Inverse Filtering for ECG Signal Reconstruction-A Deep Learning Approach.

Large-scale screening for cardiac arrhythmia, such as atrial fibrillation, using wearable and hand-held devices gain importance in modern diagnostics. However, due to the electrode characteristics, these devices often incorporate signal conditioning that potentially distorts the typical ECG morphology, significantly aggravating the subsequent diagnosis. To overcome this issue, we propose a deep learning-based inverse filtering approach using a U-Net autoencoder to reconstruct clean single-lead ECG signals. A synthetic dataset was generated using the MIMIC-IV database and an object-oriented model simulating the device's characteristics, producing paired noisy and clean ECGs for training the neural network. Evaluation of synthetic data showed effective reconstruction capabilities, preserving key ECG features with an MSE of 0.052 on the validation set and 0.134 on the test set. Additionally, the performance was evaluated on real-world data recorded from the MyDiagnostick single-lead ECG device and a reference patient monitor, yielding an overall improvement in extracting the most important temporal ECG features such as PQ interval, QRS duration and QTc time.Clinical relevance- The proposed deep learning-based inverse filtering approach enhances the diagnostic utility of single-lead ECG devices by mitigating signal distortions introduced by dry electrodes and filtering artifacts. By improving the extraction of important temporal ECG features, such as PQ interval, QRS duration, and QTc time, this method can help to enhance clinical decision-making, particularly in large-scale screening for atrial fibrillation and other cardiac conditions.