ORCID

Abstract

BACKGROUND: Diagnosis of autism spectrum disorder (ASD) significantly improves quality of life, yet current diagnostic practices rely on subjective behavioural assessments. While machine learning models using functional Magnetic Resonance Imaging (fMRI) show promise for objective diagnosis, existing approaches have critical limitations. High-accuracy models often neglect interpretability, increasing clinical distrust, while interpretable approaches frequently suffer from low accuracy and lack neuroscientific validation of identified biomarkers. Furthermore, no studies have systematically benchmarked which interpretability methods are most effective for fMRI data, hindering the translation of Artificial intelligence (AI) findings into clinical practice.

METHODS: In this observational study, we used the Autism Brain Imaging Data Exchange I (ABIDE I) dataset comprising 884 participants (408 with ASD; 476 controls with typical development) aged 7-64 years from five countries (Belgium, Germany, Ireland, the Netherlands, and the USA) across 17 international sites after applying mean framewise displacement (FD) filtering (>0.2 mm). Data were collected at the respective sites prior to the release of the ABIDE I dataset in August 2012. We described an explainable deep learning (DL) pipeline using a Stacked Sparse Autoencoder (SSAE) with a softmax classifier, trained on functional connectivity data. We systematically benchmarked seven interpretability methods using the Remove And Retrain (ROAR) technique and cross-validated our findings across three preprocessing pipelines. Critically, we validated identified biomarkers against independent neuroscientific literature from genetic, neuroanatomical, and functional studies.

FINDINGS: Filtering head movement (FD > 0.2 mm) increased classification accuracy from 91% to 98.2% (F1-score: 0.97), achieving state-of-the-art performance. ROAR analysis revealed gradient-based methods, particularly Integrated Gradients, as most reliable for fMRI interpretation. The model consistently identified visual processing regions (calcarine sulcus, cuneus) as critical for ASD classification. These findings aligned with independent genetic studies and neuroimaging research, confirming our model captured genuine neurobiological ASD markers rather than overfitting to dataset artifacts.

INTERPRETATION: We addressed three key research gaps by developing a highly accurate, interpretable ASD classification model, establishing a benchmarking framework for interpretability methods in the fMRI modality, and validating our identified biomarkers against established neuroscience literature. The consistent importance of visual processing regions suggests a fundamental biomarker potentially present across the ASD spectrum, offering new directions for diagnostic biomarker research, though translation to individual-level clinical applications requires further development.

FUNDING: This work was supported by the Engineering and Physical Sciences Research Council, UK [EP/W524554/1].

Publication Date

2025-09-19

Publication Title

eClinicalMedicine

Volume

88

Acceptance Date

2025-08-06

Deposit Date

2025-09-20

Funding

The Engineering and Physical Sciences Research Council, UK [EP/W524554/1]

Keywords

Deep learning, Functional connectivity, Neuroimaging, Explainable AI, Magnetic resonance imaging, Interpretability, Biomarkers, Sparse autoencoders, Functional MRI (fMRI), Remove and Retrain (ROAR)

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Download

Share

COinS