ASD Neuro Agentic (ANA): An Agentic Framework for Feature Discovery and Cross Site Generalization in Autism fMRI Classification

An agentic AI framework (ANA) for ASD resting-state fMRI classification achieved AUC 0.814 (95% CI 0.773 to 0.841) in leave-one-center-out testing across 17 ABIDE I sites.

• Why it matters: ASD classifiers from resting-state fMRI often fail to generalize across sites, and reported performance can be inflated by data leakage.
• What they did: ANA uses language-model agents to propose features from literature and researcher input, while fixed tools run calculations, then audits leakage-controlled pipeline search across 71 options in 1,028 participants.
• The result: ANA selected a ComBat harmonized ensemble of tangent space connectivity models from six parcellations and 98 public connectivity derivatives, and predictive connectivity was mainly reduced in ASD.

AI summary of the abstract · may contain errors · not medical advice · read the paper.