What Robustness Evaluation Is Needed Toward Practical Usage of EEG Decoding Models?

Electroencephalogram (EEG)-based brain computer interfaces (BCIs) are moving toward practical deployment, robustness to channel variability becomes essential due to unstable multichannel EEG acquisition. In real-world scenarios, electrode detachment can lead to missing channels. This situation can be simulated by channel removal. In addition to missing channels, channel identity mismatch could happen, where the channel order or labels do not correctly match the default electrode layouts. This situation can be simulated by channel permutation. This study investigates the model robustness in both channel removal and channel permutation using four representative EEG decoding models. The results demonstrate that the performance decreases as the proportion of channel removal increases. Even with 50% channel removal, accuracy remains above 0.29, showing that the model can still maintain reasonable performance under substantial channel removal. In contrast, channel permutation results in a sharp performance degradation to near-chance levels, with accuracy falling to approximately 0.262-0.293 across all models. Although both cases impair performance, the models are considerably more sensitive to channel permutation. These findings suggest that practical deployment of EEG-based BCI should assess robustness to both missing channels and channel-order variability, alongside conventional classification performance metrics.