ETD

• electrocardiography
• electroencephalography
• Gaussian processes
• heart rate variability
• heart–brain axis
• machine learning
• RR- intervals
• Source Power Comodulation (SPoC)

The heart--brain axis (HBA) describes the continuous bidirectional interaction between the central nervous system and the cardiovascular system. Recent advances in artificial intelligence have created new opportunities for estimating cardiac activity directly from electroencephalography (EEG), potentially allowing short-term cardiac dynamics to be inferred without concurrent ECG measurements during inference. This thesis presents a physiologically informed machine learning framework for estimating beat-to-beat RR intervals from EEG recordings. The proposed pipeline combines signal preprocessing, diversity-aware EEG channel selection, beat-aligned filter-bank analysis, Source Power Comodulation (SPoC) spatial feature extraction, and Gaussian Process Regression. The framework was evaluated on recordings from 52 healthy subjects using both SUbject-Specific (SUS) and Leave-One-Subject-Out (LOSO) experiments.

The results demonstrated that the proposed framework accurately estimates beat-to-beat RR intervals while preserving the temporal characteristics of the underlying cardiac dynamics. In the SUS evaluation, the proposed approach achieved a mean RMSE of 18.7ms and a mean MAE of 15.5ms, with 33 of the 52 participants obtaining an RMSE below 20ms. The LOSO experiment further showed that 31 out of 52 previously unseen subjects achieved an RMSE below 20ms, indicating that part of the learned EEG--RR relationship can generalise across individuals. The predicted RR sequences were also used to reconstruct cardiac timing and derive time-domain heart rate variability (HRV) measures, providing an additional assessment of whether physiologically relevant cardiac dynamics were preserved.

Overall, the results support the feasibility of physiologically informed machine learning as a promising approach for non-invasive EEG-based estimation of short-term cardiac dynamics and for preserving RR dynamics sufficiently to support cardiac timing reconstruction, improving our understanding of the heart--brain interplay.