Multiple Deep Learning Architectures Achieve Superior Performance Diagnosing Autism Spectrum Disorder Using Features...

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Autism Spectrum Disorder (ASD) diagnosis relies on subjective clinical expertise, which is not universally accessible.
• Automated diagnosis using machine learning (ML) can enhance the availability, reproducibility, and objectivity of ASD diagnoses.

Purpose of the Study:

• To systematically compare the diagnostic performance of various ML classifiers for Autism Spectrum Disorder (ASD).
• To evaluate shallow and deep learning models using neuroimaging features from a large subject database.

Main Methods:

• 12 classifiers were compared, including 6 nonlinear shallow ML, 3 linear shallow ML, and 3 deep learning models.
• Data from over 900 subjects in the IMPAC database, utilizing derived anatomical and functional imaging features.
• Performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC ROC) metric.

Main Results:

• Linear shallow ML models outperformed nonlinear shallow models.
• Deep learning models demonstrated superior performance compared to shallow ML models.
• A dense feedforward network achieved the highest AUC of 0.80, surpassing previous top-performing methods.

Conclusions:

• Deep learning models offer enhanced predictive accuracy for ASD diagnosis, even when using imaging-derived features.
• Automated ML-based diagnosis holds significant potential to augment clinical assessments for Autism Spectrum Disorder.