LML-GAN: Latent Generative Adversarial Network for Time-Series Heart Rate Signal Prediction

Arrhythmia as an important disease among cardiovascular diseases is a serious threat to human health and life all over the world. The electrocardiogram (ECG) provides a wealth of information for the diagnosis and treatment of cardiovascular disease, but traditional diagnostic methods are time-consuming, labor-intensive, and difficult to provide early clinical warning. Moreover, due to the complex information contained in the ECG signals and the strong nonlinearity, the current timing prediction models for ECG signals are often highly oscillatory and cannot capture the waveform features well. To address these challenges, this paper introduces LML-GAN (Latent MIX-LSTM Generative Adversarial Network), an ECG signal prediction model based on potential low-dimensional spatial generative adversarial networks. The model decomposes high-dimensional complex ECG signals into low-dimensional simple representations for adversarial training, and finally maps back to the high-dimensional space to complete the prediction, and adopts the mixed time-frequency domain stepwise supervised loss to improve the model's feature extraction ability for time-series dynamic data. A new residual model MIX-LSTM with larger storage capacity and more flexible storage is proposed for feature extraction based on traditional LSTM. In this paper, experiments are conducted using the MIT-BIH dataset, and the comparison with the same type of model shows that the proposed method shows significant superiority among the existing methods in the current related fields, and its application in the field of ECG signal prediction eliminates the serious problem of oscillatory nature that exists in the same type of model.