[Paper Review] CNN-based Approaches For Cross-Subject Classification in Motor Imagery: From The State-of-The-Art to DynamicNet

Alberto Zancanaro, Giulia Cisotto+3 more|arXiv (Cornell University)|Jan 1, 2021

EEG and Brain-Computer Interfaces25 references31 citations

TL;DR

This paper proposes DynamicNet, a Python-based toolbox for rapidly implementing CNN-based deep learning models for EEG signal classification in motor imagery (MI) BCI systems. It demonstrates that an EEGNet model implemented via DynamicNet achieves ~25% higher cross-subject classification accuracy (59.35%) compared to the standard FBCSP method (34.01%) on a 4-class MI dataset, significantly outperforming FBCSP under cross-subject validation schemes.

ABSTRACT

Motor imagery (MI)-based brain-computer interface (BCI) systems are being increasingly employed to provide alternative means of communication and control for people suffering from neuro-motor impairments, with a special effort to bring these systems out of the controlled lab environments. Hence, accurately classifying MI from brain signals, e.g., from electroencephalography (EEG), is essential to obtain reliable BCI systems. However, MI classification is still a challenging task, because the signals are characterized by poor SNR, high intra-subject and cross-subject variability. Deep learning approaches have started to emerge as valid alternatives to standard machine learning techniques, e.g., filter bank common spatial pattern (FBCSP), to extract subject-independent features and to increase the cross-subject classification performance of MI BCI systems. In this paper, we first present a review of the most recent studies using deep learning for MI classification, with particular attention to their cross-subject performance. Second, we propose DynamicNet, a Python-based tool for quick and flexible implementations of deep learning models based on convolutional neural networks. We show-case the potentiality of DynamicNet by implementing EEGNet, a well-established architecture for effective EEG classification. Finally, we compare its performance with FBCSP in a 4-class MI classification over public datasets. To explore its cross-subject classification ability, we applied three different cross-validation schemes. From our results, we demonstrate that DynamicNet-implemented EEGNet outperforms FBCSP by about 25%, with a statistically significant difference when cross-subject validation schemes are applied.

Motivation & Objective

• To survey recent deep learning approaches for cross-subject motor imagery EEG classification.
• To introduce DynamicNet, an open-source Python toolbox for flexible and rapid implementation of deep learning models in EEG-based BCI.
• To evaluate the performance of a DynamicNet-implemented EEGNet model against the standard FBCSP method in 4-class motor imagery classification.
• To investigate the cross-subject generalization capability of deep learning models in real-world BCI applications.

Proposed method

• DynamicNet is a modular, open-source Python toolbox designed for quick implementation of CNN-based deep learning models for EEG data.
• The toolbox was used to implement EEGNet, a lightweight CNN architecture optimized for EEG signal classification.
• Three training strategies were applied: subject-dependent, mixed-up (all subjects' data for training), and cross-subject (training on all but one subject, testing on the held-out subject).
• The study used the BCI Competition IV dataset 2a, with 22-channel EEG recordings from 9 subjects performing 4-class motor imagery tasks.
• Preprocessing included bandpass filtering (0.5–100 Hz) and notch filtering at 50 Hz to remove line noise.
• Classification performance was evaluated using 10-fold cross-validation under each strategy, with accuracy and standard deviation reported.

Experimental results

Research questions

• RQ1 How do recent deep learning models perform in cross-subject motor imagery EEG classification compared to traditional FBCSP?
• RQ2 Can a flexible deep learning toolbox like DynamicNet effectively support rapid prototyping of EEG classification models?
• RQ3 Does a CNN-based model such as EEGNet outperform FBCSP in 4-class cross-subject motor imagery classification?
• RQ4 How does data preprocessing affect the generalization of deep learning models across subjects in MI-BCI systems?

Key findings

• The DynamicNet-implemented EEGNet achieved a mean cross-subject classification accuracy of 59.35% across all subjects, significantly outperforming FBCSP, which achieved only 34.01%.
• Under subject-dependent training, EEGNet achieved 67.88% accuracy, outperforming FBCSP’s 62.56%.
• The mixed-up training strategy increased EEGNet’s accuracy to 72.50%, while FBCSP dropped to 45.85%, indicating that data diversity improves deep learning generalization.
• The study found that preprocessing (filtering and normalization) may reduce subject-independent features, potentially harming cross-subject performance, as raw data yielded higher accuracy than preprocessed data under the same strategy.
• EEGNet’s performance was statistically significantly better than FBCSP in cross-subject scenarios (p < 0.005), confirming its robustness for real-world deployment.
• The results suggest that deep learning models like EEGNet can achieve substantial improvements in cross-subject MI classification, with a ~25% performance gain over FBCSP, especially under challenging cross-subject validation.