Wearable Electrocardiography-Derived Respiration: Use of W-Net for Signal Separation

Continuous cardiorespiratory monitoring has become increasingly important in modern healthcare, spanning from clinical diagnosis to remote patient monitoring through wearable technologies. Among physiological signals, the electrocardiogram (ECG) is one of the most widely acquired due to its non-invasive nature, ease of acquisition, and clinical relevance. Vice versa, respiratory activity, despite providing important information about an individual's health status, is often reduced to breath rate (BR) estimation. To address this limitation, considerable research has focused on deriving respiratory information directly from ECG recordings through Electrocardiogram-Derived Respiration (EDR). Respiration induces physiological modulations in the ECG signal, enabling the extraction of respiratory information without dedicated respiratory sensors. This approach is particularly attractive for wearable and ambulatory monitoring systems, where reducing the number of sensors can improve usability and facilitate continuous monitoring in daily-life settings. This thesis addresses respiratory signal extraction from single-lead ECG recordings through two contributions. The first consists of a systematic review of biomedical signal processing techniques for EDR extraction from single-lead ECG signals. Following a structured methodology, the review highlights the wide range of methods proposed in the literature, including amplitude, frequency, morphology, area, and statistical-based techniques. Despite the large number of available approaches, the analysis reveals substantial heterogeneity in experimental protocols, acquisition settings, and validation procedures, complicating direct comparisons among studies. The second contribution presents a novel deep learning approach for the simultaneous reconstruction of ECG and EDR signals from raw single-lead ECG recordings. To this end, a W-Net architecture was developed. The proposed method aims to learn the physiological coupling between cardiac and respiratory activity directly from data, reducing reliance on manually engineered features commonly used in conventional EDR methods. A multi-objective loss function was designed to jointly optimise temporal reconstruction accuracy, spectral consistency, and correlation with reference signals. Experimental results demonstrate effective ECG reconstruction (Pearson correlation up to 0.9), while EDR reconstruction remains more challenging but achieves promising correlations (up to 0.6).