SHAP-based explainable machine learning analysis of reward-related neural connectivity to predict preadolescent irritability

Preadolescent irritability is a robust transdiagnostic neurodevelopmental predictor of later psychopathology, linked to altered reward processing-a common neurocognitive substrate across psychiatric disorders. However, its neurobiological mechanisms remain unclear. Deep learning (DL) excels in predicting neurodevelopmental vulnerabilities and detecting nonlinear brain-behavior relationships but often lacks explainability. Here, we integrate optimized prediction with explainability to characterize neural mechanisms of irritability using task-based fMRI from a large preadolescent sample (N = 1934; mean age = 9.95 years, 101 Persistently High Irritability [PHI], 1833 Persistently Low Irritability [PLI]). We trained three classifiers-artificial neural network (ANN), random forest (RF), XGBoost-to distinguish PHI from PLI using functional connectivity (FC) during reward anticipation. FC was assessed between four seeds (bilateral amygdala, ventral striatum) and 18 cortical/subcortical regions. Shapley additive explanations (SHAP) identified key connectivity predictors accounting for nonlinear effects. ANN (AUC = 0.73, p < .001) outperformed RF (AUC = 0.63, p = .03), XGBoost (AUC = 0.65, p = .01). SHAP revealed increased contralateral FC (e.g., right amygdala-left middle frontal gyrus) and decreased ipsilateral FC (e.g., left ventral striatum-left insula) generally predicted PHI, except amygdala connectivity, where higher ipsilateral FC predicted PHI. These findings highlight interplay between reward and emotion regulation circuits in persistent irritability, underscoring the potential of explainable DL to improve irritability prediction and enhance understanding of its neural mechanisms.