Analysis and Classification of Resting-State EEG for Neurological Conditions Using Machine Learning and Deep Learning

This thesis is part of a project conducted by EPFL team and HES-SO Valais (Sion, Switzerland), that combines virtual reality (VR), temporal interference (TI) stimulation and electroencephalography (EEG) to perform a spatial navigation task in mild Traumatic Brain Injury (mTBI) patients. This task was specifically chosen as mTBI patients present spatial memory and navigation deficits, so the study investigates if neuromodulation can improve these impairments. As part of the project, I contributed to the design of the EEG acquisition protocol to minimize artifacts through personalized TI electrode placement and participant preparation, ensuring high-quality signals during simultaneous VR, TI, and EEG. Resting-state EEG (rsEEG) is commonly used to assess neurological conditions, but robust and interpretable deep learning (DL) models are still needed for classifying mTBI EEG. In this context, ShallowNet and EEGNet are the main convolutional neural networks used for EEG analysis, but they still present some important interpretability limitations. For this reason, xEEGNet, an interpretable and white-box neural network, could show strong potential in EEG classification, and this work wants to explore its possible application in this field. Due to recruitment limitations of mTBI patients, this study is a preliminary analysis on a public rsEEG dataset of 88 subjects (36 Alzheimer’s disease, 23 Frontotemporal Dementia, 29 healthy controls), aiming to be a reproducible pipeline for future mTBI datasets. After preprocessing, features including band power, phase synchronization, sample entropy, and Hjorth complexity were used as input for Random Forest (RF), Logistic Regression (LR), and a Multilayer Perceptron (MLP). The results obtained were compared with ShallowNet, EEGNet and xEEGNet, using Nested Leave-N-Subjects-Out cross-validation. RF and LR achieved median weighted accuracies of 56.2% and 57.0%, respectively, while MLP reached 53.8%. As regards the DL models, ShallowNet and EEGNet showed 66.5% and 56.1% median weighted accuracy respectively, compared with xEEGNet that reached 57.2%. The results demonstrate that xEEGNet has comparable performance but greater robustness than traditional Machine Learning (ML) approaches, with an autonomous extraction of the most relevant features. As regards DL models comparison, even if ShallowNet reached higher performance (that can be possibly due to overfitting), xEEGNet showed increased interpretability power and stability, making it a more suitable model for clinical applications. All models were able to distinguish Alzheimer’s patients from healthy controls, while Frontotemporal Dementia classification was more challenging due to overlapping EEG patterns with Alzheimer. To demonstrate this aspect and the generalizability of the pipeline in case of 2 classes (i.e. mTBI and HC), another analysis was performed without considering the FTD class. In this case all the models’ accuracy were clustered around 75%, where xEEGNet performed the best (79.1%), having at the same time a small accuracy variability range across splits (IQR= 0.094). These results show that xEEGNet provides a viable trade-off between performance, stability and interpretability, suggesting that this pipeline can be a solid foundation for future mTBI rsEEG studies. This work also aims to explore EEG integration into VR-TI neuromodulation protocols, with the future goal of creating personalized neurorehabilitation programs to optimize the treatment effectiveness in mTBI patients.