RT info:eu-repo/semantics/doctoralThesis
T1 Enhancing EEG-based brain-computer interfaces with deep learning and explainable artificial intelligence
A1 Pérez Velasco, Sergio
A2 Universidad de Valladolid. Escuela de Doctorado
K1 Neuroingeniería
K1 Neural engineering
K1 Neuroingeniería
K1 Brain-computer interface (BCI)
K1 Interfaz cerebro-computadora
K1 Electroencephalography (EEG)
K1 Electroencefalografía
K1 Deep learning (DL)
K1 Aprendizaje profundo
K1 2490 Neurociencias
AB For decades, researchers have envisioned utilizing brain signals to interact with computers and assistive devices directly through thought alone. However, electroencephalography (EEG)-based brain–computer interfaces (BCIs) still face reliability issues, prolonged calibration, and incomplete knowledge of neural dynamics. These obstacles hamper the practical deployment of BCIs in real-world scenarios, including robust assistive systems for people with severe motor impairments.In this context, this Doctoral Thesis focuses on improving motor imagery (MI) EEG-based BCIs by leveraging state-of-the-art deep learning (DL) methods and explainable artificial intelligence (XAI) techniques. It first introduces EEGSym, a novel DL architecture developed to tackle the well-known variability in EEG signals that often leads to "BCI inefficiency". EEGSym incorporates hemispheric symmetry, inception modules, and residual connections to robustly decode MI commands directly from raw EEG. Furthermore, it is enhanced with data augmentation strategies and a multi-dataset transfer learning pipeline, eliminating the need for lengthy calibration across different users and recording sessions.Alongside the proposed DL architecture, a XAI-driven methodology was introduced to reveal the spatio-temporal features most relevant for MI decoding. By adapting Shapley additive explanations (SHAP) to the EEG domain, this work uncovers the neural patterns and cortical regions the model relies upon, highlighting not only primary sensorimotor areas but also the prefrontal and parietal cortices. These findings broaden our understanding of brain activation beyond conventional sensorimotor paradigms, helping to optimize electrode setups and improve BCI usability. In particular, an eight-channel montage was selected through SHAP analysis to minimize equipment needs without compromising accuracy.Lastly, this thesis explores motor execution (ME) to MI direct transfer learning as a means to improve MI sessions. Models exclusively trained on ME data were found to classify MI tasks with comparable performance to MI-trained models, pointing to a shared neural substrate that may help circumvent challenges, such as low user engagement or unobservable MI attempts. By unifying these insights (robust DL architectures, explainable feature attribution, and ME to MI transfer), this work provides a pathway toward more efficient, reliable, and versatile BCIs.In the three studies that comprise this Doctoral Thesis, we evaluated the proposed solutions for novelty, robustness, and flexibility.  First, in a strictly inter-subject evaluation that merged five public datasets and 280 participants, EEGSym achieved mean accuracies of 88.6 ± 9.0% on Physionet, 83.3 ± 9.3% on OpenBMI, 85.1 ± 9.5% on Kaya2018, 87.4 ± 8.0% on Meng2019, and 90.2 ± 6.5% on the Carnegie Mellon dataset while relying on only 16 electrodes; consequently, 268 of the 280 users (95.7%) surpassed the 70% BCI-control threshold, reducing the proportion of `BCI-inefficient'' users to below 5%.  Second, coupling EEGSym with SHAP-based explainability revealed that motor imagery engages a distributed network extending from the sensorimotor cortex to prefrontal and posterior-parietal regions.  Based on this insight, we selected an eight-channel montage centered on F7/F8 and bilateral centro-parietal sites that maintained high accuracy (86.5 ± 10.6% on Physionet and 88.7 ± 7.0% on the Carnegie Mellon dataset), thus enabling reduced EEG setups for MI-based BCIs with comparable performance. Finally, we show for the first time that a model trained exclusively on ME data can decode MI with comparable accuracy, demonstrating direct ME to MI transfer that eliminates user-specific MI calibration.Taken together, these results outline a path toward MI-based BCIs that are accurate, interpretable, and easy to deploy while remaining resilient to inter-subject and inter-session variability. Although these tests were conducted offline with healthy volunteers, the 95% control rate, portable eight-electrode configuration, and calibration-free ME to MI transfer establish a firm basis for future clinical and real-time studies, accelerating the transition of BCIs from laboratory prototypes to reliable assistive technologies for daily use.
YR 2026
FD 2026
LK https://uvadoc.uva.es/handle/10324/84125
UL https://uvadoc.uva.es/handle/10324/84125
LA eng
NO Escuela de Doctorado
DS UVaDOC
RD 19-abr-2026