Signal processing and analytics of multimodal biosignals

Abstract

Biosignals have been extensively studied by researchers for applications in diagnosis, therapy, and monitoring. As these signals are complex, they have to be crafted as features for machine learning to work. This begs the question of how to extract features that are relevant and yet invariant to uncontrolled extraneous factors. In the last decade or so, deep learning has been used to extract features from the raw signals automatically. Furthermore, with the proliferation of sensors, more raw signals are now available, making it possible to use multi-view learning to improve on the predictive performance of deep learning. The purpose of this work is to develop an effective deep learning model of the biosignals and make use of the multi-view information in the sequential data. This thesis describes two proposed methods, namely: (1) The use of a deep temporal convolution network to provide the temporal context of the signals to the deeper layers of a deep belief net. (2) The use of multi-view spectral embedding to blend the complementary data in an ensemble. This work uses several annotated biosignal data sets that are available in the open domain. They are non-stationary, noisy and non-linear signals. Using these signals in their raw form without feature engineering will yield poor results with the traditional machine learning techniques. By passing abstractions that are more useful through the deep belief net and blending the complementary data in an ensemble, there will be improvement in performance in terms of accuracy and variance, as shown by the results of 10-fold validations.