Improved Prediction of Activities of Daily Living from Wrist Electromyography Using Intermediate Gesture Classification

The long-term goal of this research is to automatically, remotely, and noninvasively identify Activities of Daily Living (ADLs) using muscle activity at the wrist. Alzheimer's Dementia is a widespread condition affecting patients' memory and ability to care for themselves. A wrist-worn device capable of monitoring precise ADLs could increase patient independence and provide caretakers with detailed information regarding the patient's wellbeing. Here, we investigate the use of an intermediate hand gesture classifying algorithm to predict ADLs using wrist electromyography (EMG). We show that predicted gestures are well-represented within the span of ADLs, and that the repertoire of predicted gestures appears distinct among ADLs. Importantly, we show that a simple linear discriminant analysis of predicted gestures provides better and more efficient classification accuracy relative to a state-of-the-art neural network that predicts ADLs directly from EMG. Accurate and efficient classification of ADLs from a wrist-worn device can provide a foundation for remote monitoring of patients in a socially acceptable formfactor. More broadly, a better understanding of human behavior via ADL tracking can enable new assistive technologies that improve quality of life.Clinical Relevance-Monitoring activities of daily living with an electromyographic smart watch can provide insights into patient behavior and wellbeing.