The ageing population and the preference to remain at home pose significant challenges for ensuring the safety and well-being of older adults, particularly those living alone. This study investigates whether deviations in daily behaviour can be identified through a non-invasive monitoring system based on simple, low-cost presence sensors combined with unsupervised deep learning models. The objective is to detect behaviour changes potentially related to cognitive or functional decline, for later assessment, without performing any clinical diagnosis. Unsupervised convolutional autoencoders are used to model typical activity patterns and identify deviations through reconstruction errors. The approach was validated using three datasets: two public datasets (CASAS Aruba and HH120) and a proprietary dataset collected from 50 real-world households under long-term monitoring conditions (MIRATAR dataset). As none of these datasets contain clinically labelled anomalous events, system evaluation relies on the controlled injection of synthetic anomalies into the test data. The results demonstrate consistent performance across the datasets, with improvements in F1-score as more stringent anomaly detection thresholds are applied. At the 98th percentile threshold, the system achieved average F1-scores of 0.96 (MIRATAR), 0.98 (CASAS Aruba), and 0.99 (CASAS HH120). In addition, the proposed autoencoder-based approach was compared with classical unsupervised anomaly detection methods, including Isolation Forest and One-Class SVM, across all evaluated datasets.
Llumiguano et al. (Fri,) studied this question.