Generative Adversarial Networks for Imputation of Heart Rate Time Series

Missing data in heart rate (HR) and heart rate variability (HRV) recordings are a common challenge in wearable and ambulatory ECG monitoring, often caused by motion artifacts, sensor displacement, or transmission issues. Traditional interpolation methods can fill short gaps but frequently fail to reproduce realistic beat-to-beat dynamics, limiting the reliability of HRV-based physiological interpretation. This thesis proposes a generative adversarial network (GAN) for reconstructing missing segments in ECG-derived HR signals, using a U-Net generator to capture local temporal structure and maintain physiological consistency. Artificially missing segments were introduced under a Missing Completely at Random (MCAR) mechanism at 5%, 15%, and 30% levels of missingness. Reconstruction quality was assessed using MAE, RMSE, correlation, and HRV feature preservation. Results show progressive degradation in reconstruction performance with increasing missingness, with the GAN achieving MAE of 3.85 bpm, 4.66 bpm, and 5.38 bpm for 5%, 15%, and 30% missingness, respectively. Despite higher error at larger gaps, the model retained reasonable correlation and reconstructed HR patterns sufficiently well to support HRV estimation, demonstrating that GAN-based imputation offers a promising approach for recovering physiologically meaningful HR signals. Beyond reconstruction accuracy, future studies could assess whether signals containing imputed segments preserve performance in downstream clinical tasks, such as arrhythmia screening, respiration estimation, or wearable-based anomaly detection. Additionally, adapting these models for real-time ECG reconstruction in wearable systems may enable on-device correction of signal dropout and further enhance the reliability of consumer and clinical ECG monitoring.