A Comparative Study of VGG-19 and a Custom CNN for Binary EEG-State Classification Using Composite Time-Frequency Imaging

Accurately distinguishing eyes-open (EO) from eyes-closed (EC) brain states using electroencephalography (EEG) is a foundational task for brain–computer interfaces and clinical neuro-diagnostics. This study provides a head-to-head benchmark of two distinct deep-learning strategies for this task. We convert multichannel EEG recordings into composite time–frequency images to compare a very-deep, pre-trained VGG-19 network against a lightweight, six-layer custom Convolutional Neural Network (CNN). On the large PhysioNet "EEG Motor/Imagery" corpus, VGG-19 achieved superior accuracy (87.88 percent), demonstrating the power of transfer learning. However, a cross-validation experiment on a second, distinct dataset (OpenNeuro ds005420) revealed a performance inversion, with the lightweight CNN (80.12 percent) outperforming VGG-19 (74.96 percent). These findings highlight a critical trade-off: while very-deep models excel on large datasets, lightweight architectures may offer more robust generalization and superior efficiency f or applications with limited data. We conclude by outlining a roadmap for future work, emphasizing generative data augmentation and transformer-based models to bridge this accuracy–efficiency gap.

Ҳали таржима қилинмаган