BCIwiki (bciwiki.com) — Researchers at Hengyang Normal University and Chongqing University of Education, China, introduced NCEL, a lightweight EEG decoder with about 8,000 parameters. Published in Biomedical Physics & Engineering Express on October 27, 2026, the study evaluated five benchmark datasets and separately reported session-disjoint tests.
The proposed Neural Convolution-Enhanced Lightweight network (NCEL) strengthens spatiotemporal electroencephalography (EEG) representations while retaining a compact structure. A dynamic neural convolution module learns input-dependent temporal filtering strategies to address trial-to-trial variability and session-related distribution shifts. Parallel multiscale temporal convolution branches capture both short-term waveform transients and longer-interval rhythmic dependencies. A lightweight channel-attention gate adaptively recalibrates intermediate representations, aiming to emphasize task-relevant signals and suppress noise-dominated channels.
The researchers evaluated NCEL on five benchmark datasets: BCIC IV-2a, BCIC IV-2b, ERN, SMR, and P300. Mean classification accuracy reached 82.75% on BCIC IV-2b. Compared with EEGNet reference values, mean performance improved by 12.53 percentage points on P300 and 9.54 percentage points on BCIC IV-2b, with approximately 0.008M parameters. These comparisons concern the reported benchmark evaluations and should not be read as cross-session gains.
Additional subject-specific, session-disjoint evaluations found cross-session performance comparable to EEGNet on BCIC IV-2a and a mean improvement of 1.80 percentage points on BCIC IV-2b. The authors describe NCEL as a parameter-efficient framework for resource-constrained EEG decoding, while noting that broader cross-session generalization requires further validation.