Analysis of active watch data after stroke

Abstract

Due to the rapid development of modern technology and the low cost of wearable devices, a very large amount of accelerometer data has been recorded and used in many different areas, particularly in medical research. However, the analysis of accelerometer data is very challenging due to its complex structure, i.e., the large noise-signal ratio, the very large amount of data and the heterogeneity in different data sets. In this thesis, we use wavelet in a functional data analysis (FDA) framework to analyse the data and apply this method to evaluate upper limb function after stroke, a difficult task in medical research. In addition to the commonly used features (Preece et al., 2009; Sekine et al., 1998) based on the wavelet energy preserving condition for accelerometer data under the discrete wavelet transform (DWT), we propose two new types of scalar features. They extract different types of information from the accelerometer data and use to predict upper limb function for stroke patients. To further investigate the ‘details’ based on wavelet-domain, under a Bayesian hierarchical model (NIG-MT), the wavelet coefficients with small values can be eliminated properly and efficiently with negligible loss from the total information. We will use the slide window approach and multivariate functional principal component analysis (fPCA) based on the small DWT tree structure. This further reduces the size of the data set and extracts the useful information from the pattern of small DWT tree structures in wavelet-domains. Classification and regression models are developed based on the small DWT tree structure. The models have been applied to distinguish between the different activities in the designed data and refine the new features in free-living data to assess the patients’ upper limb function respectively.