Open-Source Retinal Electrophysiology Analysis Platform: Interpreting ERG Signals with Machine Learning

ERG-Analysis-API is a complete open-source solution that integrates signal processing, machine learning, and clinical decision support into a unified platform, providing end-to-end support for automated analysis of retinal electrophysiology data.

ERG signals have unique characteristics: small amplitude (usually at the microvolt level), narrow frequency range (0.3-300 Hz), and susceptibility to noise interference. Traditional ERG analysis relies on manual interpretation, which is not only time-consuming and labor-intensive but also has subjective differences. With the development of machine learning technology, automated and standardized ERG analysis has become possible.

The ERG-Analysis-API project emerged as the times require; it is a complete open-source solution that integrates signal processing, machine learning, and clinical decision support into a unified platform. Supported by the Apress/Springer-Nature publishing group, this project is accompanied by the textbook "ERG Signal Processing with Python" to be published in 2028, providing reproducible analysis workflows for ophthalmologists and researchers.

The project implements a signal filtering process compliant with the standards of the International Society for Clinical Electrophysiology of Vision (ISCEV):

• Butterworth Bandpass Filter: Precisely extracts the effective frequency band of ERG signals, removing baseline drift and high-frequency noise
• Notch Filter: Eliminates 50/60 Hz power line interference
• Median Filter: Handles transient noise and artifacts

These filters are carefully tuned to maximize the signal-to-noise ratio while preserving the waveform characteristics of ERG signals.

The project provides two complementary time-frequency analysis methods:

1. Short-Time Fourier Transform (STFT) Spectrogram: Displays the energy distribution of ERG signals in both time and frequency domains, helping to identify waveform abnormalities
2. Wavelet Transform: Offers multi-resolution analysis capabilities, especially suitable for capturing transient features in ERG signals

The project builds a comprehensive feature extraction framework covering three dimensions:

• Time-Domain Features: Classic parameters such as amplitude, latency, and peak time of a-waves and b-waves
• Frequency-Domain Features: Power spectral density, main frequency components, etc.
• STFT Statistical Features: Texture features like mean, variance, and entropy of the spectrogram

These features form the input foundation for machine learning models.

The project adopts two complementary machine learning models:

1. Random Forest Baseline Model: As a highly interpretable and computationally efficient benchmark method, it is suitable for rapid screening
2. Vision Transformer: Uses attention mechanisms to capture long-range dependencies in ERG waveforms, performing excellently in complex cases

The interpretability of model predictions is crucial for clinical applications. The project integrates the SHAP (SHapley Additive exPlanations) framework, providing three levels of interpretation:

• Feature-Level Interpretation: Identifies which ERG parameters contribute the most to classification decisions
• Spectrogram-Level Interpretation: Visualizes the model's attention areas in the time-frequency domain
• Natural Language Summary: Converts technical SHAP values into clinical descriptions understandable to doctors

A key highlight of the project is the clinical decision support API with a four-layer report structure:

1. Traffic Light Indication: Intuitively displays risk levels using red, yellow, and green colors
2. Clinical Summary: Summarizes key findings in plain language
3. Specialist Report: Provides detailed technical parameters and professional interpretations
4. Audit Trail: Records the complete analysis process and parameter settings to meet medical compliance requirements

This layered design not only meets the rapid decision-making needs of clinicians but also provides specialists with the ability to conduct in-depth analysis.