ETD

• brain-heart interplay
• deep learning
• domain generalization
• electroencephalography
• r-peak detection

Electroencephalography (EEG) is a non-invasive technique for recording brain activity. Cardiac activity is reflected in EEG recordings through two distinct manifestations: as QRS artifacts contaminating some electrode signals, and through the tight bidirectional brain-heart coupling, as reflected, for example, in the Heartbeat-Evoked Potential (HEP). Current literature, which largely reduces the problem to artifact suppression, is limiting for Brain-Heart Interplay (BHI) studies, which require precise R-peak localization even when artifacts are subtle or absent.

This thesis reframes the problem as continuous R-peak temporal localization directly from EEG, without requiring an auxiliary ECG channel. The work began with a rigorous benchmarking phase in which several deep learning architectures were compared. Building on this, targeted architectural modifications and domain generalization strategies were introduced to improve performance and address inter-subject physiological variability.

The proposed approach achieves an F1-score of approximately 70% on unseen subjects in WP2, a dataset without visible cardiac artifacts, rising to approximately 86% and 88% on datasets featuring more structured brain-heart coupling. On the MIT-BIH Polysomnographic Database, the model outperforms the artifact-detection method of Dora & Biswal (2019). These results establish a foundation for future Brain-Heart Interplay studies.