Modelling Physical Activity Behaviour Changes for Personalised Feedback in a Health Education Application

Open-ended domains, where the focus is not about learning specific expert movements but about adopting healthy physical activity behaviours, require the use of unsupervised algorithms and artificial intelligence in education techniques for modelling evolving patterns from physical activity sensor data to enable feedback and personalisation. We present a suite of unsupervised window-based algorithms that detect physical activity changes aligned with learning objectives from accelerometer data. These are translated into learner model attributes and used to generate timely feedback. We illustrate our method in the context of a health education program that teaches adolescents about healthy physical activity behaviours through an application connected to a wrist-worn activity tracker. We present the feedback generated by our algorithms and report on the qualitative evaluation with four experts. We conclude that the automated feedback is useful, important and timely to leverage adolescents' physical activity learning.