Wearable Sensors and AI to Recognize and Evaluate IADLs

Project Summary/Abstract: Mild cognitive impairment (MCI) reportedly affects up to 24% of older adults and involves an associated decline in functional mobility. Individuals with MCI experience decreased balance, decreased gait speed, altered gait parameters, and even a greater risk of falling. Currently, clinical measures

of balance and mobility only moderately predict dysfunction associated with MCI. Recent studies using cognitive-motor dual-tasks were promising. This is done by attempting to increase the complexity brain processing demand by combining a movement task, such as gait, with a cognitive task, such as counting down

from a random number by 3's. Current studies exploring dual-task assessments offer conflicting results in their ability to detect MCI, limiting their reliability. We hypothesize that current clinical testing paradigms lack ecological validity and functional task performance. This oversight limits the complexity of performing self-

selected movements and the associated cognitive overlay required for instrumental activities of daily living (IADLs) engagement. It may be this additional real-world complexity that results in performance difficulty due to MCI and/or altered functional movement. The objective of this project is to combine the expertise of physical

and occupational therapy and biomedical engineering to use advancing wearable technology of inertial measurement units (IMU) and advanced deep learning algorithms to develop a framework for recognizing and determining ability to perform naturalistic movements in an ecologically valid setting. To accomplish this, we

will recruit individuals with MCI (n=15) and cognitively healthy (n=15) adults from 60-75-years old to perform a simulated IADL involving a series of tasks that include at least 10 repetitions of discrete activities that are involved in typical grocery shopping (e.g. carrying a basket, reaching up for an item, etc.). IMU data will be

labeled using video ground truth, allowing files consisting of a full activity stream (the complete grocery shopping task) as well as files segregating discrete activities (retrieving a can of soup from a shelf). We will then develop and validate a deep learning framework in order to identify each discrete activity performed in the

IADL task in both those with MCI and cognitively normal older adults (Aim 1). Additionally, we will use feature extraction methods to identify specific kinematic performance parameters of each gait and non-gait based activity (Aim 2). We then use this pilot kinematic data to identify sample sizes of future studies with adequate

power and effect size to provide a robust framework to use naturalistic movements to detect movement dysfunction in those with MCI. By achieving these aims, we establish a state-of-the-art framework that may ultimately be used for detecting and measuring performance and safety of IADL engagement in older adults.

Our long-term goal is to develop a naturalistic and highly reliable method that may provide early identification of cognitive and movement dysfunction in order to initiate treatment before the onset of dementia, as well as to provide a functional test to measure potential longitudinal functional changes.