AI-based Psychiatric Prediction in Youth: Neuroimaging Provides Minimal Gains Beyond Confounds

Abstract

Recent advances in artificial intelligence (AI) have raised interest in its potential to similarly progress biological psychiatry. This study investigates the current utility of AI models in predicting psychiatric phenotypes in youth - a critical window for psychiatric diagnosis - using neuroimaging data from a large developmental clinical cohort. We assessed the predictive performance of machine learning models on diverse psychiatric and non-psychiatric targets. We show that while models are able to predict various targets from EEG and fMRI data, simple models using only readily-available factors such as demographics and recording site match their performance for clinical phenotypes. This pattern holds across clinical targets and replicates for state-of-the-art deep neural networks, suggesting that either neuroimaging data contains limited disease-specific information or that current methods cannot reliably identify such patterns. These benchmarking results provide important context regarding promises of modern AI in the field of biological psychiatry with a focus on youth.