Toward Robust EEG Classification Using Adaptive CNN-Transformer and Inception Architectures and Signal Level Augmentation.

Non-invasive electroencephalography (EEG) enables practical brain-state monitoring for applications such as emotion recognition and event-related potential (ERP)-based deception detection. However, robust EEG classification remains challenging because of noise, non-stationarity, limited labeled data, and substantial inter-subject variability. In this work, we present a sensor-density-aware framework that applies different deep architectures to low- and high-channel EEG acquisition settings and augments the training data using a physiologically constrained signal-level procedure. For the 5-channel LieWaves dataset, the CNN-Transformer achieved 97.14±1.36% subject-dependent accuracy with augmentation, compared with 92.91±4.34% without augmentation. For the 62-channel SEED dataset, the Inception CNN achieved 98.52±0.79% with augmentation and 98.44±0.83% without augmentation. The improvement on LieWaves was statistically significant, whereas the small improvement on SEED was not statistically significant. Under subject-independent evaluation, performance was 57.83±8.96% on LieWaves with augmentation and 57.95±8.38% on SEED. These results demonstrate strong subject-dependent performance while confirming that cross-subject generalization remains challenging. Overall, the proposed framework combines sensor-density-aware architecture selection with signal-level augmentation and provides a systematic comparison of subject-dependent and subject-independent EEG classification.