Synthesising Interictal Epileptiform Discharges With Generative Adversarial Network

Interictal epileptiform discharges (IEDs) are reliable biomarkers in electroencephalograms for epilepsy. To automate IED detection, deep learning (DL) has been applied to treat EEG signals as time-series data. However, collecting sufficient IEDs for training DL models is expensive; public and private datasets are often overwhelmed with background activities. To address this issue, generative adversarial networks (GANs) have been used to synthesise IEDs. Previous works using GANs achieved some improvements, but the quality metrics of synthetic samples were not contextualised, resulting in suboptimal model tuning. In this work, we proposed to use density and coverage to evaluate synthetic data and to tune a GAN model. We also utilised soft dynamic time warping (soft-DTW), a differentiable version of DTW, as the reconstruction loss to learn temporal alignments of the IED's waveforms and improve the diversity of synthetic samples. Our proposed approach achieved a 10% improvement in the AUPRC score on the Temple University Events (TUEV) dataset by combining synthetic and real data, demonstrating the potential of this method to enhance the accuracy of IED detection.