Domain adaptation for epileptic EEG classification using adversarial learning and Riemannian manifold

The epileptic electroencephalography (EEG) classification technique has been extensively adopted in epilepsy diagnosis and management due to its powerful capacity in distinguishing brain dysfunction. Since EEG patterns vary significantly from patient to patient, conventional studies usually perform the training and testing processes on the same subject. However, the challenging issue regarding domain shift among different individuals remains unsolved, which leads to low popularization of clinical application. To alleviate such problem, a domain adaptation model is proposed in this paper. This method intends to learn the universal feature space between different patients via constructing an adversarial autoencoder. By performing an adversarial training procedure, the aggregated posterior of the embedding space is matched with a Riemannian manifold-based prior that contains cross-domain information. This variational inference process can increase the generalization ability while circumventing the overfitting to source domains. Moreover, a distance measure is developed to align the distributions among various domains using the Riemannian mean. It can make the extracted features present higher domain adaptability by minimizing domain gaps in hidden space. Two common classification tasks, seizure prediction and seizure detection, are utilized for model evaluation. In the tests performed on the CHB-MIT scalp EEG dataset, the proposed model achieves a sensitivity of 82.2%, a false alarm rate (FPR) of 0.13 h−1 for seizure prediction and a sensitivity of 86.4%, an FPR of 0.08 h−1 for seizure detection. Experimental results indicate that this method can effectively reduce the domain disparity compared with state-of-the-art baselines.