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

Area of Science:

• Neuroscience
• Developmental Psychology
• Computational Psychiatry

Background:

• Preadolescent irritability predicts future psychopathology and is linked to altered reward processing.
• The neurobiological mechanisms underlying irritability remain largely unclear.
• Deep learning (DL) shows promise in predicting neurodevelopmental issues but often lacks explainability.

Purpose of the Study:

• To integrate optimized prediction with explainability using DL to characterize neural mechanisms of irritability.
• To identify functional connectivity (FC) patterns associated with persistent high irritability (PHI) versus persistently low irritability (PLI) in preadolescents.
• To leverage Shapley Additive Explanations (SHAP) for understanding nonlinear brain-behavior relationships in irritability.

Main Methods:

• Task-based functional magnetic resonance imaging (fMRI) data from a large preadolescent sample (N=1934).
• Trained three DL classifiers (ANN, RF, XGBoost) to distinguish PHI from PLI using FC during reward anticipation.
• Assessed FC between amygdala/ventral striatum seeds and cortical/subcortical regions, using SHAP for feature importance.

Main Results:

• Artificial neural network (ANN) achieved the highest predictive accuracy (AUC=0.73).
• SHAP analysis identified specific FC patterns differentiating PHI from PLI.
• Increased contralateral FC and decreased ipsilateral FC (except for amygdala) predicted PHI.

Conclusions:

• Findings highlight the interplay between reward and emotion regulation circuits in persistent irritability.
• Explainable DL can improve irritability prediction and deepen understanding of its neural underpinnings.
• Specific functional brain connectivity alterations are associated with persistent irritability in preadolescence.