Elichatiti, V. V.; Basari, B.; Arif, M.; Ikhsan, M.

Transformer-based deep learning models have shown great potential for decoding visual EEG signals. However, their internal attention mechanisms are often evaluated primarily on optimization objectives, leaving their alignment with biological brain connectivity an open question. This study empirically evaluates how variations in EEG preprocessing strategies affect these attention representations using the Adaptive Thinking Mapper (ATM) model as a framework. We compared a baseline pipeline (MVNN only) against a comprehensive cleaning pipeline integrating ICA and notch filtering. The models were evaluated through cross-generalization, noise robustness, and spectral-temporal ablation analyses. Furthermore, we investigated the structural correspondence between the model's data-driven attention weights and neurophysiological reference networks (GPDC, PDC, and DTF) using Node Strength Correlation and Representational Similarity Analysis (RSA). The results show that the comprehensive preprocessing successfully suppresses non-neural artifacts, such as frontal noise and electrical interference, while maintaining comparable decoding accuracy and baseline robustness. Alignment analyses revealed that the broad spatial organization of the learned attention patterns remains highly stable across pipelines, capturing key directed connectivity dynamics with subtle, metric-dependent variations in global representational geometry. This work provides an empirical exploration into bridging data-driven attention weights with neurophysiological consistency, offering insights toward more transparent brain-computer interfaces.