Assessing Daily Activity Routines Using an Unsupervised Approach in a Smart Home Environment

Abstract

When the mental acuity of older adults deteriorates (e.g., dementia), irregular patterns manifest within their activities of daily living (ADL), which renders an effective opportunity for healthcare providers to monitor patients' mental status. Although successful, such studies depended on supervised learning approaches to recognize ADLs, which require tedious human observation and manual annotation of data. To establish a more efficient alternative, this study develops an unsupervised data segmentation process by modifying a Superpixels Extracted via Energy Driven Sampling (SEEDS) algorithm and a hierarchical clustering method effective for high-dimensional temporal sensor data. The proposed approaches consider the spatiotemporal features (e.g., start time, duration, location, and sequence) and activity-oriented features (e.g., motion intensity and appliance usages) to identify ADL routines without necessitating predefined rules or limiting the scope of features. The results showed that the proposed approaches have comparable accuracy (0.788) to benchmark models that require a priori knowledge (e.g., ontology). Our proposed methodology can be extended to high-dimensional, nonintrusive sensing data to capture the variability of ADL routines in the future. This study contributes a methodological advance for efficiently assessing ADL routines via high-dimensional sensor data and supports future opportunities for capitalizing on smart home technologies that enable older adults to live alone safely, aging-in-place.