TFMCAD: anomaly detection algorithm based on medical characteristics of ventricular ECG signals in time-frequency domains

Ventricular arrhythmia (VA) is a common type of arrhythmia in clinical practice. Traditional electrocardiogram (ECG) detection methods rely on labeled data and ignore the medical characteristics of the ECG in the time and frequency domains. Therefore, this paper proposes an autoencoder optimization method that integrates time-frequency domain medical features to solve the ECG signal reconstruction problem of VA. In the signal preprocessing stage, the Q, R, and P wave features are masked and input into the backbone network together with the original signal for feature extraction. In order to improve the consistency between the masked signal features and the original signal features, the mutual information module is introduced to effectively enhance the accuracy of signal reconstruction. In addition, by extracting the high-frequency and low-frequency features of the training data in the frequency domain and calculating the mean square error with the reconstructed signal, the signal reconstruction task is further assisted. Extensive experiments on multiple real ECG datasets validate the effectiveness of our proposed method over the state-of-the-art (SOTA) anomaly detection baselines.