JMIR Publications

ABSTRACT

Background:

Delirium superimposed on dementia is associated with poor outcomes yet remains underdetected in home settings. Current detection relies on face-to-face clinical assessment (e.g., the Confusion Assessment Method (CAM) criteria) rarely applied outside hospitals.

Objective:

This proof-of-concept study developed a theory-driven framework for detecting delirium-consistent anomalous patterns in home-dwelling people with dementia, using passive smart home sensor data.

Methods:

Individualized anomaly detection algorithms, including Isolation Forest and Long Short-Term Memory (LSTM) models, were applied to identify delirium-related anomalies within each participant. Predictor features consisted of theory-driven digital markers approximating key CAM criteria, including agitation, disrupted sleep–wake cycles, and disorientation (indexed by activity entropy), along with clinically relevant indicators such as physiological instability (early warning scores) and urinary tract infections. Multimodal smart-home sensor data from 17 individuals with dementia were analyzed.

Results:

Using matched thresholds, Isolation Forest and LSTM models each identified 71 anomalies, with the Isolation Forest detecting a median of 10.2% anomalous days per individual and anomalies typically occurring in short temporal clusters; agreement between methods was 17%. Feature importance analyses indicated that activity entropy, sleep quality, and early warning scores were the most influential features, with stronger inter-feature correlations observed during anomaly versus non-anomaly periods.

Conclusions:

This study demonstrates technical feasibility of detecting delirium-related anomalies through passive smart home monitoring. While lacking ground truth validation, the approach shows promise for early intervention in community settings. Future validation studies with clinically confirmed delirium labels are essential. Clinical Trial: not applicable