Analysis of physiological signals using machine learning methods

Oroojeni Mohamad Javad, Hooman. 2021. Analysis of physiological signals using machine learning methods. Doctoral thesis, Goldsmiths, University of London [Thesis]

Text (Analysis of physiological signals using machine learning methods)
COM_thesis_OroojeniM_2021.pdf - Accepted Version
Available under License Creative Commons Attribution Non-commercial No Derivatives.

Download (2MB) | Preview

Abstract or Description

Technological advances in data collection enable scientists to suggest novel approaches, such as Machine Learning algorithms, to process and make sense of this information. However, during this process of collection, data loss and damage can occur for reasons such as faulty device sensors or miscommunication. In the context of time-series data such as multi-channel bio-signals, there is a possibility of losing a whole channel. In such cases, existing research suggests imputing the missing parts when the majority of data is available. One way of understanding and classifying complex signals is by using deep neural networks. The hyper-parameters of such models have been optimised using the process of back propagation. Over time, improvements have been suggested to enhance this algorithm. However, an essential drawback of the back propagation can be the sensitivity to noisy data. This thesis proposes two novel approaches to address the missing data challenge and back propagation drawbacks: First, suggesting a gradient-free model in order to discover the optimal hyper-parameters of a deep neural network. The complexity of deep networks and high-dimensional optimisation parameters presents challenges to find a suitable network structure and hyper-parameter configuration. This thesis proposes the use of a minimalist swarm optimiser, Dispersive Flies Optimisation(DFO), to enable the selected model to achieve better results in comparison with the traditional back propagation algorithm in certain conditions such as limited number of training samples. The DFO algorithm offers a robust search process for finding and determining the hyper-parameter configurations. Second, imputing whole missing bio-signals within a multi-channel sample. This approach comprises two experiments, namely the two-signal and five-signal imputation models. The first experiment attempts to implement and evaluate the performance of a model mapping bio-signals from A toB and vice versa. Conceptually, this is an extension to transfer learning using CycleGenerative Adversarial Networks (CycleGANs). The second experiment attempts to suggest a mechanism imputing missing signals in instances where multiple data channels are available for each sample. The capability to map to a target signal through multiple source domains achieves a more accurate estimate for the target domain. The results of the experiments performed indicate that in certain circumstances, such as having a limited number of samples, finding the optimal hyper-parameters of a neural network using gradient-free algorithms outperforms traditional gradient-based algorithms, leading to more accurate classification results. In addition, Generative Adversarial Networks could be used to impute the missing data channels in multi-channel bio-signals, and the generated data used for further analysis and classification tasks.

Item Type:

Thesis (Doctoral)

Identification Number (DOI):


DFO, GANS, Swarm Intelligence, Deep Learning, Dispersive Flies Optimisation

Departments, Centres and Research Units:



31 October 2021

Item ID:


Date Deposited:

24 Nov 2021 15:15

Last Modified:

07 Sep 2022 17:19


View statistics for this item...

Edit Record Edit Record (login required)