<< Back
Can machine learning algorithms detect agitation in dementia patients from video data?
Shehroz Khan, Ph.D.
University of Toronto
Toronto, Canada
Background
Studies show that as Alzheimer’s advances, individuals may experience symptoms such as mood and behavior changes including anxiety, aggression, and agitation. Scientists are studying systems capable of monitoring these symptoms in order to better understand their nature and pattern, provide objective measures of their frequency and severity, and track responses to interventions.
Many of these studies focus on the use of wearable sensors to identify patterns of activity associated with the symptoms of agitation or aggression. However, wearable devices are intrusive and long-term adherence is challenging. In addition, these devices are not able to capture all behaviors associated with agitation and aggression, nor are they able to differentiate between some normal activities (such as dancing) and agitation and other risk-associated behaviors.
Video cameras are widely used to monitor the common areas of residential care environments housing people living with dementia. Video cameras may be an inexpensive and unobtrusive alternative to monitor and detect behaviors associated with agitation or aggression in real-time.
Research Plan
Overcoming issues of security and storage of large amounts of data, Dr. Shehroz Khan and colleagues created a large and unique databank of videos in a dementia care environment and will study this annotated video data using an advanced computer science technique called machine learning.
The researchers will use the data collected in their previous Detecting Agitation Study, which used two types of sensors, video cameras and wearable devices. They will develop and evaluate machine learning algorithms that are able to classify whether individuals in 5-second video snippets show agitation or not. The research team will study a number of machine learning algorithms and identify the most accurate one. They will also develop new approaches to combine previously collected data from wearable devices and video data to create a more accurate system for the detection of agitated behavior.
Impact
This project may lead to the development of software capable of monitoring video streams in real-time to detect agitation in residential care environments, without the need for humans to monitor the video stream. In addition, capturing the moments leading up to and during an agitation event may help clarify triggers and provide context for the event, helping caregivers understand the behavior and personalize the appropriate intervention.
Back to Top