Deep learning identification of reward-related neural substrates of preadolescent irritability: A novel 3D CNN application for fMRI

Abstract

The recent emergence of deep learning methods, particularly convolutional neural networks (CNNs), applied to fMRI data presents a promising avenue in psychiatry research, offering advantages over traditional analyses by requiring minimal assumptions and enabling detection of higher-level patterns and intricate, nonlinear relationships within inherently complex fMRI data. Irritability, defined as a lowered threshold for angry responses to blocked rewards, is a promising neurodevelopmental marker for mental health risk due to its robust, transdiagnostic predictive power in youth. In this study, data from the Adolescent Brain and Cognitive Development (ABCD) baseline sample (N = 6065) were utilized for a novel application of a 3D CNN to whole-brain fMRI data acquired during the reward anticipation period of the monetary incentive delay task to predict parent-reported youth irritability severity, measured dimensionally. Regression activation mapping (RAM) was employed to extract feature maps of brain regions most predictive of irritability severity from the model. The model demonstrated satisfactory accuracy, with a mean squared error (MSE) of 1.82, and predicted irritability severity scores with a mean absolute error (MAE) of 0.48 ± 1.54 SD from the true scores. Notably, feature maps revealed bilateral representation of key regions implicated in emotional response and reward processing, including the caudate nucleus, amygdala, parahippocampal gyrus, and hippocampus. This study underscores the potential for 3D CNNs to predict significant, dimensional clinical outcomes such as irritability severity using fMRI data.