Long-Term Time Series Prediction with Extreme Peak and Interval Calibration

Prediction of electrocardiogram (ECG) recordings has great significance for early diagnosis of cardiac diseases. However, the long-term ECG prediction task aims to output future ECG recordings with multiple PQRST complexes. Leveraging deep learning methods in the long-term prediction task is still faced with the challenges of extreme peak points and varying periodic intervals. Thus, we address these issues by proposing an ECG prediction framework with extreme peak and interval calibrations. Our framework is capable of learning time-series features via a triple-channel prediction network and emphasizing amplitude residuals of extreme peaks in our designed training loss. In addition, we further design a calibration network to reduce the deviation of periodic intervals, which is capable of alleviating the problem of unaligned predictions resulting from the inaccurate periodic intervals. The experimental results indicate that our framework has an excellent performance in the long-term ECG prediction task.