MINDGAN: EEG-Based Motor Imagery Neural Decoding via Hybrid Deep Learning Network with GAN-Based Augmentation

Abstract

Motor Imagery (MI)-based Brain-Computer Interface (BCI) systems, relying on Electroencephalography (EEG), are constrained by limited training data and high inter-subject variability. To overcome these critical issues, this paper introduces MINDGAN, a novel hybrid framework that integrates a conditional Deep Convolutional Generative Adversarial Network (cDCGAN) with a specialized hybrid classifier. The cDCGAN generates class-conditioned, high-fidelity synthetic EEG samples, effectively addressing data scarcity and enhancing feature diversity. The classifier employs a hybrid architecture with convolutional layers for spatial features, recurrent layers for temporal dynamics, and attention for discriminative signatures. MINDGAN, assessed using BCI Competition IV Datasets 2A and 2B, attained mean accuracies of 85.38% and 90.04%, with kappa values of 0.81 and 0.80, respectively. It exhibited remarkable consistency, producing the lowest cross-subject standard deviations among the assessed models (6.36% and 7.83%). A systematic ablation study revealed that the generative augmentation contributed the largest marginal performance gain. These results position MINDGAN as a robust and generalizable solution for developing reliable MI-BCI applications.

Related Publications