Impact of Spatial Context on Characterization of Activities of Daily Living Using sEMG Signals

Assessment of activities of daily living is an important method for evaluating a) the overall health of individuals living in smart health-care homes and b) progress in the rehabilitation process for stroke patients. The classification of ADL is an important component of such assessment methods. In practice, surface electromyography (sEMG) signals can be used to automatically classify ADLs in real-world scenarios. Even though, under laboratory conditions, the classifiers perform well, their widespread usage is limited by the measurement context. Specifically, models that perform well on data from a specific context may fail with data from a different context. In this study, the focus is on the spatial context of measurement in terms of the environment of ADL, reaching distance between the subject and the objects of ADL, and body positions. sEMG signals are collected from 10 subjects while performing ADL in two distinct environments, such as the kitchen and workspace, under two body postures and three reaching distances. Features based on time domain (TD) statistics and log moments in the frequency domain are extracted and classified using the random forest model. The machine learning experiments on ADL classification are conducted in two scenarios; when the environmental context is 1) assumed to be known and 2) estimated from the data. The key finding in this work is that the ADL classification is consistently higher in the first scenario and basic DL frameworks fail to perform well in both scenarios. Finally, from the statistical tests, it is determined that i) reaching distance has a statistically significant impact on the ADL classification in the workspace environment and ii) body position has a significant impact on classification performance of the most of the activities.