Synthetic electroretinogram signal generation using a conditional generative adversarial network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• The electroretinogram (ERG) measures retinal function and can reveal alterations in neurological conditions.
• ERG waveform analysis holds potential for biomarker discovery.
• Challenges in data acquisition for rare or heterogeneous populations necessitate novel approaches.

Purpose of the Study:

• To investigate the utility of AI-generated synthetic ERG waveforms in augmenting real data.
• To enhance the performance of deep learning models for classifying neurological conditions using ERGs.
• To address data limitations in heterogeneous or rare patient cohorts.

Main Methods:

• A dataset of real ERGs from individuals with Autism Spectrum Disorder (ASD) and controls was augmented with synthetic waveforms generated by a Conditional Generative Adversarial Network.
• Two deep learning models, a Time Series Transformer and a Visual Transformer, were employed to classify groups using real and/or synthetic ERG data.
• Model performance was evaluated using Balanced Accuracy (BA).

Main Results:

• The inclusion of synthetic ERGs improved the Balanced Accuracy (BA) of the Time Series Transformer from 0.756 to 0.879 across all recordings.
• The Time Series Transformer achieved a peak BA of 0.89 when trained with both real and synthetic waveforms from a specific flash strength.
• The study demonstrated enhanced classification performance with the integration of AI-generated synthetic data.

Conclusions:

• The enhanced performance of deep learning models utilizing synthetic ERG waveforms validates the application of AI in improving ERG data classification.
• AI-driven synthetic data generation can effectively mitigate data scarcity issues in neurological research.
• This approach shows promise for advancing biomarker discovery and diagnostic capabilities through ERG analysis.