Dynamic Active Learning for Objective Qualification of Physical Activity in Free-Living Environments

Wearable sensors can provide valuable information about movement and everyday behaviour outside clinical settings. However, developing reliable human activity recognition systems normally requires large amounts of accurately labelled data, which can be costly and difficult to collect in real-life environments. This thesis investigates methods for reducing the need for detailed manual labelling, using Parkinson’s disease as a case study.

Three complementary machine-learning approaches are evaluated. The first uses broad subject-level labels to learn interpretable representations of activity patterns from wrist-worn sensor data. The second uses self-supervised learning to identify unusual movement patterns without requiring manual labels during the initial learning stage. The third uses active learning to adapt an activity-recognition model to a new individual by selecting only the most informative sensor segments for labelling.

The findings show that weak supervision, self-supervised anomaly detection, and targeted active learning can each reduce reliance on densely labelled wearable-sensor data. These approaches may support the development of more practical and personalised activity-recognition systems for use in free-living environments. The studies are proof-of-concept machine-learning evaluations and should not be interpreted as clinical diagnostic validation. The three approaches were assessed independently rather than as a single end-to-end system.