Explainable AI techniques for dynamic functional brain imaging: validation and analysis of E/I imbalance in autism
Deep neural networks are increasingly crucial for analysing dynamic functional brain imaging data, offering unprecedented accuracy in distinguishing brain activity patterns across health and disease. However, they often function as black boxes obscuring the neurobiological features driving classifications between groups. This study systematically investigates explainable AI (xAI) methods to address this challenge, employing two complementary simulation approaches: recurrent neural networks for controlled parameter exploration, and The Virtual Brain for biophysically realistic whole-brain modelling. These simulations generate fMRI datasets with known regional alterations in excitation/inhibition (E/I) balance, mimicking mechanisms implicated in psychiatric and neurological disorders. Our comprehensive validation demonstrates that Integrated Gradients and DeepLift successfully identify ground-truth affected regions across challenging conditions, including high noise (-10dB SNR), low prevalence (1% of regions), and subtle E/I alterations. This performance remains robust across three different attribution methods and baseline choices, establishing the reliability of xAI for functional neuroimaging analysis. Critically, successful cross-species validation using both human (68-region) and mouse (426-region) connectomes demonstrates the approach's ability to detect mechanistic alterations across different scales of brain organization. Application to the multisite ABIDE resting-state fMRI dataset (N=834) reveals that regions within the default mode network, particularly the posterior cingulate cortex and precuneus, most clearly differentiated children with autism from neurotypical controls. The convergence between these empirical findings and our biophysical simulations of E/I imbalance provides computational support for mechanistic theories of E/I imbalance in autism while demonstrating how xAI can bridge cellular-level mechanisms with clinical biomarkers. This work establishes a framework for reliable interpretation of deep neural network models in functional neuroimaging, with implications for understanding brain disorders and developing targeted brain stimulation strategies.