Machine Learning-Enhanced EEG Analysis for Predicting Perceptual Learning Efficacy in Amblyopia

Abstract

Purpose : This study investigates the potential of resting-state electroencephalography (EEG), particularly the dominant alpha peak activity, as a tool for capturing neural visual activities indicative of visual plasticity. While perceptual learning (PL) is known to improve visual function in amblyopia, it does so with varying degrees of individual visual plasticity. Our objective is to explore whether EEG signals can serve as objective indicators and predictive biomarkers of individualized PL outcomes in amblyopia patients.

Methods : This study enrolled 176 children (112 males; mean age = 8.08 years, SD = 3.44) diagnosed with amblyopia, comprising 98 with refractive amblyopia and 78 with visual deprivation amblyopia. Out of these, 140 participants underwent 14 days of PL training using a monocular cut-off spatial frequency (cut-off SF) paradigm, followed by a 3-month follow-up period. Based on their response to the PL, participants were divided into two groups: most responsive (MR) and least responsive (LR). Data collection involved measuring the alpha peak frequency and amplitude in resting-state EEG, as well as visual function metrics from the quick contrast sensitivity function test, both before and after the training. The data was analyzed using an automated EEG analysis method at both sensor and source levels. Correlation analyses were performed to assess the relationship between visual function metrics and EEG signals. Additionally, machine learning (ML)-based algorithms were employed to investigate whether EEG signals could predict the classification into MR and LR groups.

Results : The alpha peak frequency and amplitude exhibited a significant correlation with the Cut-off SF(P<0.05) in the children with amblyopia. The effectiveness of PL training was predictably determined using a combination of EEG features at both the sensor and source levels, yielding a maximum accuracy of 85.71% and an Area Under Curve (AUC) value of 0.92. Furthermore, through both correlation analysis and ML algorithms, significant activity was identified in the parietal lobes. It was also noted that the type of amblyopia did not significantly influence these outcomes.

Conclusions : EEG signals demonstrate promise as objective predictors for PL outcomes in amblyopia, with visual electrophysiological activity in the parietal lobe being particularly relevant to individualized visual plasticity in amblyopic individuals post-PL.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.