IDENTIFICATION AND EVALUATION OF BEHAVIORAL SYMPTOMS IN DEMENTIA USING PASSIVE RADIO SENSING AND MACHINE LEARNING

Introduction Behavioral symptoms of Alzheimer's disease (e.g. delusions, wandering, aggression, sleep disturbance) lead to increased emergency room visits, caregiver burden, and transfers to memory care facilities. Sensor technologies may hold the potential to facilitate early detection and pre-emptive intervention for these symptoms by enabling continuous passive monitoring in a way that in-person monitoring may not be able to. We present preliminary data for such an approach using a device called the Emerald, developed at MIT, which emits low-powered radio signals and can identify and track parameters related to human behavior (sleep, motion, spatial motion, and respiratory rate) based on how these waves reflect off the human body. Artificial Intelligence (AI) algorithms elicit behavioral markers from sensor data. The device does not require any contact or direct interaction by the person being monitored, thus representing true passive sensing. Methods The Emerald device was installed in the rooms of two dementia patients (N=2) with behavioral symptoms residing in an assisted living facility (ALF). Motion data was gathered continuously for a period of three months and was mapped on to spatial location and time frame. Data processing and analysis occurred simultaneously during the collection period. Additionally, study staff administered weekly standardized assessments to both the participant (MMSE) and ALF staff (NPI-NH, CMAI, PAS) to augment data collected from the Emerald. Device data was compiled and made available to the study clinician for clinical analysis and identification of emergent behavioral complications. Results In both participants, device data were used to identify specific behavioral patterns. The device detected variations in behavior by time of day, escalations in pacing, and moments of restlessness throughout the night for both participants. For one participant, clinical interpretation of device data led to the proposition that the participant was experiencing Periodic Limb Movement Disorder, which was unbeknownst to the participant or clinician prior to study participation. The device was able to identify periodic spasms, which occurred when the person was asleep, and localize these to the patient's legs. The second participant showed increase pacing, wandering, and motor agitation before being hospitalized for heightened anxiety and aggression. Device data indicates the period prior to hospitalization featured increased movement episodes relative to this participant's baseline. Conclusions We propose that behavioral phenotyping using an AI-backed passive sensing approach is feasible and safe, and that this approach can help digitally phenotype behavior symptoms in dementia. While the device merits validation against the current standard of behavior measurement in dementia, its advantages include low cost and ongoing engagement, and continuous monitoring while giving patients the option of stopping monitoring at their discretion. Further studies evaluating sensitivity and reliability are warranted to validate the clinical utility of this device. This research was funded by This project is supported by an Innovations grant from the Massachusetts Institute of Technology.