An Unsupervised Data-Driven Anomaly Detection Approach for Adverse Health Conditions in People Living With Dementia:...

Area of Science:

• Health Informatics
• Machine Learning
• Gerontology

Background:

• Remote health monitoring using sensors aids early detection of health deterioration in dementia patients.
• Existing anomaly detection methods struggle with noisy, multivariate data and lack generalizability.
• Need for lightweight, online unsupervised learning for real-time health anomaly detection in dementia care.

Purpose of the Study:

• To develop an online, lightweight unsupervised learning approach for detecting adverse health conditions in dementia patients using activity changes.
• To evaluate the effectiveness of the proposed method against state-of-the-art techniques on real-world data.
• To apply the approach to household movement data for detecting urinary tract infections (UTIs) and hospitalizations.

Main Methods:

• Utilized Contextual Matrix Profile (CMP), an ultrafast, distance-based anomaly detection algorithm.
• Generated CMPs from daily aggregated household movement data (sensor counts, duration, hourly patterns).
• Computed normalized anomaly scores using univariate and multidimensional CMP, comparing against other outlier detection methods.

Main Results:

• The multidimensional CMP achieved an average recall of 84.3% with a 5.1% alert rate for detecting UTIs and hospitalizations.
• Demonstrated superior balance of recall and precision compared to Angle-Based Outlier Detection, Copula-Based Outlier Detection, and Lightweight Online Detector of Anomalies.
• Identified midnight to 6 AM bathroom activity as a key digital biomarker for UTIs, contributing ~30% to the anomaly score.

Conclusions:

• This is the first real-world study adapting CMP for continuous anomaly detection in healthcare for dementia patients.
• CMP offers speed, accuracy, simplicity, noise reduction, pattern detection, and explainability for clinical practitioners.
• The multidimensional CMP effectively addresses anomaly scoring in multivariate time-series health data, showing clinical promise for dementia and other conditions.