Electroretinogram Analysis Using a Short-Time Fourier Transform and Machine Learning Techniques.

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Signal Processing

Background:

• Electroretinography (ERG) is a crucial non-invasive technique for evaluating retinal function.
• Accurate classification of ERG waveforms is essential for diagnosing various retinal disorders.
• Optimizing signal processing and machine learning approaches can improve ERG analysis.

Purpose of the Study:

• To enhance electroretinography (ERG) waveform signal classification.
• To investigate the effectiveness of Short-Time Fourier Transform (STFT) spectrogram preprocessing combined with machine learning (ML).
• To compare different window functions and ML algorithms for optimal feature extraction and classification.

Main Methods:

• Utilized Short-Time Fourier Transform (STFT) for spectrogram generation from ERG signals.
• Compared various window functions (e.g., Hamming, Boxcar, Bartlett) and sizes for STFT preprocessing.
• Trained deep learning models (Visual Transformer) and classical ML models (e.g., RF) with both spectrograms and manually extracted features.

Main Results:

• The Visual Transformer architecture combined with the Hamming window function demonstrated superior performance in ERG signal classification.
• For manual feature extraction, the RF algorithm performed best, particularly with Boxcar or Bartlett window functions.
• Optimized STFT preprocessing significantly improved the accuracy of ML-based ERG classification.

Conclusions:

• The study identified optimal STFT preprocessing and ML methodologies for ERG signal classification.
• The Visual Transformer with Hamming window offers a powerful approach for automated ERG analysis.
• Recommended RF algorithm for manual feature extraction scenarios, enhancing diagnostic capabilities in clinical electroretinography.