Truace
Subscribe

AI model told autistic from typically developing people with 90.47% accuracy using brain scans

Study title: Diagnostic classification of ASD improves with positively sparsified DTI inter-regional correlation and graph-theoretical artificial intelligence

Researchers built an autism diagnostic model from diffusion tensor imaging scans of 283 people, 154 with ASD and 129 typically developing. They computed correlations between measures from 50 white matter regions, split them into positive and negative networks, and sparsified these into graphs.

Read the study abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by impairments in social interaction and behavioral functioning. ASD diagnosis primarily relies on behavioral assessments and can vary across individuals. In this study, an ASD diagnostic model was developed using diffusion tensor imaging (DTI) derived inter-regional correlations, graph-theoretical analysis and artificial intelligence. Initially, the DTI data from 283 individuals (154 ASD and 129 typically developing) were pre…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2233· cite & verify
Peer-reviewed study: Neuroscience

In brief

Graph features from each region were fed to several machine learning models, and the best result came from a positively correlated network with a support vector machine. This matters because autism diagnosis currently relies on behavioral assessments that vary between individuals, so an imaging-based approach could add objectivity. The study reports cross-validated accuracy only, and the text does not show how it would work in routine clinical practice.

Main points

  1. The dataset covered 283 individuals: 154 with ASD and 129 typically developing.
  2. DTI parameters were extracted from 50 white matter regions, and nine graph features per region produced 450 features.
  3. Models tested included logistic regression, SVM, random forest, extreme gradient boosting, and a one-dimensional convolutional neural network.
  4. The best model used 75 features from a positively correlated Pearson network thresholded at 0.2 with an SVM.
  5. Top discriminative features included centrality measures of the left and right cingulum, left superior corona radiata, and right inferior cerebellar peduncle.

The gain

An SVM trained on graph features from positively correlated white matter networks classified autism versus typically developing individuals with 90.47% average 5-fold accuracy, suggesting a more objective diagnostic aid.

The rundown

The study compared three correlation methods, Pearson, biweight midcorrelation, and Spearman rank, to build inter-regional networks from DTI-derived parameters. The fully correlated matrices were decomposed into positive and negative matrices and thresholded at various levels.

Nine graph-theoretical features were calculated for every white matter region and passed to five types of AI models. The top configuration achieved 90.47% average 5-fold accuracy using a reduced set of 75 features.

The most discriminative features were eigenvector and pagerank centrality of the left cingulum and eigenvector centrality of the left superior corona radiata, right inferior cerebellar peduncle, and right cingulum. The authors describe the framework as promising for accurate ASD classification.

Sources

  1. Peer-reviewedNeuroscience2026-10-21

The debate