Comparative study of machine learning and deep learning methods on ASD classification

TL;DR

This study compares machine learning and deep learning methods for classifying autism spectrum disorder using rs-fMRI data, analyzing connectivity patterns to identify the most accurate and interpretable models.

Contribution

It introduces a framework that evaluates multiple classification models on multisite autism data, highlighting the tradeoff between accuracy and interpretability.

Findings

Best model achieved 71% classification accuracy

Analyzed tradeoff between model interpretability and precision

Compared statistical inference and deep learning approaches

Abstract

The autism dataset is studied to identify the differences between autistic and healthy groups. For this, the resting-state Functional Magnetic Resonance Imaging (rs-fMRI) data of the two groups are analyzed, and networks of connections between brain regions were created. Several classification frameworks are developed to distinguish the connectivity patterns between the groups. The best models for statistical inference and precision were compared, and the tradeoff between precision and model interpretability was analyzed. Finally, the classification accuracy measures were reported to justify the performance of our framework. Our best model can classify autistic and healthy patients on the multisite ABIDE I data with 71% accuracy.