Related Experiment Video
Updated: May 9, 2026
11:21
Methodology for Establishing a Community-Wide Life Laboratory for Capturing Unobtrusive and Continuous Remote Activity and Health Data
Published on: July 27, 2018
Unsupervised Convolutional Neural Networks Using Home Presence Sensors for Behavioural Anomaly Detection in Older
Henry Llumiguano1, Jesús Martinez-Del-Rincon2, Jesús Fernandez-Bermejo1
1School of Computer Engineering, University of Castilla-La Mancha, Ciudad Real 13071, Ciudad Real, Spain.
International Journal of Neural Systems
|May 8, 2026
Summary
This study introduces a noninvasive monitoring system using presence sensors and deep learning to detect behavioral changes in older adults living alone. The system effectively identifies deviations potentially linked to cognitive or functional decline, aiding early assessment.
Area of Science:
- Gerontology
- Artificial Intelligence
- Health Informatics
Background:
- Aging population and desire to age in place present safety challenges for older adults, especially those living alone.
- Early detection of cognitive or functional decline is crucial for timely intervention and maintaining well-being.
- Noninvasive monitoring methods are needed to assess behavior changes without clinical diagnosis.
Purpose of the Study:
- To investigate the feasibility of a noninvasive monitoring system for detecting behavioral deviations in older adults.
- To utilize unsupervised deep learning models to identify changes potentially indicative of cognitive or functional decline.
- To validate the system's performance across diverse datasets, including real-world household monitoring.
Main Methods:
- Employed a noninvasive monitoring system with low-cost presence sensors.
- Utilized unsupervised convolutional autoencoders to model typical activity patterns and detect anomalies via reconstruction errors.
- Validated the approach on public (CASAS Aruba, HH120) and proprietary (MIRATAR) datasets, using synthetically injected anomalies for evaluation.
Main Results:
- The system demonstrated consistent performance across all three datasets.
- Higher F1-scores were achieved with more stringent anomaly detection thresholds.
- Average F1-scores at the 98th percentile threshold reached 0.96 (MIRATAR), 0.98 (CASAS Aruba), and 0.99 (CASAS HH120).
- The autoencoder approach outperformed classical unsupervised methods like Isolation Forest and One-Class SVM.
Conclusions:
- Unsupervised deep learning with presence sensors offers a viable noninvasive method for detecting behavioral changes in older adults.
- The proposed system can identify deviations potentially related to cognitive or functional decline, supporting early assessment.
- The approach shows promise for enhancing the safety and well-being of aging individuals living independently.
