Machine learning model for the detection of autism spectrum disorder using electroretinogram signals

Area of Science:

• Neuroscience
• Ophthalmology
• Artificial Intelligence

Background:

• Autism Spectrum Disorder (ASD) presents diagnostic challenges due to subtle early symptoms.
• Electroretinogram (ERG) signals offer a potential biomarker, but interpretation requires expertise.
• Objective AI-driven tools are needed to support the clinical detection of ASD.

Purpose of the Study:

• To develop and validate an AI-enabled framework, UMAP-ERG, for objective ASD detection using ERG signals.
• To investigate the efficacy of Uniform Manifold Approximation and Projection (UMAP) in transforming ERG data for improved classification.
• To address class imbalance in ASD datasets using SMOTE oversampling techniques.

Main Methods:

• Utilized ERG signals from 106 participants, applying UMAP for feature extraction and pattern transformation.
• Implemented SMOTE for oversampling the minority class to handle data imbalance.
• Integrated UMAP and SMOTE with machine learning classifiers (SVM) within the UMAP-ERG framework.
• Evaluated model performance across various flash strengths, focusing on lower intensity signals.

Main Results:

• The UMAP-ERG framework, particularly with SVM, achieved high diagnostic performance (AUC-ROC 0.98, accuracy 0.96).
• Lower intensity flash strengths yielded superior classification results compared to mid and high strengths.
• UMAP-ERG demonstrated superior accuracy (7-15% improvement) over existing state-of-the-art models.
• The model showed excellent sensitivity (0.93) and specificity (0.98) in differentiating ASD cases.

Conclusions:

• The UMAP-ERG framework offers a robust and accurate AI-driven approach for ASD detection via ERG signals.
• This AI tool can significantly enhance clinical decision support systems for ophthalmologists and neuro-specialists.
• The findings suggest ERG analysis, enhanced by AI, is a promising avenue for early ASD identification.