Predicting Glaucoma Before Onset Using a Large Language Model Chatbot.

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Informatics
• Glaucoma Research

Background:

• Ocular hypertension (OHT) is a precursor to glaucoma, necessitating accurate prediction methods.
• Early detection of glaucoma conversion is crucial for timely intervention and vision preservation.

Purpose of the Study:

• To evaluate the efficacy of ChatGPT, a large language model (LLM), in predicting the conversion of ocular hypertension to glaucoma.
• To assess the performance of ChatGPT 4.0 and 3.5 in forecasting glaucoma development using data from the Ocular Hypertension Treatment Study (OHTS).

Main Methods:

• A retrospective case-control study utilizing 3008 eyes from 1504 OHTS participants.
• Demographic, clinical, ocular, optic nerve head, and visual field parameters from one year prior to glaucoma onset were extracted.
• Tabular data was converted to textual format for ChatGPT prompting via its API to predict glaucoma conversion.

Main Results:

• ChatGPT 4.0 achieved 75% accuracy, an AUC of 0.67, 56% sensitivity, 78% specificity, and a weighted F1 score of 0.77 in predicting glaucoma conversion.
• ChatGPT 3.5 yielded lower performance with 61% accuracy, an AUC of 0.62, 64% sensitivity, 59% specificity, and a weighted F1 score of 0.63.

Conclusions:

• ChatGPT 4.0 demonstrated reasonable performance in forecasting glaucoma development one year before onset, outperforming ChatGPT 3.5.
• LLMs show significant promise for enhancing glaucoma research and clinical practice.
• Future development of ophthalmology-specific LLMs integrating multimodal data and active learning could improve clinical integration.