Deep Learning-Based Modeling of the Dark Adaptation Curve for Robust Parameter Estimation.

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Dark adaptation (DA) is a critical functional test in age-related macular degeneration (AMD) research.
• Accurate modeling of DA curves is essential for assessing disease progression and treatment efficacy.
• Traditional methods for fitting DA curves may lack robustness and predictive accuracy.

Purpose of the Study:

• To investigate deep learning (DL) sequence modeling for fitting DA curves in AMD patients.
• To enhance the robustness and predictive capabilities of DA curve parameter estimation.
• To compare DL performance against classical nonlinear regression methods.

Main Methods:

• Utilized a long-short-term memory autoencoder as the DL model for DA curve fitting.
• Compared DL model performance with a classical nonlinear regression approach.
• Assessed goodness-of-fit, repeatability, and prediction accuracy of rod intercept time (RIT).

Main Results:

• DL and classical models showed comparable goodness-of-fit (RMSE: DL=0.13±0.06 LU, Classical=0.11±0.04 LU).
• DL method demonstrated superior repeatability, particularly for cone decay parameters.
• DL significantly outperformed the classical method in predicting RIT from early curve data (3.1±3.1 min vs. 19.1±18.6 min error).

Conclusions:

• DL-derived parameters offer enhanced robustness and predictability for DA curves.
• This approach holds potential for improved characterization of AMD severity using DA parameters.
• DL-based DA curve modeling can advance clinical trial endpoints in AMD research.