Deep Learning versus Classical Methods in BCI: A Systematic Evaluation of Motor Imagery Classification

Brain-Computer Interfaces (BCI) based on motor imagery (MI) relyon efficient extraction of oscillatory features and cross-subject generalization.This paper presents a systematic comparative evaluation between the classi-cal Common Spatial Patterns with Linear Discriminant Analysis (CSP-LDA)method and a deep learning approach using Compact Convolutional NeuralNetworks (EEGNet), focusing on the cross-subject generalization problem. Us-ing the BCI Competition IV-2b dataset, we establish a rigorous data leak-age prevention pipeline with training-fold-only normalization and a Leave-One-Subject-Out (LOSO) cross-validation protocol. Our central hypothesisis that, under correctly implemented training conditions, compact CNNs canlearn inter-subject invariant representations that generalize beyond the perfor-mance of classical covariance-based methods. Results support this hypothesis:EEGNet achieved a mean accuracy of 71.10% (κ = 0.4220) versus 68.23%(κ = 0.3646) for CSP-LDA. A multi-seed robustness analysis further confirmsthe consistency of this advantage across different random initializations.