Inferring Physical Agitation in Dementia using Smartwatch and Sequential Behavior Models

Abstract

Caregivers for community-dwelling persons with dementia (PWD) often struggle to handle agitation episodes of the PWD. Such episodes pose a major health risk for both PWD and caregivers. Timely detection can prevent escalation of such events and their hazardous consequences. Wearable sensors are often employed for continuously sensing physiological variables, however, reliable inference of behavioral events using such signals is still an open research. Inferring such behavior in uncontrolled residential environments is challenging, especially due to the prevalence of unpredictable and wide-variety activity patterns. This paper presents a novel methodology to infer the onset of agitation episodes from PWD inertial motion data. As part of a transdisciplinary study, inertial sensors on smart watches are used to unobtrusively capture motion patterns during month-long deployments from eight clinically diagnosed PWD residing in their homes. These patterns are analyzed to build a sequential behavior model using long short-term memory (LSTM) based recurrent neural network. The performance of this model in inferring the onset of agitation episodes is evaluated using data from real deployments. This paper shows the potential of such models in behavior inference using wearable sensors and in developing intervention systems for real-world deployments.