Feature Selection and Dimension Reduction of Social Autism Data

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Autism Spectrum Disorder (ASD) presents a complex and heterogeneous neuropsychiatric condition.
• Existing diagnostic approaches often rely on comprehensive assessments, such as the Social Responsiveness Scale (SRS).

Purpose of the Study:

• To determine if a reduced subset of SRS questions can accurately distinguish ASD from non-ASD.
• To explore feature redundancy within the SRS for potential simplification of diagnostic and intervention strategies.

Main Methods:

• Item-level question selection was performed on SRS data.
• Filter, wrapper, and embedded feature selection analyses were utilized to identify key questions.
• Dimensionality reduction techniques including PCA, t-SNE, and denoising autoencoders were applied to the SRS data.
• A multilayer perceptron (MLP) classifier was trained and evaluated using selected features and reduced representations.

Main Results:

• Classification using the single top-rated SRS question achieved an AUC over 92% for SRS-derived diagnoses and over 83% for dataset-specific diagnoses.
• Feature selection analyses revealed high redundancy among SRS questions.
• A denoising autoencoder for dimensionality reduction slightly outperformed PCA and t-SNE in MLP classifier performance.

Conclusions:

• A minimal set of SRS questions can effectively differentiate ASD, suggesting potential for streamlined diagnostic processes.
• The high feature redundancy in the SRS highlights opportunities for digital quantification of social behaviors, even with privacy considerations.
• Machine learning models, particularly those employing denoising autoencoders, show promise in analyzing SRS data for ASD identification.