Autism Spectrum Disorder Studies Using fMRI Data and Machine Learning: A Review

Meijie Liu, Baojuan Li, Dewen HuFrontiers in Neuroscience 2021 · DOI

Abstract

Machine learning methods have been frequently applied in the field of cognitive neuroscience in the last decade. A great deal of attention has been attracted to introduce machine learning methods to study the autism spectrum disorder (ASD) in order to find out its neurophysiological underpinnings. In this paper, we presented a comprehensive review about the previous studies since 2011, which applied machine learning methods to analyze the functional magnetic resonance imaging (fMRI) data of autistic individuals and the typical controls (TCs). The all-round process was covered, including feature construction from raw fMRI data, feature selection methods, machine learning methods, factors for high classification accuracy, and critical conclusions. Applying different machine learning methods and fMRI data acquired from different sites, classification accuracies were obtained ranging from 48.3% up to 97%, and informative brain regions and networks were located. Through thorough analysis, high classification accuracies were found to usually occur in the studies which involved task-based fMRI data, single dataset for some selection principle, effective feature selection methods, or advanced machine learning methods. Advanced deep learning together with the multi-site Autism Brain Imaging Data Exchange (ABIDE) dataset became research trends especially in the recent 4 years. In the future, advanced feature selection and machine learning methods combined with multi-site dataset or easily operated task-based fMRI data may appear to have the potentiality to serve as a promising diagnostic tool for ASD.

Introduction

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by rising prevalence and biological causes that remain largely elusive. Currently, diagnosis relies heavily on subjective behavioral observations. To move toward more objective diagnostic tools, researchers are increasingly turning to functional Magnetic Resonance Imaging (fMRI) to identify "biomarkers"—measurable biological indicators of the disorder.

However, fMRI data presents significant challenges: it is high-dimensional, noisy, and highly variable across different patients and scanning sites. This review synthesizes a decade of research (2011–2021) applying machine learning (ML) to fMRI data for ASD detection. It highlights a shift from traditional classifiers on small, task-based datasets toward deep learning models trained on large, multi-site repositories like the Autism Brain Imaging Data Exchange (ABIDE).

Methodology and Data Processing

The standard pipeline for ASD classification involves several critical stages: data preprocessing, feature construction, feature selection, and pattern classification.

General workflow for ASD classification using fMRI and machine learning.

Feature Construction and Selection

Most studies utilize Functional Connectivity (FC), which calculates the correlation of activity between different brain regions (ROIs) defined by anatomical (e.g., AAL) or functional (e.g., Power) atlases. Advanced techniques include "Dynamic FC," which tracks connectivity changes over time, and "High-Order FC," which examines interactions between multiple networks.

To combat the "curse of dimensionality," feature selection methods such as t-tests, Recursive Feature Elimination (RFE), and LASSO are employed to prune thousands of potential features down to the most discriminative ones. Studies incorporating a dedicated feature selection step typically see an accuracy boost of approximately 9%.

Classification Algorithms

  • Traditional Machine Learning: Support Vector Machines (SVM) remain the most popular choice due to their robustness with smaller datasets. Other common methods include Random Forests and Logistic Regression.
  • Deep Learning (DL): Recent trends favor Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), and Autoencoders (AEs). These models can automatically learn complex features from raw or semi-processed data, often outperforming traditional methods on large-scale datasets.

Datasets

The ABIDE repository is the most significant resource in the field, providing resting-state fMRI (rs-fMRI) data from over 24 independent sites and appearing in roughly 70% of the reviewed literature. While multi-site data is essential for generalization, it introduces significant noise due to varying scanner protocols. In contrast, smaller, site-specific or task-based datasets often yield higher accuracies due to more consistent parameters and the "stressing" of relevant neural circuits.

Key Insights and Results

The review identified several critical factors influencing classification performance and neurophysiological findings.

Informative brain regions identified as critical for ASD classification.

Reported accuracies vary widely, from 48.3% to 97%. An inverse relationship exists between sample size and accuracy: small, highly controlled task-based studies report near-perfect results, while large-scale multi-site studies typically achieve between 60% and 80% accuracy.

Neurophysiological "Hotspots"

A consistent finding across various methodologies is the atypical connectivity within the Default Mode Network (DMN), particularly in the Posterior Cingulate Cortex (PCC) and Precuneus (PCUN). Furthermore, several studies noted a right-hemisphere dominance for discriminative brain regions, suggesting potential functional asymmetry in ASD.

Discussion and Future Directions

Despite significant progress, several hurdles remain before these tools can be used clinically:

  • Standardization: The lack of a "gold standard" pipeline for brain atlas selection and data processing makes cross-study comparisons difficult.
  • Interpretability: Deep learning models often function as "black boxes." Future research must prioritize explainable AI to provide clinicians with understandable rationales for diagnostic flags.
  • Age Sensitivity: ASD manifestations change across the lifespan, necessitating age-specific models for improved precision.

In conclusion, the integration of advanced feature selection and deep learning methods with multi-site datasets or standardized task-based fMRI holds great promise. Future efforts should focus on creating robust, interpretable, and clinically applicable diagnostic tools for ASD.

Summary updated

Abstract

Machine learning methods have been frequently applied in the field of cognitive neuroscience in the last decade. A great deal of attention has been attracted to introduce machine learning methods to study the autism spectrum disorder (ASD) in order to find out its neurophysiological underpinnings. In this paper, we presented a comprehensive review about the previous studies since 2011, which applied machine learning methods to analyze the functional magnetic resonance imaging (fMRI) data of autistic individuals and the typical controls (TCs). The all-round process was covered, including feature construction from raw fMRI data, feature selection methods, machine learning methods, factors for high classification accuracy, and critical conclusions. Applying different machine learning methods and fMRI data acquired from different sites, classification accuracies were obtained ranging from 48.3% up to 97%, and informative brain regions and networks were located. Through thorough analysis, high classification accuracies were found to usually occur in the studies which involved task-based fMRI data, single dataset for some selection principle, effective feature selection methods, or advanced machine learning methods. Advanced deep learning together with the multi-site Autism Brain Imaging Data Exchange (ABIDE) dataset became research trends especially in the recent 4 years. In the future, advanced feature selection and machine learning methods combined with multi-site dataset or easily operated task-based fMRI data may appear to have the potentiality to serve as a promising diagnostic tool for ASD.
BibTeX
@inproceedings{liu2021autism,
  author = {Meijie Liu and Baojuan Li and Dewen Hu},
  title = {{Autism Spectrum Disorder Studies Using fMRI Data and Machine Learning: A Review}},
  booktitle = {Frontiers in Neuroscience},
  year = {2021},
  doi = {10.3389/fnins.2021.697870},
}
Figure