Reliable Connectome-Based Autism Classification: Calibration and Explanation Stability
How far can brain-connectivity classifiers for autism be trusted? This study looks past accuracy to calibrated probabilities and explanations that stay stable across cohorts and sites.
Beyond accuracy
Scroll to watch it unfold. Each beat is one step of the question the paper asks.
- 01
Brain connectivity
Each scan becomes a connectome: a map of how strongly brain regions are functionally connected.
- 02
Two model families
Linear and tree-based classifiers are trained on the same connectomes and compared.
- 03
Cohort and site validation
Evaluation is controlled by cohort and by site, so a model cannot score well by learning which scanner a scan came from.
- 04
Calibrated probabilities
A predicted 70% should mean 70%: the study checks whether probabilities can be taken at face value.
- 05
Stable explanations
The regions that explain a prediction should stay the same from one cohort to the next.
A schematic of the idea, drawn from the write-up below. It shows no results from the paper.
Overview
Autism classifiers built on brain-connectivity data are often reported by accuracy alone. For anything close to clinical use that is not enough: a model also has to say how sure it is, honestly, and point to the same evidence when the patients or the scanner change.
Approach
The study compares linear and tree-based models on connectome features under controlled cohort and site-level validation, so results reflect generalisation to new sites rather than memorised site effects.
Each model is judged on three things together: how well it classifies, how well its probabilities are calibrated, and how stable its explanations stay across cohorts.
Status
Under review at Computers in Biology and Medicine. The paper and results will be posted here after the decision.