Predicting Intraocular Pressure From Glaucoma Patients Receiving Medication Treatment Using Explainable Machine Learning

TL;DR

This study uses explainable machine learning to predict intraocular pressure in glaucoma patients on medication, identifying key factors like IGF-1 and LDL.

Contribution

The study introduces an explainable machine learning approach to predict treatment outcomes in glaucoma patients using combined clinical and metabolic data.

Findings

XGBoost outperformed other models with an AUC of 0.708 in predicting intraocular pressure outcomes.

IGF-1, LDL, and lymphocyte count were the top three important features for the model.

LDL and IGF-1 showed strong interactions and stable contributions across model runs.

Abstract

Glaucoma is a chronic neurodegenerative disease of the visual system, and treatment is targeted toward lowering intraocular pressure. However, some patients fail to respond to treatment and their intraocular pressure levels remain high, risking continuous vision loss. Explainable machine learning provides a mechanism for both individual prognostication and the identification of factors associated with treatment outcome. Here, we used explainable machine learning to predict intraocular pressure for glaucoma patients receiving medication treatment. We accessed the UK Biobank to obtain information on 290 eyes from 161 participants who reported a diagnosis of glaucoma and were receiving treatment. Features were divided into three distinct datasets containing demographic data only, physiometabolic parameters and medication prescription data, and all data combined. We evaluated five machine learning techniques for each feature set in terms of their ability to predict intraocular pressure at a follow‐up visit in a classification task. We then calculated SHapley Additive exPlanation (SHAP) values for the best performing model to determine feature importance, stability, and interactions. We found that eXtreme Gradient Boosting (XGBoost) outperformed all other models when trained and tested on the combined feature set with an area under receiver operating characteristic curve (AUC) of 0.708. Insulin‐like growth factor 1 (IGF‐1), low‐density lipoprotein (LDL), and lymphocyte count ranked as the three most important features for this model. LDL and IGF‐1 exhibited a low degree of global variability in contribution to the model output across all cross‐validation repeats. SHAP values demonstrated the strongest interactions being between LDL and IGF‐1. In summary, our studies indicated the importance of blood LDL and IGF‐1 in contributing to the outcomes of intraocular pressure lowering treatment and demonstrated the ability of XGBoost to predict these outcomes.

Figures