Prediction of refractive error in adolescents using a multimodal large language model

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Digital Health

Background:

• Refractive errors, such as myopia, are a growing global health concern in adolescents.
• Early detection and management are crucial for preventing vision impairment.
• Existing screening methods may lack scalability and accessibility, especially in remote areas.

Purpose of the Study:

• To develop and evaluate a multimodal large language model (LLM) for predicting refractive error (diopter) in adolescents.
• To integrate fundus images with clinical and demographic data for enhanced prediction accuracy.
• To establish the model's utility as a digital health tool for early screening and personalized management.

Main Methods:

• A dataset of 16,226 adolescent records (ages 2-18) was utilized.
• A vision-language foundation model (Qwen2.5-VL) was fine-tuned using supervised learning with imaging and clinical data.
• Performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), R-squared, and Pearson correlation.

Main Results:

• The multimodal LLM achieved a Mean Absolute Error of 0.647 diopters.
• The model demonstrated strong predictive performance across various refractive error subgroups.
• Multimodal data integration significantly outperformed single-modality approaches, with clinically interpretable predictions.

Conclusions:

• The developed multimodal LLM shows significant potential as a digital health technology for refractive error prediction in adolescents.
• This non-invasive, scalable solution can enhance early myopia screening and personalized eye care.
• The model offers a valuable tool for clinicians in both remote and urban healthcare settings.