Explainable AI for dynamic brain imaging: evaluation and cellular-to-network investigation of E/I imbalance in autism

Abstract

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 evaluate explainable AI (xAI) methods to address this challenge, using complementary simulation approaches: recurrent neural networks for controlled parameter exploration, and The Virtual Brain for biophysically realistic modeling with human and mouse connectomes. We demonstrate that xAI methods reliably identify brain features driving performance across challenging conditions, including high noise, low prevalence rates, and subtle alterations in excitatory/inhibitory (E/I) balance. This performance remains consistent across species and anatomical scales. Application to the ABIDE dataset (N=834) reveals that posterior cingulate cortex and precuneus — key nodes of the default mode network — most strongly distinguish autism from controls. The convergence between computational predictions and clinical findings provides support for E/I imbalance theories in autism while demonstrating how xAI can bridge cellular mechanisms with clinical biomarkers, establishing a framework for interpretable deep learning in clinical neuroscience.