A deep learning model for diagnosing autism using brain time series.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Early identification of autism spectrum disorder (ASD) is crucial for effective intervention.
• Distinguishing ASD from neurotypical individuals is challenging due to subtle differences.

Purpose of the Study:

• To develop an accurate and robust diagnostic model for autism using neuroimaging data.
• To enhance feature extraction and fusion for improved autism diagnosis.

Main Methods:

• A hybrid deep learning model combining Long Short-Term Memory (LSTM) networks and Attention mechanisms.
• Integration of a residual block with channel Attention for enhanced feature fusion.
• Utilized a sliding window preprocessing method, voting strategy, and 5-fold cross-validation on the ABIDE dataset.

Main Results:

• Achieved 73.1% accuracy on the DOS brain atlas and 81.1% on the HO brain atlas, outperforming baseline models.
• Demonstrated the model's generalizability across data splits through subject-level cross-validation.
• Constructed brain functional connectivity topological structures for ASD patients and healthy individuals.

Conclusions:

• The proposed hybrid LSTM-Attention model with residual and channel Attention blocks shows significant promise for accurate autism diagnosis.
• The findings provide valuable insights into brain functional connectivity in ASD and offer resources for future research.
• This approach facilitates earlier and more effective intervention strategies for individuals with autism spectrum disorder.