Using plantar pressure for free-living posture recognition and sedentary behaviour monitoring

Health authorities in numerous countries and even the World Health Organization (WHO) are concerned with low levels of physical activity and increasing sedentary behaviour amongst the general population. In fact, emerging evidences identify sedentary behaviour as a ubiquitous characteristic of contemporary lifestyles. This has major implications for the general health of people worldwide particularly for the prevalence of non-communicable conditions (NCDs) such as cardiovascular disease, diabetes and cancer and their risk factors such as raised blood pressure, raised blood sugar and overweight. Moreover, sedentary time appears to be uniquely associated with health risks independent of physical activity intensity levels. However, habitual sedentary behaviour may prove complex to be accurately measured as it occurs across different domains, including work, transport, domestic duties and even lei¬sure. Since sedentary behaviour is mostly reflect as too much sitting, one of the main concerns is being able to distinguish among different activities, such as sitting and standing. Widely used devices such as accelerometer-based activity monitors have a limited ability to detect sedentary activities accurately. Thus, there is a need of a viable large-scale method to efficiently monitor sedentary behaviour. This thesis proposes and demonstrates how a plantar pressure based wearable device and machine learning classification techniques have significant capability to monitor daily life sedentary behaviour. Firstly, an in-depth review of research and market ready plantar pressure and force technologies is performed to assess their measurement capabilities and limitations to measure sedentary behaviour. Afterwards, a novel methodology for measuring daily life sedentary behaviour using plantar pressure data and a machine learning predictive model is developed. The proposed model and its algorithm are constructed using a dataset of 20 participants collected at both laboratory-based and free-living conditions. Sitting and standing variations are included in the analysis as well as the addition of a potential novel activities, such as leaning. Video footage is continuously collected using of a wearable camera as an equivalent of direct observation to allow the labelling of the training data for the machine learning model. The optimal parameters of the model such as feature set, epoch length, type of classifier is determined by experimenting with multiple iterations. Different number and location of plantar pressure sensors are explored to determine the optimal trade-off between low computational cost and accurate performance. The model s performance is calculated using both subject dependent and subject independent validation by performing 10-fold stratified cross-validation and leave-one-user-out validation respectively. Furthermore, the proposed model activity performance for daily life monitoring is validated against the current criterion (i.e. direct observation) and against the de facto standard, the activPAL. The results show that the proposed machine learning classification model exhibits excel-lent recall rates of 98.83% with subject dependent training and 95.93% with independent training. This work sets the groundwork for developing a future plantar pressure wearable device for daily life sedentary behaviour monitoring in free-living conditions that uses the proposed ma-chine leaning classification model. Moreover, this research also considers important design characteristics of wearable devices such as low computational cost and improved performance, addressing the current gap in the physical activity and sedentary behaviour wearable market.