Introduction

Abstract

This work is aimed to demonstrate the contribution of feature learning in classification applications, especially for biomedical data. Unsupervised learning and subject-independent settings are desirable deployment manners for the area. One reason is that these abilities can help deploy classification tasks in out-of-the-lab wearable devices. Another reason is reducing labour costs and subjectivity associated with human involvement. In this thesis, three examples are studied: human body movement assessment where acceleration data is used (Case 1), respiratory artefact removal where lung function tests are carried out (Case 2), and spike sorting for electrophysiological data (Case 3). Manual classification is often considered the de facto standard practice but it is time-consuming and subjective. Existing automated efforts have been predominantly designed for subject-dependent settings. Unsupervised sorters using simple statistics have only yielded modestly accurate results.