Development of a machine learning model to predict low vision aid fitting for visually impaired patients

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Machine Learning

Background:

• Low-vision aid (LVA) fitting currently requires specialized expertise, limiting global access.
• An artificial intelligence (AI)-based model was developed to automate and assist in LVA fitting.

Purpose of the Study:

• To develop and validate machine learning models for predicting LVA prescriptions.
• To compare the performance of Random Forest (RF), Deep Neural Network (DNN), and Logistic Regression (LR) models.
• To identify key clinical factors influencing LVA selection.

Main Methods:

• Collected clinical data from 1,241 low-vision patients in southeastern China (2015-2021).
• Trained and tested RF, DNN, and LR models to predict prescriptions for Distant Optical Visual aids (DOV), Near Electronic Visual aids (NEV), and Near Optical Visual aids (NOV).
• Validated the best model on external data and compared its performance to an experienced ophthalmologist.

Main Results:

• The RF model demonstrated superior performance with AUC values of 0.93 (DOV), 0.83 (NEV), and 0.91 (NOV).
• Patient age, best-corrected visual acuity (BCVA), and consultation year were key predictors for all LVA categories.
• External validation showed the AI model's performance comparable to a mid-career ophthalmologist.

Conclusions:

• Machine learning models can effectively predict LVA prescriptions based on clinical data.
• The developed AI tool offers data-driven recommendations, potentially improving LVA fitting accessibility.
• Significant associations between patient characteristics and LVA prescription patterns were identified.