Generating Time Series Data With Real-Valued DC-GAN From Complex Time-Frequency Domain: Application to ECG Synthesis

A large amount of data is needed to increase performance in learning algorithms. In many fields such as health, dataset size is limited due to the difficulty of data collection and ethical restrictions. In these cases, synthetic data is needed to increase the performance of learning algorithms. Deep Convolutional Generative Adversarial Networks (DC-GAN) produce very successful results in obtaining synthetic images. In time series synthesis, additional signal-specific information is often needed. With Time-Frequency Transform, the change in frequency and characteristics of time series over time is obtained. In this study, time frequency domain coefficients of time series are generated with 2D real-valued DC-GAN. By embedding the real and imaginary parts of the complex time-frequency domain coefficients into the RGB channels of the real-valued images, the time-frequency domain characteristics of the time series are generated with the real-valued DC-GAN. By extracting the real and imaginary parts from the generated images and converting them into complex coefficients, time series are obtained by inverse transformation. The proposed method is applied to the synthesis of ECG beats and the results obtained are tested with evaluation metrics. The proposed method is also applied to the synthesis of Atrial Fibrillation (AF) and Ventricular Tachycardia (VT) ECG signals. The results obtained show the effectiveness of the proposed method in the synthesis of time series data.