Effective EEG analysis for advanced AI-driven motor imagery BCI systems

  • Natasha Padfield

Student thesis: Doctoral Thesis


Developing effective signal processing for brain-computer interfaces (BCIs) and brain-machine interfaces (BMIs) involves factoring in three aspects of functionality: classification performance, execution time, and the number of data channels used. The contributions in this thesis are centered on these three issues. Contributions are focused on the classification of motor imagery (MI) data, which is generated during imagined movements. Typically, EEG time-series data is segmented for data augmentation or to mimic buffering that happens in an online BCI. A multi-segment decision fusion approach is presented, which takes consecutive temporal segments of EEG data, and uses decision fusion to boost classification performance. It was computationally lightweight and improved the performance of four conventional classifiers. Also, an analysis of the contributions of electrodes from different scalp regions is presented, and a subset of channels is recommended. Sparse learning (SL) classifiers have exhibited strong classification performance in the literature. However, they are computationally expensive. To reduce the test-set execution times, a novel EEG classification pipeline consisting of a genetic-algorithm (GA) for channel selection and a dictionary-based SL module for classification, called GABSLEEG, is presented. Subject-specific channel selection was carried out, in which the channels are selected based on training data from the subject. Using the GA-recommended subset of EEG channels reduced the execution time by 60% whilst preserving classification performance.Although subject-specific channel selection is widely used in the literature, effective subject-independent channel selection, in which channels are detected using data from other subjects, is an ideal aim because it leads to lower training latency and reduces the number of electrodes needed. A novel convolutional neural network (CNN)-based subject-independent channelsselection method is presented, called the integrated channel selection (ICS) layer. It performed on-a-par with or better than subject-specific channel selection. It was computationally efficient, operating 12-17 times faster than the GA channelselection module. The ICS layer method was versatile, performing well with two different CNN architectures and datasets.
Date of Award5 May 2022
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorPaul Murray (Supervisor) & Stephen Marshall (Supervisor)

Cite this