Empirical Mode Decomposition for Deep Learning-Based Epileptic Seizure Detection in Few-Shot Scenario

The precise and automated detection of epileptic seizures has become a focal point of research due to its potential to alleviate the severe consequences experienced by patients. Recent advancements in deep learning-based detection techniques utilizing electroencephalogram (EEG) signals have yielded state-of-the-art performance. However, these methods typically require a large number of training samples to effectively train the deep neural networks. Consequently, their performance can be compromised in scenarios where only a limited number of samples are available. This paper presents two novel approaches that aim to improve seizure detection accuracy in situations with a scarcity of data by harnessing the power of empirical mode decomposition (EMD) applied to EEG signals. Specifically, in both methods, the EMD of EEG signals and the power spectral density (PSD) of the resulting EMD components are employed as inputs for subsequent neural networks. Multiple convolutional neural networks (CNNs) are purposefully designed to perform seizure detection using these inputs. Experimental results demonstrate that our proposed methods achieve superior detection accuracy compared to traditional deep learning-based detection methods that do not incorporate EMD in few-shot scenarios. In particular, when the number of training samples is reduced to 10%, our method shows an improvement of 23%, 19%, and 26% in accuracy, sensitivity, and specificity, respectively, compared to the original EEG input across different networks.