Inferred retinal sensitivity in recessive Stargardt disease using machine learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computational Biology

Background:

• Spatially-resolved retinal function assessment is crucial for evaluating macular diseases.
• Fundus-controlled perimetry (FCP) is a standard but laborious method for measuring retinal function.
• Recessive Stargardt disease impacts retinal function and requires sensitive outcome measures for clinical trials.

Purpose of the Study:

• To evaluate a machine learning (ML) approach for predicting spatially-resolved retinal function ('inferred sensitivity') in Stargardt disease.
• To assess the accuracy of ML-based inferred sensitivity using spectral domain optical coherence tomography (SD-OCT) and patient data.
• To determine if inferred sensitivity can serve as a surrogate functional outcome measure in clinical research.

Main Methods:

• Developed an ML model to predict 'inferred sensitivity' from SD-OCT microstructural imaging and patient data.
• Employed nested cross-validation to assess prediction accuracy, measured by mean absolute error (MAE).
• Investigated feature importance (e.g., IS&OS and RPE thickness) for predicting retinal sensitivity.

Main Results:

• The ML model achieved a prediction accuracy of 4.74 dB MAE using only imaging and patient data.
• Incorporating limited FCP data improved accuracy to 3.89 dB MAE, comparable to test-retest variability (3.51 dB MAE).
• Inner and outer segments (IS&OS) and retinal pigment epithelium (RPE) thickness were key predictors of retinal sensitivity.

Conclusions:

• ML-based inferred sensitivity accurately estimates spatially-resolved retinal function in Stargardt disease.
• This approach offers a potential quasi-functional surrogate marker for efficient clinical trial outcome assessment.
• Inferred sensitivity may facilitate refined investigation of treatment effects and disease progression in Stargardt disease.