Identifying Autism from Resting-State fMRI Using Long Short-Term Memory Networks.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Functional magnetic resonance imaging (fMRI) aids in understanding autism spectrum disorder (ASD) pathophysiology and biomarker development.
• Current ASD biomarker research from resting-state functional connectivity (rsFC) faces challenges with heterogeneous, multi-site data accuracy.
• Previous high-accuracy ASD identification was limited to small, homogeneous datasets.

Purpose of the Study:

• To develop and validate a novel method for classifying individuals with ASD using resting-state fMRI time-series data.
• To investigate the efficacy of Long Short-Term Memory (LSTM) recurrent neural networks for ASD classification.
• To improve classification accuracy on large, heterogeneous, multi-site datasets.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) recurrent neural networks for direct classification from resting-state fMRI time-series.
• Trained and tested LSTM models on the complete Autism Brain Imaging Data Exchange (ABIDE) I dataset, a large, multi-site cohort.
• Employed a cross-validation framework to assess model performance and generalizability.

Main Results:

• Achieved a classification accuracy of 68.5% for ASD individuals versus typical controls.
• Demonstrated a 9% improvement in accuracy compared to prior methods utilizing the entire ABIDE cohort.
• Identified specific functional networks and regions implicated in ASD through interpretation of trained LSTM weights.

Conclusions:

• LSTM networks offer a promising approach for objective ASD biomarker development using resting-state fMRI.
• The proposed method shows improved performance on large, heterogeneous, multi-site datasets, addressing a key limitation in the field.
• Interpretation of LSTM weights provides insights into the neural underpinnings of ASD, highlighting key brain regions and networks.