NODE-Former: A Hybrid Approach for Robust Long-Term Time Series Forecasting

The accurate prediction of electroencephalogram (EEG) and electrocardiogram (ECG) signals proves crucial for clinical diagnosis. However, the long-term time series prediction still faces challenges such as the strong dependence on training samples and poor extrapolation performance. To address these issues, This paper introduces NODE-Former (Neural Ordinary Differential Equations-Former), a model designed for accurate long-term prediction of EEG and ECG signals, the model uses latent states with probsparse attention to encode time series and neural ordinary differential equations (NODE) to infer correlations, improving the extrapolation performance. A fractional optimizer is introduced to retain more historical information, enhancing memory, stability, and convergence speed. NODE-Former effectively reconstructs time series and demonstrates superior inference on unseen data. In tests on the MIT-BIH Polysomnographic Database, it outperforms the ITransformer by 77% and NODE-RNN by 48%.