Explainable AI techniques for dynamic functional brain imaging: validation and analysis of E/I imbalance in autism

Deep neural networks have revolutionized functional neuroimaging analysis but remain "black boxes," concealing which brain mechanisms and regions drive their predictions--a critical limitation for clinical neuroscience. Here we develop and validate an explainable AI (xAI) framework to test whether feature attribution techniques can reliably recover brain regions affected by excitation/inhibition (E/I) imbalance--a fundamental dysregulation implicated in autism, schizophrenia, and other neuropsychiatric disorders. We employed complementary simulation approaches: recurrent neural networks for controlled parameter exploration, and The Virtual Brain simulator incorporating empirically-derived human and mouse connectomes to model E/I balance alterations with unprecedented biological realism. Through systematic validation, we demonstrate that Integrated Gradients and DeepLIFT methods reliably identify brain regions affected by E/I imbalance across challenging conditions, including high noise, low prevalence, and subtle neurophysiological alterations. This performance remains robust across species and anatomical scales, from 68-region human to 426-region mouse connectomes. Application to the ABIDE autism dataset (N=834) reveal convergence between our biophysically grounded simulations and empirical findings, providing computational support for E/I imbalance mechanisms in autism. This work establishes essential tools and data for interpretation of deep learning models in functional neuroimaging, and enables hypothesis-driven analysis of cellular mechanisms across neuropsychiatric disorders.