A framework for comparison and interpretation of machine learning classifiers to predict autism on the ABIDE dataset

Q: Why do machine learning results for autism classification vary across studies?

Performance differences are linked to variations in experimental setup, including inclusion criteria, data modalities, and evaluation pipelines, which limit direct comparison across published approaches.

Q: Which machine learning models were trained and compared on the ABIDE dataset?

The study trains and compares graph convolutional networks (GCN), edge-variational graph convolutional networks (EV-GCN), fully connected networks (FCN), an auto-encoder followed by a fully connected network (AE-FCN), and a support vector machine (SVM).

Q: What did the ensemble approach and the feature-stability analysis show?

Ensemble models achieved the highest accuracy (72.2%) with AUC reported as 0.78 using GCN classifiers, while SmoothGrad interpretation indicated the FCN model showed the highest stability in selecting relevant features.

Aug 3, 2026

Upload by Miles

Autism is a neurodevelopmental condition impacting roughly 1% of the population, and machine learning approaches have increasingly been trained to distinguish autistic and typically developing individuals using neuroimaging-derived information. Reported model performance varies because experimental design, inclusion criteria, data modalities, and evaluation pipelines differ across studies. This work trains five widely used classifiers on ABIDE using matched evaluation standards and compares their results across functional connectivity matrices, structural volumetric measures, and phenotypic data, including interpretability and feature stability via SmoothGrad.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages37

Reading Time1 h 33 min

AccessFree to download

Abstract

Introduction

Background on autism and neuroimaging biomarkers

Methods and models

GCN, EV-GCN, FCN, AE-FCN, and SVM

Results and comparison

Accuracy, AUC, and statistical testing

Interpretation and feature stability

SmoothGrad-based stability analysis

Resources

Code availability

King’s Research Portal Document Version Peer reviewed version Link to publication record in King's Research Portal Citation for published version (APA): Dong, Y. , Batalle, D. , & Deprez, M. (in press) . A framework for comparison and interpretation of machine learning classifiers to predict autism on the ABIDE dataset. Human Brain Mapping. Citing this paper Please note that where the full-text provided on King's Research Portal is the Author Accepted Manuscript or Post-Print version this may differ from the final Published version. If citing, it is advised that you check and use the publisher's definitive version for pagination, volume/issue, and date of publication details. And where the final published version is provided on the Research Portal, if citing you are again advised to check the publisher's website for any subsequent corrections. General rights Copyright and moral rights for the publications made accessible in the Research Portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognize and abide by the legal requirements associated with these rights. •Users may download and print one copy of any publication from the Research Portal for the purpose of private study or research. •You may not further distribute the material or use it for any profit-making activity or commercial gain •You may freely distribute the URL identifying the publication in the Research Portal Take down policy If you believe that this document breaches copyright please contact [librarypure@kcl.ac.uk](librarypure@kcl.ac.uk) providing details, and we will remove access to the work immediately and investigate your claim. Download date: 11. Mar. 2025 A framework for comparison and interpretation of machine learning classifiers to predict autism on the ABIDE dataset Yilan Dong 1,2, Dafnis Batalle 1,2, Maria Deprez 1 1 School of Biomedical Engineering & Imaging Sciences, King's College London, London SE1 7EH, United Kingdom 2 Department of Forensic and Neurodevelopmental Science, Institute of Psychiatry, Psychology & Neuroscience, King's College London, London SE5 8AF, United Kingdom Corresponding author: Yilan Dong School of Biomedical Engineering & Imaging Sciences, King's College London, London SE1 7EH, United Kingdom [yilan.dong@kcl.ac.uk](yilan.dong@kcl.ac.uk) Abstract Autism is a neurodevelopmental condition affecting ~1% of the population. Recently, machine learning models have been trained to classify participants with autism using their neuroimaging features, though the performance of these models varies in the literature. Differences in experimental setup hamper the direct comparison of different machine-learning approaches. In this paper, five of the most widely used and best-performing machine learning models in the field were trained to classify participants with autism and typically developing (TD) participants, using functional connectivity matrices, structural volumetric measures and phenotypic information from the Autism Brain Imaging Data Exchange (ABIDE) dataset. Their performance was compared under the same evaluation standard. The models implemented included: graph convolutional networks (GCN), edge-variational graph convolutional networks (EV-GCN), fully connected networks (FCN), auto-encoder followed by a fully connected network (AE-FCN) and support vector machine (SVM) . Our results show that all models performed similarly, achieving a classification accuracy around 70% . Our results suggest that different inclusion criteria, data modalities and evaluation pipelines rather than different machine learning models may explain variations in accuracy in published literature. The highest accuracy in our framework was obtained when using ensemble models (p<0.001), leading to an accuracy of 72.2% and AUC = 0.78 using GCN classifiers. However, an SVM classifier performed with an accuracy of 70.1% andAUC=0.77, just mar

FAQs

Why do machine learning results for autism classification vary across studies?

Performance differences are linked to variations in experimental setup, including inclusion criteria, data modalities, and evaluation pipelines, which limit direct comparison across published approaches.

Which machine learning models were trained and compared on the ABIDE dataset?

The study trains and compares graph convolutional networks (GCN), edge-variational graph convolutional networks (EV-GCN), fully connected networks (FCN), an auto-encoder followed by a fully connected network (AE-FCN), and a support vector machine (SVM).

What did the ensemble approach and the feature-stability analysis show?

Ensemble models achieved the highest accuracy (72.2%) with AUC reported as 0.78 using GCN classifiers, while SmoothGrad interpretation indicated the FCN model showed the highest stability in selecting relevant features.

Aug 3, 2026

Upload by Miles