Back to search

National Technical University of Athens (NTUA)

Τεχνικές βελτίωσης επεξηγησιμότητας συστημάτων τεχνητής νοημοσύνης σε δεδομένα ηλεκτροεγκεφαλογραφήματος

Abstract

dc:description

Electroencephalography (EEG) is a neurophysiological technique offering high temporal resolution, crucial for understanding brain function and developing Brain-Computer Interfaces (BCIs). Despite the rich information contained in EEG signals, their analysis and interpretation present significant challenges, particularly in extracting precise spatial information from scalp sensors and understanding the underlying neural mechanisms driving specific responses or behaviors. Simultaneously, the increasing use of deep learning (DL) models for EEG signal classification, while yielding impressive performance levels, often creates "black boxes" that make it difficult to justify predictions, limiting trust and clinical adoption. This thesis aims to bridge the gap between the high performance of Artificial Intelligence (AI) models on EEG data and the need for physiologically valid and anatomically informed interpretability. To this end, a comprehensive framework combining advanced signal analysis techniques, source localization, and explainable deep learning was developed. The work was structured across three main axes. Firstly, the value of Time-Frequency Analysis was explored as a powerful feature engineering tool, demonstrating its effectiveness in conjunction with a computationally efficient Convolutional Neural Network for classifying EEG signals in a mental fatigue detection context. This part highlighted the potential for achieving high accuracy (97%) with low computational cost, while also identifying, through explainability techniques like SHAP, spectral patterns (in theta and alpha bands) consistent with fatigue literature, underscoring the utility of TFR analysis and explainable models beyond the central BCI application. Subsequently, a critical review of current EEG Source Localization methods was presented. Their theoretical foundations, various algorithm categories (e.g., beamformers, minimum norm solutions), applications, and inherent limitations were analyzed. This review established the necessity of moving from sensor-level analysis to brain source level to achieve accurate anatomical interpretation, justifying the choice of a source localization approach (beamformer) for estimating cortical activity in the main application of the thesis. The third and central part of the work applied the combination of these techniques to a motor imagery BCI system, integrating explainability techniques (SHAP, Integrated Gradients). By using features extracted from source-localized EEG data (using a beamformer and Desikan–Killiany parcellation) as input to a convolutional neural network, competitive accuracy in classifying motor imagery of hands, feet, and tongue was achieved. The most significant finding emerged from applying the explainability techniques: the most important regions highlighted for each motor imagery class showed strong agreement with established neuroscientific knowledge and the somatotopic organization of the motor cortex, essentially reproducing a functional "homunculus" and laterality phenomena observed in fMRI studies. This convergence confirmed that the classifier relies on genuine neuroanatomical patterns rather than random features, resolving the ambiguity of sensor-level analysis and providing anatomically specific explanations. These results have significant implications for both clinical BCI applications and neuroscience. They enhance trust in the use of DL models in clinical settings, enable the provision of anatomically informed neurofeedback, and offer a tool for monitoring neuroplastic changes. At a scientific level, the framework can serve as a tool for verifying known patterns and discovering less understood aspects of brain function. Despite the limitations (finite source localization accuracy, sensitivity to deep sources, dataset size), the thesis demonstrates that a combined framework can achieve high performance without remaining a "black box." Future directions include improving spatial/temporal resolution, integrating multimodal data, applying advanced explainability techniques (such as interaction analysis), and application to clinical populations and other EEG tasks. Overall, the work lays the foundation for a new generation of brain-computer interfaces and neuroscientific tools where transparency and understanding of underlying neural mechanisms are central.

Degree

thesis:*
Grantor dc:publisher
National Technical University of Athens (NTUA)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Zorzos, Ioannis
  • Ζώρζος, Ιωάννης

Subjects

dc:subject × 22

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/58995

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Zorzos, Ioannis; Ζώρζος, Ιωάννης. Τεχνικές βελτίωσης επεξηγησιμότητας συστημάτων τεχνητής νοημοσύνης σε δεδομένα ηλεκτροεγκεφαλογραφήματος. National Technical University of Athens (NTUA), 2025. http://hdl.handle.net/10442/hedi/58995