NCEL: Convolution-Enhanced Lightweight Neural Architecture for Robust EEG Signal Decoding in Noninvasive Brain-Computer Interfaces

Electroencephalography (EEG) is a cornerstone of brain-computer interfaces (BCIs) and clinical neuro-diagnostics due to its non-invasiveness and high temporal resolution. However, robust EEG decoding remains challenging, particularly for resource-constrained deployments. While lightweight architectures like EEGNet offer favorable parameter efficiency, their reliance on fixed convolution kernels and restricted temporal receptive fields inherently limits their performance in complex, highly variable scenarios. To address these limitations, we propose a Neural Convolution-Enhanced Lightweight network (NCEL) that augments spatiotemporal EEG representations while preserving structural compactness. Specifically, a dynamic neural convolution module (NCM) is designed to learn input-dependent temporal filtering strategies for trial-to-trial variability and session-related distribution shifts.Furthermore, parallel multiscale temporal convolution branches are integrated to concurrently capture short-term local waveform transients and long-interval rhythmic dependencies. Finally, a lightweight channel-attention gating mechanism adaptively recalibrates intermediate representations, amplifying task-relevant neural responses while suppressing noise-dominated channels. Extensive evaluations on five benchmark datasets (BCIC IV-2a, BCIC IV-2b, ERN, SMR, and P300) demonstrate that NCEL achieves a mean classification accuracy of 82.75\% on BCIC IV-2b. Notably, NCEL achieves higher mean performance than the EEGNet reference values, with performance gains of 12.53 percentage points on the P300 dataset and 9.54 percentage points on BCIC IV-2b, while maintaining only $\sim$0.008M parameters. Additional subject-specific session-disjoint evaluations show comparable cross-session performance to EEGNet on BCIC IV-2a and a 1.80-percentage-point mean improvement on BCIC IV-2b. Overall, NCEL provides a parameter-efficient framework for resource-constrained EEG decoding, while broader cross-session generalization requires further validation.