Abstract/Summary

Healthcare systems worldwide are facing growing demands on their resources due to an ageing population and increase in prevalence of chronic diseases. Innovative residential healthcare monitoring systems, using a variety of sensors are being developed to help address these needs. Interpreting the vast wealth of data generated is key to fully exploiting the benefits offered by a monitoring system. This thesis presents the application of topic models, a machine learning algorithm, to detect behaviour patterns in different types of data produced by a monitoring system. Latent Dirichlet Allocation was applied to real world activity data with corresponding ground truth labels of daily routines. The results from an existing dataset and a novel dataset collected using a custom mobile phone app, demonstrated that the patterns found are equivalent of routines. Long term monitoring can identify changes that could indicate an alteration in health status. Dynamic topic models were applied to simulated long term activity datasets to detect changes in the structure of daily routines. It was shown that the changes occurring in the simulated data can successfully be detected. This result suggests potential for dynamic topic models to identify changes in routines that could aid early diagnosis of chronic diseases. Furthermore, chronic conditions, such as diabetes and obesity, are related to quality of diet. Current research findings on the association between eating behaviours, especially snacking, and the impact on diet quality and health are often conflicting. One problem is the lack of consistent definitions for different types of eating event. The novel application of Latent Dirichlet Allocation to three nutrition datasets is described. The results demonstrated that combinations of food groups representative of eating event types can be detected. Moreover, labels assigned to these combinations showed good agreement with alternative methods for labelling eating event types.