Fall Detection and Prevention for Memory Care through Real-Time Artificial Intelligence Applied to Video

点击翻译按钮获取中文摘要

Abstract In the US, Alzheimer’s disease (AD) is the single most expensive disease, the only one in the top six for which the number of deaths is increasing. The greatest costs are hospitalizations, where falls are the largest culprit, and frequent need for assistance with daily life activities. A fall safety system shows the potential to reduce costs and increase quality of care by reducing the likelihood of emergency events (e.g., detecting falls before a fracture occurs, reducing the number of repeat falls). Unfortunately, no fall detection and prevention technology has been developed specifically for the needs of dementia care where individuals (1) fall more frequently and (2) often cannot tell care staff how they fell, leading to increased use of Emergency Medical Services (EMS) when falls are unwitnessed to ensure affected individuals are safe. Our goal is to perform a randomized wait-list control clinical trial (n=460) of SafelyYou Guardian, an online fall detection system with wall-mounted cameras to automatically detect falls for residents with AD and related dementias (ADRD). The automation is based on algorithms that push the frontier of deep learning, a subfield of Artificial Intelligence (AI), with a human-in-the- loop (HIL). SafelyYou Guardian is designed to primarily operate in memory care facilities (defined herein as assisted living and skilled nursing facilities providing ADRD care). Deep learning has already revolutionized several fields: robotics, self-driving cars, social networks in particular. Our approach is anchored in novel algorithms developed at the Berkeley AI Research Lab (BAIR) and extended by SafelyYou for real-time detection of rare events in video. The HIL is operating from a call center, confirms the fall detection alerts provided by our artificial intelligence algorithms, and places a call to the communities, so an intervention can happen within minutes of the fall detection. Subsequently, an Occupational Therapist (OT) working from our office in San Francisco reviews the fall videos with the front-line staff over video conference and using our web portal to make recommendations on how to re-organize the resident space (intervention) to prevent future falls. We leverage our HIL paradigm, in which our deep learning approach identifies and pre-filters falls with high sensitivity followed by a human who confirms the fall with high specificity and calls the communities in case of detected fall. This project leverages past small scale clinical and technical pilots including 87 residents from 11 partner communities, and our experience with paid commitments for 480 residents from three partner networks. Past pilots leading to this NIH Phase II proposal include: · Pilot 1: Technical proof of concept with healthy subjects (200 acted falls). · Pilot 2: We demonstrated acceptance of privacy/safety tradeoffs by residents, family and staff, through the collection of 3 months of video data at WindChime of Marin, our first partner facility; we identified 4 total hours of fall data. This led to clinical benefits including an 80% fall reduction through the intervention of OT. · Pilot 3: We demonstrated scalability and acceptance by deploying the system in 11 communities, for 87 residents monitored by our system (offline, no HIL intervention). · Pilot 4: Small scale NIH Phase I clinical trial. We demonstrated the ability to perform real-time fall detection, with real-time intervention of the HIL through our partner company Magellan-Solutions which provides the 24/7 monitoring service for the facilities. We demonstrated that 93% of 89 falls were detected, that time on the ground was reduced by 42%, that the likelihood of EMS use was 50% lower with video available, and the that total facility falls including participants and non-participants decreased by 38%. The trial proposed for this NIH SBIR Phase II will provide clinical evidence that the preliminary trends observed experimentally (pilot 2) and at small scale (pilot 4) are true phenomena. It will use a wait-list control population (230 residents) to be compared to the population monitored with SafelyYou Guardian (230 residents). After crossover, the wait-list population will also benefit from the technology and be compared to itself before crossover.