Digital phenotyping of agitation in AD/ADRD: A systematic review of multimodal sensing and machine learning pipelines

Document Type

Article

Publication Date

8-19-2026

Publication Title

Smart Health

Abstract

Agitation is a common and burdensome neuropsychiatric symptom among people living with Alzheimer's disease and related dementias (AD/ADRD), affecting approximately 40–60% of this population. It is associated with increased caregiver burden, elevated injury risk, earlier institutionalization, and reduced quality of life. Conventional assessment instruments, including the Cohen-Mansfield Agitation Inventory, Pittsburgh Agitation Scale, and Neuropsychiatric Inventory, remain widely used in clinical and care settings; however, their episodic and retrospective nature limits timely detection of short, context-dependent agitation events. To address this gap, recent research has investigated continuous monitoring approaches based on wearable and physiological sensing, ambient and in-home sensors, vision and depth systems, audio analytics, radar/RF sensing, real-time location systems, and multimodal fusion.

This systematic review examines 25 peer-reviewed studies published between 2015 and 2025 that applied sensing technologies and machine learning methods to detect or predict agitation, aggression, or related challenging behaviors in AD/ADRD populations. The review is structured around three questions: how agitation is defined and labeled, what modality-specific markers are used to capture agitation-related behaviors, and how machine learning pipelines are implemented and evaluated. Although many studies report encouraging results, the evidence base remains fragmented. Cross-study comparison is limited by heterogeneous ground-truth definitions, coarse temporal labels, small and predominantly single-site cohorts, incomplete reporting of annotation procedures, and inconsistent validation designs. Future work should prioritize temporally explicit and standardized labeling, deployment-oriented evaluation using temporal and cross-site holdouts, and privacy-conscious multimodal systems that are robust, interpretable, and feasible for integration into routine AD/ADRD care.

Volume

41

Publisher

Elsevier

Rights

© 2026 Published by Elsevier Inc.

Link to Full Text

Share

COinS