From perception to imagination: a robust deep learning architecture for real-time EEG decoding and neurophysiological validation

BCI clinical viability is hindered by “BCI illiteracy” (15–30% failure rate) and high latency. We propose CSOANet2026, a lightweight CNN with a FIR/ICA preprocessing pipeline, evaluated on 46 subjects. The 3,028-parameter model: (1) reaches 98.67% ± 2.46% within-subject accuracy and 96.73% ± 6.31% leave-one-subject-out (LOSO) accuracy, recovering decoding for “noisy” subjects who fail under linear baselines; (2) runs in 3.99 ± 0.60 ms (CPU)/0.81 ± 0.16 ms (GPU) per trial with a ~17 KB memory footprint; and (3) uses Grad-CAM and gate-weight analyses to verify reliance on neurophysiologically valid occipital Alpha-band (10–13 Hz) modulation—suppression during perception and synchronization during imagery. CSOANet2026 occupies a Pareto-optimal point of the accuracy–capacity frontier, matching deeper CNN baselines at 1/40 to 1/105 of their parameter count, while meeting real-time constraints for closed-loop neurofeedback.