Uncertainty-guided Fourier-based domain generalization for seizure prediction

Although deep neural networks have shown promise in the patient-specific context closely related to training and testing data distributions, they face challenges when applied to real-world applications' dynamic and diverse environments. The inherent individual differences and the dynamic nature of EEG signals can lead to substantial domain shifts, hindering the models' ability to generalize from training data (source domain) to new, unseen patient data (target domain). To tackle this issue, we investigate the domain generalization (DG) setting, which aims to learn a model that can generalize to an arbitrary unknown target domain by learning from multiple source domains. In this paper, we propose a novel Fourier-based data augmentation method called UFA, which approaches DG from an uncertainty perspective. UFA works by nonlinearly perturbing the feature statistics of the frequency spectrum obtained by the Fourier transform, thus increasing the diversity of the data distribution in the source domain. In particular, UFA first decomposes the spatial features of the network modules into phase and amplitude spectrums using the Fourier transform (FT). It then randomly samples the feature statistics of frequency spectrums from a specific distribution and uses inverse normalization to generate new frequency spectrums. Finally, we use the inverse Fourier transform to synthesize new spatial feature variants, simulating potential domain shifts in the unknown target domain. Our UFA can be seamlessly integrated as a plug-and-play module within network architectures, positioned immediately after each submodule. The UFA operates in training and is inactive in testing. Extensive experiments on two publicly available benchmarks show that our method achieves significant performance improvements over existing Fourier-based data augmentation methods.