Year: 2026
Type: dissertation
Abstract: Dementia is a progressive neurodegenerative condition that significantly impairs cognitive function, autonomy, and the ability to perform daily activities, representing a growing global public health challenge. As the aging population increases, there is an urgent need for intelligent and scalable solutions capable of supporting continuous monitoring and early detection of behavioural changes in elderly individuals living with dementia. This dissertation presents the development of an Artificial Intelligence (AI) driven wearable monitoring system designed to recognize daily activities and detect anomalies in routine behaviour. The proposed approach integrates multimodal sensor data, including inertial and physiological signals, and evaluates both classical machine learning models and deep learning architectures for time-series analysis. The experimental work is structured into four stages. First, activity recognition models are evaluated using public and real-world datasets, where ensemble methods, particularly Light Gradient Boosting Machine (LightGBM) and eXtreme Gradient Boosting (XGBoost) consistently achieve the best performance, with accuracy reaching 0.98% in controlled scenarios, although subtle and static activities remain challenging to classify due to limitations of wrist-based sensing. Second, a synthetic framework generates realistic daily routines and simulate anomalous behaviour. Additionally, a sequence-based anomaly detection model based on a Bidirectional Gated Recurrent Unit (GRU) trained under a semi-supervised paradigm using only normal routines, achieves 99.4% accuracy and F1-scores above 0.95, showing that modelling behaviour as temporal sequences improves the detection of deviations from routine patterns. Third, a real-world multimodal dataset is collected to validate the data acquisition and processing pipeline under naturalistic conditions. The bestperforming model, LightGBM, achieved an accuracy of 0.73 and a weighted F1-score of 0.70, confirming the approach’s feasibility while revealing limitations in recognizing subtle and low-movement activities. Finally, building on this dataset, an integrated framework combines an XGBoost activity classifier with two complementary, fully unsupervised anomaly detectors, a Long Short-Term Memory (LSTM) Autoencoder and a rule-based duration detector, applied to continuous routines reconstructed from real sensor segments, achieving a mean F1-score of 0.798 and a false alarm rate of 2.9% on unseen routines without requiring any labelled anomalous data. Overall, this work demonstrates that integrating activity recognition with routinebased modelling provides a more effective framework for behavioural monitoring in dementia care. By moving beyond isolated activity classification towards sequence-based analysis, the proposed approach enables more meaningful detection of behavioural changes, supporting the development of proactive and context-aware monitoring systems.
Source: CiencIPCA
Author: Carvalho, Mariana da Costa
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221735356