Autism Classification Using Topological Features and Deep Learning: A Cautionary Tale.

Area of Science:

• Neuroscience
• Data Science
• Medical Imaging

Background:

• Autism spectrum disorder (ASD) diagnosis can be enhanced by objective methods using resting state functional connectivity (rs-fC) networks.
• Current deep learning models achieve 70.2% classification accuracy on the ABIDE dataset for ASD identification.

Purpose of the Study:

• To investigate the effectiveness of topological features derived from rs-fC networks for classifying ASD.
• To evaluate different topological representations (persistence diagrams, images, landscapes) and machine learning models for ASD classification.

Main Methods:

• Utilized persistence diagrams, persistence images, and persistence landscapes to represent topological features from fMRI data.
• Employed neural networks, support vector machines, and random forests for classification.
• Developed a hybrid model combining topological features with functional correlations.

Main Results:

• The hybrid approach, augmenting topological features with functional correlations, achieved 69.2% accuracy with a simple neural network on the ABIDE dataset.
• Topological features alone did not consistently yield statistically significant improvements over models using only functional correlations.
• Different topological representations showed varying performance across classification models.

Conclusions:

• Topological features offer a complementary but not always statistically significant source of information for ASD classification from fMRI data.
• Practitioners should be cautious about the limited discriminative power of solely relying on topological features for ASD diagnosis.
• Hybrid approaches combining topological and functional correlation features show potential but require further validation.