PRP034: Automated sensor-based detection of neuropsychiatric symptoms in dementia using wearable devices and machine learning
Elena Guseva, MA; Isabelle Vedel, MD, PhD; Zahinoor Ismail, MD; Walter Wittich, PhD, FAAO CLVT; Ovidiu Lungu, PhD; Sondra Goldman, BA, MA; Machelle Wilchesky, PhD
Context: The development of automatic real-time detection of neuropsychiatric symptoms (NPS) using objective physiological data has the potential to revolutionize individualized dementia care. Current NPS detection tools are retrospectively applied observational scales that rely on subjective informed caregiver judgment or clinician assessment, and require considerable resources. Objective: To identify instances of NPS such as agitation/aggression and apathy in persons with dementia (PwD) using automated sensor-based detection methods. Study Design and Analysis: Our feasibility pilot study will employ wearable sensor technology to detect episodes of agitation/aggression and apathy in PwD. Wearable devices provide physiological signals related to movement and activity patterns and sympathetic system activation measured via electrodermal activity (EDA) and heart rate variability (HRV). An accelerometer activity tracker will be worn by study participants for the 28-week study follow-up period. EDA and HRV data will be recorded using an Empatica E4 device. Scoring algorithms derived from raw accelerometry and heart rate signal data, will provide estimates of activity levels and sleep-wake patterns in relation to NPS outside of therapeutic sessions. EDA and HRV as known indicators of stress and disengagement respectively will monitor states of agitation/aggression and apathy during the sessions. To validate our approach, NPS outcomes identified via these wearable devices will be compared with outcomes obtained using the Neuropsychiatric Inventory-Nursing Home version. The Engagement of a Person with Dementia Scale will be used to code video recorded therapy sessions. Episodes of NPS identified by these scales will be cross verified with signal processing of the wearable sensor data streams using machine learning algorithms and statistical techniques. Setting: A large 387-bed long-term care facility.
Population Studied: 10-12 residents with moderate to advanced Alzheimer’s disease (AD) with agitation/aggression and/or apathy in the 2-week period prior to enrolment. Instruments: 3-axis Accelerometer and Empatica E4 devices. Outcome Measures: 1) Aggression/agitation; 2) apathy; and 3) engagement Expected Outcomes: We expect to demonstrate that automated sensor-based detection methods can feasibly, acceptably and accurately detect episodes of NPS in PwD.