Graph-Based Analysis of Electroretinograms for Reducing Computational Complexity and Classifying Neurodevelopmental Disorders

Luis Roberto Mercado Diaz, Javier O. Pinzon-Arenas, Paul A Constable, Hugo F. Posada-Quintero

Confirmation: I have read and agree with the IEEE BSN 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: electroretinogram, graph signal processing, autism spectrum disorder, ADHD, machine learning

TL;DR: This network based ERG analysis transforms retinal signals into networks for better autism/ADHD classification, achieving superior multi-class performance with 83% fewer features than traditional methods.

Abstract: Electroretinogram (ERG) signals show distinctive patterns in neurodevelopmental disorders including autism spectrum disorder (ASD) and attention deficit/hyperactivity disorder (ADHD). Traditional ERG analysis relies primarily on time-domain features, limiting the capture of complex nonlinear relationships. We propose ERG-Graph, a novel graph signal processing (GSP) approach that transforms ERG signals into graph networks to extract topological features for improved classification. Using a dataset of 5,838 ERG recordings from 278 subjects across four groups (Control, ADHD, ASD, ASD+ADHD), we applied quantization and k-nearest neighbor graph construction to create ERG-graphs. We extracted 25 graph-level features including centrality measures, spectral properties, and connectivity metrics. Seven machine learning algorithms were evaluated, including Random Forest, Support Vector Machine, and Gradient Boosting, with hyperparameter optimization performed using 3-fold cross-validation. Leave-one-subject-out cross-validation achieved balanced accuracies of 0.77 for ADHD vs. Control classification and 0.76 for ASD vs. Control classification using Random Forest, outperforming traditional ERG features. The ERG-Graph approach demonstrates superior performance in capturing subtle topological patterns associated with neurodevelopmental conditions, offering a promising advancement in automated ERG-based diagnosis.

Track: 3. Signal processing, machine learning, deep learning, and decision-support algorithms for digital and computational health

NominateReviewer: Luis R. Mercado Diaz, [e-mail]

Submission Number: 66