion sensor features via machine learning models. Step counts, cadence, regularity of steps and strides, longest uninterrupted walking duration, and the proportion of walks ≥ 8 seconds were derived through frequency domain analyses and peak detection methods. Classification accuracy was evaluated using ten-fold validation, and criterion validity for step time and walking speed was established against electronic walkway measurements. Further validation and normative data extraction were conducted with 73,438 UK Biobank participants who wore the wrist sensor for five days or more, assessing test-retest reliability and concurrent validity against self-reported walking pace and mobility issues.
The classification accuracy for walking achieved a sensitivity of 94% and a precision of 93%. Within identified walking events, those with arm swing had a sensitivity of 94% and specificity of 97%. The mean absolute percentage error (MAPE) for step time ranged from 2.9±4.3% to 5.2±11.2%, and for walking speed, it was 5.7±8.5%. Test-retest reliability (Intraclass correlation coefficient, 2-way random, absolute agreement, average) ranged from 0.89 to 0.98. Real-world walking speed corresponded well with self-reported walking pace (F(2,73230)=3052, p0.001), while the longest duration of uninterrupted walking correlated with self-reported mobility issues (H(3,51792)=1915.4, p0.001).
The Watch Walk method generates reliable and valid measures of real-world gait speed, quality, distribution, and step counts. Normative data from UK Biobank participants provide a benchmark for comparisons.
The Watch Walk Platform is a valuable resource for assessing real-world gait performance as a predictor of adverse health conditions and an endpoint for intervention effectiveness.
wearable sensor
Measurement
