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

Background: At present, the fitting of low-vision aids (LVA) for patients globally necessitates the intervention of highly skilled ophthalmologists and certified rehabilitation specialists. To mitigate this limitation, we employed machine learning algorithms to develop an artificial intelligence (AI)-based model for automated LVA fitting assistance. Patients and Methods: Clinical characteristics and diagnostic data from patients with low vision in southeastern China were collected between October 26, 2015, and October 6, 2021, to establish the training and test datasets. We developed and compared three machine learning models-Random Forest (RF), Deep Neural Network (DNN), and Logistic Regression (LR)-to predict prescriptions for three LVA categories selected based on compliance with the World Health Organization's basic specifications for assistive products: Distant Optical Visual aids (DOV), Near Electronic Visual aids (NEV), and Near Optical Visual aids (NOV). Hyperparameter optimization was conducted through four rounds of internal cross-validation. Following model training, the best-performing model was identified and subsequently validated on external data to assess its predictive accuracy and sensitivity. Results: The dataset comprised a total of 1,241 patients diagnosed with low vision. Our model displayed satisfactory performance in LVA fitting when evaluated on the test set. Comparative analysis revealed the RF model as the optimal choice, achieving area under the curve (AUC) values of 0.93 for DOV, 0.83 for NEV, and 0.91 for NOV. Furthermore, feature importance analysis derived from the RF model weights indicated that patient age, best-corrected visual acuity (BCVA of the left eye), and consultation year were the predominant factors influencing LVA fitting decisions across all three aid categories, while visual disability grade specifically impacted DOV prescriptions. In external validation involving 112 prospective cases, the model demonstrated performance comparable to that of a mid-career ophthalmologist (5 years' experience). Conclusion: This study identified significant associations between clinical characteristics and LVA prescription patterns. Leveraging historical LVA fitting data, we developed a machine learning-based decision support system capable of predicting optimal fittings for the three fundamental LVA categories. The proposed tool demonstrates potential for clinical application by generating data-driven prescription recommendations.