ReHeartNet: Reconstruct Electrocardiogram From Photoplethysmography by Using Dense Connected Deep Learning Model

Goal: To enable comfortable and non-invasive heart rhythm monitoring, this work aims to reconstruct electrocardiogram (ECG) signals from photoplethysmogram (PPG) signals, eliminating the need for multiple electrode attachments, which are often inconvenient and may cause skin irritation. Method: To achieve high-fidelity ECG reconstruction from PPG inputs, we propose ReHeartNet, a novel neural network that formulates the task as a regression problem. To capture the multi-scale temporal and frequency relationships between PPG and ECG signals, the model employs densely connected bidirectional long short-term memory (DC-BiLSTM) blocks. To enhance reconstruction accuracy, hierarchical features from different BiLSTM layers are fused within the network architecture. Results: To validate the proposed method, experiments were conducted on four datasets: MIMIC-III, BIDMC, TBME-RR, and CBIC-Heart. ReHeartNet consistently outperforms baselines based on generative adversarial networks (GAN), recurrent neural networks (RNN), and transformers. Conclusions: To support reliable cardiac monitoring in various populations, ReHeartNet demonstrates strong generalization and robustness in ECG reconstruction for healthy individuals and patients with circulatory diseases and arrhythmias, using only wearable PPG signals.