ROBUST DATA-DRIVEN PREDICTIVE MODEL FOR BRAIN-COMPUTER INTERFACE

Abstract

A Brain-Computer Interface (BCI) system enables communication and control between a user and an external device without relying on peripheral and muscular activity. Effective control of such a device hinges on accurately recognizing and decoding intricate brain activity patterns generated by the user. The goal of this PhD project is to develop a robust model for predicting human mental intentions using electroencephalography (EEG) signals. EEG, a widely used non-invasive method for monitoring brain activity, is considered due to its ethical considerations, relatively low cost, and its ability to provide a high temporal resolution of received signals. The robustness of the system is verified based on the classification accuracy with respect to the previously unknown subjects such that the performance of subject-independent (SI) BCI system could be evaluated. An essential challenge in BCI research is developing a classifier capable of interpreting users' mental states from EEG data collected from independent subjects. The focus on SI classification is justified because it can lead to BCIs that eliminate the need for individual calibration processes. Over the past few years, deep neural networks (DNNs) in general, and in particular Convolutional Neural Networks (CNNs), have shown impressive training efficiency and performance, leading to the development of state-of-the-art architectures for accurate EEG classification. In this Thesis, to further enhance the performance of the CNN in SI classification, multi-subject ensemble CNN (MS-En-CNN) models are designed. These are the ensembles of CNN classifiers where each base classifier is built using data aggregated from multiple subjects. Based on the distribution of subject-specific data for training and tuning the base learners of the ensemble, three design strategies for MS-En-CNN are introduced: Subject-Specific Training and Model Selection (SS-TM), Subject Pairs Training and Model Selection (SP-TM), and Delete-a-Subject-Jackknife (DASJ) approach. The predictive performance of the proposed techniques is evaluated across two BCI paradigms, namely motor imagery (MI) and P300, using various publicly available datasets. Empirical results show that with any of the presented strategies constructing MS-En-CNN leads to a significantly better SI classification performance with respect to the average performance of the base CNN classifiers. Moreover, MS-En-CNN notably enhances average classification accuracy compared to a single CNN trained on pooled data from training subjects. Among the three strategies, the latter approach, a jackknife-inspired deep learning technique, emerges as the most promising one. It is then benchmarked against state-of-the-art methods, highlighting its superior performance in single-trial SI classification. While these results show potential for datasets with a small number of subjects, addressing computational requirements for large-scale datasets involves extending this approach through the consideration of K-fold cross-validation (CV). In this extended approach, instead of deleting a single subject to form a jackknife sample, a group of K subjects is set aside. On one of the largest MI datasets a K-fold CV-based MS-En-CNN demonstrated a statistically significant improvement (p < 0.001) over the best previously reported results. In addition to MS-En-CNN, proven as a simple yet effective method to enhance the performance of existing CNN models, a new adaptive boosting strategy on the basis of CNN base classifiers (AdaBoost-CNN) with iterative oversampling is proposed. This innovative approach is contrasted with the conventional sample reweighting method, showcasing its potential. Encouraged by promising results, the AdaBoost-CNN warrants further investigation. Overall, this study highlights the effectiveness of MS-En-CNN and AdaBoost-CNN and offers valuable insights that pave the way for further advancements in SI classification within BCI applications.