Key result
The self-supervised neural network ECG de-noising method improved downstream cardiac disease classification, achieving a Micro F1 score of 0.909 compared to 0.845 without de-noising and 0.885 with the best traditional method.
Why the study?
Previous ECG de-noising approaches did not profoundly consider the ECG signals' specific data structure and were not adaptive to signals recorded by various devices or different skin electrodes, leading to inaccurate cardiac disease analyses.
Does a self-supervised neural network de-noising method improve ECG signal quality and classification performance compared to traditional filters?
Does a self-supervised neural network de-noising method improve ECG signal quality and classification performance compared to traditional filters?
Absolute Event Rate: 0.909% vs 0.845%
A novel self-supervised neural network approach for ECG de-noising improves signal quality and downstream disease classification performance compared to traditional filtering methods.
The electrocardiogram (ECG) records heart-beats and is potentially life-saving. However, the ECG signals (e.g., recorded by the standard ECG monitoring system or the Holter ECG monitoring system) heavily suffered from the noises. Thus, the recorded signals involve meaningful cardiac deflections, other biological waves (e.g., caused by the muscle or electrode movements), and even some noises from the monitoring devices (e.g., the power cables). These noise sources would result in inaccurate analyses of cardiac diseases, thus requiring the ECG de-noising methods in data pre-processing phase before the diagnoses. Previous work provided various ECG de-noising approaches, typically based on some filter algorithms or some wave decomposition algorithms. Most of these approaches did not profoundly consider the ECG signals’ specific data structure, and were not adaptive to the signals recorded by various devices or different skin electrodes. Inspired by the ECG recording theory, we find it available to extract noise information from noisy ECG signals directly. We propose a new ECG de-noising method implemented by the neural network, which de-noise the ECG signals without the supervision of the clean signals. The procedure in the self-supervision is straightforward: we first estimate and simulate the noise signals according to the given noisy ECG signals, and then “subtract” the simulated noises to obtain the de-noised ECG signals by the neural network. Experiments on a public dataset verify that our approach is adaptive to ECG signals from different patients and devices. Also, it is proven that the classification on the ECG signals de-noised by the proposed de-noising methods outperforms those with the traditional de-noising methods.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2024) studied Electrocardiogram (ECG) noise (n=18,604). Self-supervised neural network ECG de-noising vs. No de-noising and traditional de-noising methods (e.g., NeuroKit2, Band-pass filter) was evaluated on Classification performance (Micro F1 score) on raw ECG signals. The self-supervised neural network ECG de-noising method improved downstream cardiac disease classification, achieving a Micro F1 score of 0.909 compared to 0.845 without de-noising and 0.885 with the best traditional method.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: