Predicting 24-hour intraocular pressure peaks and averages with machine learning

Area of Science:

• Ophthalmology
• Medical Informatics
• Artificial Intelligence

Background:

• Accurate 24-hour intraocular pressure (IOP) monitoring is crucial for diagnosing and managing glaucoma.
• Current methods may not fully capture diurnal IOP variations, impacting treatment efficacy.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting 24-hour peak and average IOP.
• To identify optimal time points and ML algorithms for enhanced prediction accuracy.
• To uncover key features influencing IOP predictions for improved clinical insights.

Main Methods:

• Retrospective analysis of electronic medical records (January 2014 - May 2024) including 24-hour IOP data from 517 patients.
• Training and evaluation of predictive models using five ML algorithms.
• Optimization of prediction accuracy by testing combinations of five specific time points (10 AM, 12 PM, 2 PM, 4 PM, 6 PM).
• Feature importance assessment using Shapley Additive Explanations (SHAP).

Main Results:

• The Random Forest Regression (RFR) model demonstrated optimal performance for both peak and average IOP prediction.
• For 24-hour peak IOP, RFR using 10 AM, 12 PM, 2 PM, and 4 PM data achieved R²=0.823.
• For 24-hour average IOP, RFR using 10 AM, 12 PM, 4 PM, and 6 PM data achieved R²=0.918.

Conclusions:

• Developed ML models, particularly RFR, effectively predict 24-hour peak and average IOP.
• Identified specific combinations of time points that significantly improve prediction accuracy.
• Findings offer potential for more effective glaucoma management and personalized treatment strategies.