It is challenging to routinely assess gait unless dedicated measuring devices are available. Inspired by a recent study that reported high classification performance of activity recognition tasks using smartwatches [1], we hypothesized that the recognition of gait-related activities in older adults can be formulated as a supervised learning problem. To quantify the complex gait motion, we focused on hand motion because disturbed hand motions are frequently reported as typical symptoms of neurodegenerative diseases [2].

We employed the accuracy and macro average of the F₁-score, F_m, as a measure of performance. For the both-hands condition, the accuracy (F_m) was 0.9757 (0.9728), 0.9736 (0.9699), and 0.9771 (0.9738) when Conv1D, LSTM, and attention-based LSTM were employed, respectively. In the case of the left-hand and right-hand conditions, the accuracies (F_m) obtained in the left-hand condition were 0.9652 (0.9623), 0.9611 (0.9583), and 0.9630 (0.9592), respectively. In the right-hand condition, the accuracies (F_m) were 0.9724 (0.9706), 0.9673 (0.9643), and 0.9673 (0.9635) for the same employed models, respectively. We also examined the learned representations as shown in Figure 2 using t-distributed stochastic neighbor embedding (t-SNE) [6], which visualizes the high-dimensional vectors by projecting them into a 2D space in such a way that similar points cluster together.

The experimental results demonstrated an acceptable classification performance (ie, both accuracy and the F_m score were higher than 0.95). However, there is systematic confusion, such as recognizing C3 as C2 (0.03-0.04 for the left hand, 0.05-0.07 for the right hand, and 0.05-0.06 for both hands, respectively) as shown in Figure 2 (see the region highlighted in black). It is noteworthy that the classification performance of the single-hand condition was similar to that of the both-hands condition, suggesting that wearing a single smartwatch is sufficient for the proposed gait assessment task. From the t-SNE plot, it was found that points from the LSTM and attention-based LSTM exhibit a more clustered distribution than those from the Conv1D model. We expect that the proposed approach can be applied to various health care applications for older adults (eg, wearable detection of gait disturbances).