A convolutional neural network algorithm for step and gait bout estimation from wristband accelerometry

• Published in: Alzheimer's & dementia Vol. 17; no. S5
• Main Authors: Peraza, Luis R., Kinnunen, Kirsi M., Joules, Richard, Wolz, Robin
• Format: Journal Article
• Language: English
• Published: 01.12.2021
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.053487

Background Walking depends on an intricate interplay between cognitive and motor functions that can be impaired by neurodegeneration. Recent developments on the use of wearable sensors permit the measurement of multiple gait markers to establish their association with symptoms of neurodegeneration. To estimate gait markers, accurate algorithms have been developed for sensors worn around the trunk (e.g., waist) or on lower limbs. However, there has been limited research to develop gait algorithms for wristband accelerometry, despite their superior wear compliance compared to other body‐worn sensors. Our work aims to investigate the feasibility of gait metrics estimated from wrist‐worn accelerometry devices. Method Healthy adults (N=26) were recruited in a multiactivity study which also included a 25‐feet walking exercise along a corridor. Participants wore a lumbar wearable sensor (AX3, Axivity) and a wristband sensor on the non‐dominant hand (GENEActiv, Activinsights). Both devices were mechanically synchronised before and after each study session. Steps were detected and labelled with a heel‐impact detection algorithm applied on the lumbar sensor accelerometry data. Posterior visual inspection was implemented to correct for non‐steps or missing steps. Step labels were posteriorly used to train a convolutional neural network (CNN) algorithm to detect individual steps from the wristband device as well as standing and walking segments. Algorithm performance was assessed with a leave‐one‐subject‐out (LOSO) validation procedure. Result The LOSO validation resulted in an average global accuracy of 88% (95% confidence interval, CI: 87 – 90%) for the three labels of standing, walking, and steps. For the detection of individual steps, the percentage of absolute error was 6.5% (95% CI: 4.9 ‐ 8.0%), or an error of 6.5 steps per 100. The heel impact time error detection showed a mean delay of +9.7 msecs (95% CI: ‐2.2 ‐ 21.7 msecs). Conclusion Our investigation shows that it is feasible to assess gait from wristband devices using CNNs with high accuracy and an acceptable level of model generalisation. Our future work will focus on further improvements to our gait models, and on transferring the trained CNN models to estimate gait in populations affected by neurodegenerative conditions such as different forms of dementia, Parkinsonian disorders, and Huntington’s disease.