Synthetic electroretinogram signal generation using a conditional generative adversarial network DOI Creative Commons
Mikhail Kulyabin, Aleksei Zhdanov, Irene Lee

et al.

Documenta Ophthalmologica, Journal Year: 2025, Volume and Issue: unknown

Published: April 16, 2025

Abstract Purpose The electroretinogram (ERG) records the functional response of retina. In some neurological conditions, ERG waveform may be altered and could support biomarker discovery. heterogeneous or rare populations, where either large data sets availability a challenge, synthetic signals with Artificial Intelligence (AI) help to mitigate against these factors classification models. Methods This approach was tested using publicly available dataset real ERGs, n = 560 (ASD) 498 (Control) recorded at 9 different flash strengths from 18 ASD (mean age 12.2 ± 2.7 years) 31 Controls 11.8 3.3 that were augmented waveforms, generated through Conditional Generative Adversarial Network. Two deep learning models used classify groups only combined ERGs. One Time Series Transformer (with waveforms in their original form) second Visual model utilizing images wavelets derived Continuous Wavelet Transform Model performance classifying evaluated Balanced Accuracy (BA) as main outcome measure. Results BA improved 0.756 0.879 when ERGs included across all recordings for training Transformer. also achieved best 0.89 single strength 0.95 log cd s m −2 . Conclusions supports application AI improve group recordings.

Language: Английский

Synthetic electroretinogram signal generation using a conditional generative adversarial network DOI Creative Commons
Mikhail Kulyabin, Aleksei Zhdanov, Irene Lee

et al.

Documenta Ophthalmologica, Journal Year: 2025, Volume and Issue: unknown

Published: April 16, 2025

Abstract Purpose The electroretinogram (ERG) records the functional response of retina. In some neurological conditions, ERG waveform may be altered and could support biomarker discovery. heterogeneous or rare populations, where either large data sets availability a challenge, synthetic signals with Artificial Intelligence (AI) help to mitigate against these factors classification models. Methods This approach was tested using publicly available dataset real ERGs, n = 560 (ASD) 498 (Control) recorded at 9 different flash strengths from 18 ASD (mean age 12.2 ± 2.7 years) 31 Controls 11.8 3.3 that were augmented waveforms, generated through Conditional Generative Adversarial Network. Two deep learning models used classify groups only combined ERGs. One Time Series Transformer (with waveforms in their original form) second Visual model utilizing images wavelets derived Continuous Wavelet Transform Model performance classifying evaluated Balanced Accuracy (BA) as main outcome measure. Results BA improved 0.756 0.879 when ERGs included across all recordings for training Transformer. also achieved best 0.89 single strength 0.95 log cd s m −2 . Conclusions supports application AI improve group recordings.

Language: Английский

Citations

0