Multiple-Instance Learning for Sparse Behavior Modeling from Wearables: Toward Dementia-Related Agitation Prediction.

Area of Science:

• Gerontology
• Behavioral Science
• Wearable Technology

Background:

• Agitation in persons with dementia (PWD) presents significant risks to both individuals and caregivers.
• Passive sensing and continuous behavior tracking offer potential for early intervention.
• Predicting agitation from sensor data in real-world settings remains a research challenge due to data sparsity and weak annotations.

Purpose of the Study:

• To propose and evaluate a novel approach for predicting agitation episodes in PWD using wrist motion data.
• To address challenges of sparsity, unpredictability, and weak annotations in behavior prediction.
• To compare the efficacy of multiple-instance learning (MIL) models against single-instance models.

Main Methods:

• A transdisciplinary study involving dementia dyads in home settings.
• Continuous motion sensing using smartwatch inertial sensors from PWD.
• Active marking of agitation episodes by caregivers and analysis using MIL models.

Main Results:

• MIL-based models demonstrated potential in inferring agitated behavior from sparsely labeled data.
• Comparison showed MIL models outperformed single-instance models in this context.
• The study analyzed data from 10 residential deployments, each lasting 30 days.

Conclusions:

• MIL models are effective for behavior inference from wearable sensor data in real-world, 'in-the-wild' settings.
• This approach holds promise for preventing the escalation of agitation episodes in PWD.
• Wearable technology combined with advanced machine learning can improve dementia care.