Beyond Accuracy: Counterfactual Auditing of fMRI Representations

Abstract

Self-supervised learning is increasingly applied to functional neuroimaging, yet existing evaluations rely mainly on downstream prediction accuracy, offering limited insight into whether learned representations preserve biologically meaningful structure. We propose Generative Manifold Auditing (GMA), a framework that probes the intrinsic structural sensitivity of latent representations via ROI-level counterfactual interventions. Across the multi-site ABIDE benchmark, a denoising autoencoder (DAE) consistently exhibits higher sensitivity to biologically grounded perturbations than contrastive (MoCo) and masked autoencoder (MAE) baselines, which show flattened structural sensitivity despite apparent stability. Applying GMA to Autism Spectrum Disorder (ASD), we observe heterogeneous, lobe-dependent sensitivity deviations relative to typically developing controls, with consistent distributional patterns on an independent Korean Brain Network (KBN) cohort. Longitudinal analysis further reveals a dissociation between behavioral improvement and neural sensitivity dynamics: while all three longitudinal subjects show reduced K-CARS scores after intervention, their neural sensitivity trajectories remain region- and subject-specific. GMA reframes representation evaluation from prediction-centric benchmarks to mechanism-aware scientific auditing.