Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.745314
Title: Learning from biosignals
Author: Supratak, Akara
ISNI:       0000 0004 7223 6098
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
A long-standing goal of the biosignal analysis is to develop tools that can continuously collect quality biosignals, and algorithms to extract meaningful information from them. This can lead to a better understanding of our health that allows us to adjust our daily activity to be suitable for our well-being, and provide treatments promptly. However, the conventional approaches to analyze biosignals rely on the development of algorithms to extract features from the signals (i.e., hand-engineering features). It can be labor-intensive and time-consuming to develop algorithms to extract such features for particular applications repeatedly. Also, the existing machine learning based systems assume that there are annotations associated with the particular patterns of biosignals; such annotations are very expensive to obtain. In this thesis, our objective is to develop models that can automatically learn features from biosignals without utilizing any hand-engineering features and can extract meaningful information from biosignals even when the labels of biosignals are not available. Our first contribution is that we propose a method that can remotely and accurately estimate walking speeds for people with walking impairments. This work also motivates us to use deep learning to automate the expensive feature engineering process in the remaining contributions. Our second contribution is that we propose a model that can automatically extract meaningful features from raw scalp EEG signals for epileptic seizure detection. We demonstrate that it can detect seizures without utilizing any seizure annotations. Our third contribution is that we propose a model that can automatically learn features that are useful for sleep stage scoring from raw single-channel EEG data, and achieved similar performance compared to the state-of-the-art hand-engineering ones. We also demonstrated that our model can generalize to two sleep datasets that have different properties without any modifications to the model architecture. Our final contribution is that we propose a novel data-driven approach that employs a signal transcription model to capture the relationships between signals from multiple domains to detect desynchronized biosignals as anomalies without utilizing any annotations.
Supervisor: Guo, Yike ; Deisenroth, Marc Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.745314  DOI:
Share: