Establishment of a corneal ulcer prognostic model based on machine learning

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Corneal infection is a leading cause of unilateral blindness globally.
• Keratitis can lead to corneal perforation and severe visual impairment, necessitating early risk assessment.
• Current treatment strategies require timely analysis of corneal ulcer patient prognosis.

Purpose of the Study:

• To develop a fully automated prognostic model system for corneal ulcer patients.
• To predict the risk of corneal perforation and visual impairment.
• To aid in early treatment strategy development for better patient outcomes.

Main Methods:

• A two-part modeling approach using deep learning and machine learning algorithms (XGBoost, LightGBM).
• Part 1: Deep learning for segmenting and classifying five corneal lesions from 4973 slit lamp images.
• Part 2: Integrating clinical data with image analysis for prognostic modeling using 240 patients' data.

Main Results:

• High accuracy rates for lesion segmentation: hypopyon (96.86%), descemetocele (91.64%), corneal ulcer (90.51%), neovascularization (93.97%).
• Corneal scar classification accuracy reached 69.76%.
• XGBoost model showed strong predictive performance: 1-month AUC of 0.81 for perforation and 0.77 for visual impairment; 3-month AUC of 0.97 for perforation. LightGBM achieved 0.98 AUC for 3-month visual impairment.

Conclusions:

• The developed AI system demonstrates high accuracy in predicting corneal ulcer perforation and visual impairment.
• The prognostic model can significantly aid clinicians in formulating timely and effective treatment strategies.
• This automated system offers a promising tool for managing corneal ulcer patients and preventing blindness.