A Methodological Framework for AI-Based Brain Decoding: From Controlled Benchmarks to Interactive Dual-EEG Scenarios

At the intersection of artificial intelligence (AI), computational neuroscience, and signal processing lies a central challenge: how to design data-driven models that can effectively decode complex neural signals while preserving interpretability and neurophysiological plausibility. This thesis addresses this challenge through a systematic methodological investigation of multivariate electroencephalographic (EEG) decoding. The research develops a structured framework organized around four main contributions. First, it establishes a controlled benchmark for EEG decoding using openly accessible motor execution and motor imagery datasets, enabling rigorous comparison across machine learning and deep learning approaches. Second, it systematically evaluates alternative EEG data representations, including time–frequency and electrode–frequency organizations, demonstrating how the structure of the input signal critically shapes classification performance, robustness, and generalization. Third, the thesis introduces explainability as a core validation layer rather than a post hoc analysis. Gradient-based methods such as Grad-CAM are applied to EEG-derived representations to identify the temporal, spatial and spectral features driving model decisions, providing a principled way to assess the neurophysiological plausibility of learned patterns. Fourth, the proposed framework is extended to a more complex dual- EEG socio-affective paradigm within a hyperscanning context, where the increased variability and interdependence of interacting brains highlight the limits of standard decoding pipelines and reveal the boundary conditions for reliable generalization. Overall, this work advances the methodological foundations of AI-based EEG decoding by integrating performance evaluation, representation design, and interpretability within a unified framework. Beyond controlled laboratory settings, the results provide insights into the challenges of transferring AI models to ecologically valid interactive scenarios, contributing to ongoing efforts in computational neuroscience, digital biomarkers, and human-centered AI.