ROI-pair importance for autism classification.

Cortex learned to tell autism-like and typical-like connectivity apart. To understand what it learned, we ask Cortex: "For each connection between brain regions, how much did this connection matter in your decision?" The answer is an importance score per ROI-pair. Higher score = Cortex paid more attention to that connection.

Of the 15 top ROI-pairs, 7 show stronger connectivity in autism (hyperconnectivity) and 8 show weaker connectivity (hypoconnectivity).

DMN contributes the most importance across all top ROI-pairs. Default Mode Network — active when you're at rest, self-reflecting, or mind-wandering. Central to self-referential thinking.

This is consistent with a large body of autism neuroimaging literature, which has repeatedly identified atypical connectivity in DMN, social-attention networks (VAN), and frontoparietal control systems.

L_DMN_45 ↔ R_DMN_95 is Cortex's most important single connection (importance = 0.92).

It involves the DMN network. Cortex expects this connection to be stronger in autistic brains than typical — a finding consistent with reported DMN hyperconnectivity patterns.

Treat these as targets for follow-up, not discoveries. Good next steps:

(1) Cross-check literature — do the top edges match published autism connectivity findings? If yes, you've corroborated your dataset and Cortex's learning. (2) Targeted testing — design a focused imaging study on the top 3-5 edges with a larger, stratified cohort. (3) Mechanism — what could cause this specific edge to differ? Tract-tracing, neurotransmitter mapping, genetics can answer that downstream.

Gradient-based importance tells you what Cortex used to classify, not necessarily what causes autism. A feature can be highly predictive because it's an artifact (motion, age, sex confound) rather than a biological signal. Always check your results against known confounds, and prefer edges that appear robustly across multiple cohorts.