EEGNet With Ensemble Learning to Improve the Cross-Session Classification of SSVEP Based BCI From Ear-EEG

Abstract

Ear-electroencephalography (ear-EEG) using electrodes placed above hairless areas around ears is a convenient and comfortable method for signal recording in practical applications of steady-state visual evoked potential (SSVEP) based brain-computer interface (BCI). However, due to the constraint of electrode distribution behind the ear, the amplitude of SSVEP in ear-EEG signals is relatively low, which hinders the application of ear-EEG in SSVEP-based BCI. This study was aimed to improve the performance of ear-EEG in SSVEP-based BCI through re-implementing a compact convolutional neural network (EEGNet) with ensemble learning. We first evaluated the feasibility of applying widely used EEGNet models with different kernel numbers to decode SSVEP in ear-EEG signals. Then we applied an ensemble learning strategy to combine EEGNet models with different kernel numbers to improve the classification of ear-EEG signals. The ear-EEG data was from an open dataset, which acquired three sessions of SSVEP data induced by three flicker stimuli from eleven subjects. The average accuracy of EEGNet with ensemble learning for ear-EEG signals in cross-session validations at 1 s window length was 81.12% (from session 1 to session 2) and 81.74% (from session 1 to session 3), which significantly outperformed canonical correlation analysis (CCA). In addition, the network visualization indicated that EEGNet extracted features related to stimulation frequencies. The results showed promise for accurate classification of SSVEP in ear-EEG signals using deep learning models with strategies, helping to promote the SSVEP based BCI from laboratory to practical application. INDEX TERMSBrain-computer interface (BCI), convolutional neural network (CNN), earelectroencephalography (ear-EEG), steady-state visual evoked potential (SSVEP).

A Systematic Review of Using Deep Learning Technology in the Steady-State Visually Evoked Potential-Based Brain-Computer Interface Applications: Current Trends and Future Trust Methodology

“…Additionally, [ 20 ] proposes the multiharmonic linkage CNN (MHLCNN) model for SSVEP and SSMVEP signal classifications. Furthermore, studies such as [ 10 , 15 , 26 , 28 , 53 , 56 , 63 , 64 ] utilize CNNs for person identification and improving the performance of SSVEP-based BCI systems, respectively. Researchers have employed various feature extraction methods in conjunction with their deep learning models for extracting SSVEP brain signal features.…”

“…Unlike traditional methods, CNNs can perform automatic feature extraction and classification as an end-to-end decoding process. Several studies have demonstrated the effectiveness of using CNNs to improve the classification of scalp-EEG signals by combining EEGs from different channels and at different times through nonlinear operations on signals [ 26 ]. In the context of SSVEP, CNNs are structured with five layers: an input layer, a convolutional layer, a linear unit layer, a pooling layer, and a fully connected layer [ 14 , 15 ].…”

A CNN-Based Deep Learning Approach for SSVEP Detection Targeting Binaural Ear-EEG

“…For ear-EEG, most took the multi-channel approach using the cEEGed platform (Debener et al, 2015 ). With up to 18 channels reading from two ears, Zhu et al ( 2021 ) achieved 81–84% accuracy with 1s window length, slightly less than >90% accuracy from scalp reading. In the work, a CCA-based SSVEP measuring from ears provided as reference was found to be only around 40–50% accurate.…”

An Analysis of Deep Learning Models in SSVEP-Based BCI: A Survey

“…Zhu applied an ensemble learning strategy to combine multiple EEGNet models with different kernel numbers together to enhance the classification accuracy of ear-EEG signals from 50.61% to 81.12% at a 1 s window length of the EEG signal. Zhu also demonstrated that the classification accuracy of the average ensemble model surpasses the accuracy of a single EEGNet model with different kernel numbers [ 54 ]. These studies show that EEGNet is an effective building block in CNN model design.…”