The rise of mobile electrocardiogram (ECG) devices came with the rise of frequent large magnitudes of noise in their recordings. Several artificial intelligence (AI) models have had great success in denoising, but the model's generalizability and enhancement in clinical interpretability are still questionable. We propose Cardio-NAFNet, a novel AI-based approach to ECG denoising by employing a modified version of the Non-Linear Activation Free Network (NAFNET). We conducted three experiments for quantitative and qualitative evaluation of denoising, clinical implications, and generalizability. In the first experiment, Cardio-NAFNet achieved a 53.74dB average signal-to-noise ratio across varying magnitudes of noise in beat-to-beat denoising, which is a significant improvement over the current state-of-the-art model in ECG denoising. In the second experiment, we tested the enhancement in clinical interpretation of the ECG signals by utilizing a pre-trained ECG classifier using 8-second long noise-free ECG signals. When the classifier was tested using noisy ECG signals and their denoised counterparts, Cardio-NAFNet's denoised signals provided a 26% boost in classification results. Lastly, we provide an external validation dataset composed of single-lead mobile ECG signals along with signal quality evaluation from physician experts. Our paper suggests a settling method to capture and reconstruct critical features of ECG signals not only in terms of quantitative evaluation but also through generalizable qualitative evaluation.

UNSTRUCTURED The rise of mobile electrocardiogram (ECG) devices came with the rise of frequent large magnitudes of noise in their recordings. Several artificial intelligence (AI) models have had great success in denoising, but the model’s generalizability and the enhancement in clinical interpretability are still questionable. We propose Cardio-NAFNet, a novel AI-based approach to ECG denoising by employing a modified version of Non-Linear Activation Free Network (NAFNET). We conducted three experiments for quantitative and qualitative evaluation of denoising, clinical implications and generalizability. In the first experiment, Cardio-NAFNet achieved 53.74dB average signal to noise ratio across varying magnitude of noise in beat-to-beat denoising, which is a significant improvement over the current state of the art model in ECG denoising. In the second experiment, we tested the enhancement in clinical interpretation of the ECG signals by utilizing a pretrained ECG classifier using 8 second long noise-free ECG signals. When the classifier was tested using noisy ECG signals and their denoised counterparts, Cardio-NAFNet's denoised signals provided 26% boost in classification results. Lastly, we provide an external validation dataset composed of single-lead mobile ECG signals along with signal quality evaluation from physician experts. Our paper suggests a settling method to capture and reconstruct critical features of ECG signals not only in terms of quantitative evaluation, but also through generalizable qualitative evaluation.