Clinical Manifestations

Area of Science:

• Gerontology
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Agitation and aggression (AA) in dementia present significant care challenges.
• Current detection methods are often delayed, leading to increased psychotropic medication use.
• Deep learning offers potential for enhanced early detection using multimodal data.

Purpose of the Study:

• To improve the prediction of agitation and aggression (AA) in advanced dementia patients.
• To develop a privacy-compliant method for early AA detection using fused data.
• To evaluate the efficacy of a deep learning model integrating video and wearable data.

Main Methods:

• Collected physiological data via EmbracePlus wristbands and skeletal keypoints from privacy-compliant video.
• Aligned and preprocessed multimodal data, including sampling, normalization, and feature extraction.
• Implemented a deep learning framework (RNN with GRU and attention) for temporal pattern analysis.

Main Results:

• A fused model integrating video and wearable data achieved 0.98 accuracy.
• The combined approach outperformed single-modality models (video: 0.94, wearable: 0.91).
• Preliminary analysis included five patients with advanced dementia and AA.

Conclusions:

• Integrating wearable physiological data with video-derived skeletal keypoints improves early agitation detection in dementia care.
• The developed approach is feasible, privacy-compliant, and enhances predictive accuracy.
• Further research aims to refine and scale this model for clinical application.