Explainable fMRI decoding reveals neurobiological signatures of emotional arousal complementary to traditional analyses.

OBJECTIVE

Understanding how emotional arousal is represented in the brain is fundamental for elucidating affective regulation and its alterations in autism spectrum disorder. Classical neuroimaging approaches, such as the General Linear Model (GLM), have provided important insights into emotional processing. However, the extent to which data-driven decoding and traditional model-based analyses provide complementary information about emotional arousal remains unclear.

APPROACH: We introduce an explainable artificial intelligence framework that combines the lightweight functional magnetic resonance imaging (fMRI) decoder (fMRINet) with DeepLIFT to derive data-driven neurobiological signatures of emotional arousal. The framework enables interpretation of spatiotemporal neural contributions during emotional processing. We applied the framework to fMRI data acquired while neurotypically and autistic participants viewed naturalistic emotional videos and compared the resulting signatures with categorical and parametric GLM analyses.

MAIN

RESULTS

The framework identified distinct spatiotemporal neurobiological signatures associated with emotional arousal. Neurotypical participants engaged a distributed perceptual-limbic-prefrontal network under high arousal and frontoparietal executive regions under low arousal. In contrast, autistic participants exhibited a more spatially restricted and posterior-biased pattern under high arousal, with reduced ventromedial prefrontal and limbic contributions, while low-arousal patterns resembled TD, suggesting preserved executive mechanisms during low-arousal processing. The GLM analyses revealed partly convergent and partly distinct findings: categorical GLM detected no significant group differences, whereas parametric GLM identified widespread cortical-subcortical modulation in neurotypical individuals and cerebellar-restricted effects in autistic, consistent with compensatory reliance on motor-timing circuits.

SIGNIFICANCE: The proposed framework provides an interpretable, data-driven perspective on emotional processing and reveals neurobiologically plausible patterns consistent with existing literature. Together, the explainable decoding and GLM results highlight complementary aspects of emotional arousal processing, suggesting that explainable decoding approaches may provide additional insights alongside traditional model-based neuroimaging analyses.