ESCRS - O8 - PREDICTION OF PERIMETRIC PROGRESSION IN OPEN ANGLE GLAUCOMA PATIENTS BASED ON HIGH-SPEED CORNEAL IMAGING (CORVIS ST) MEASUREMENTS USING ARTIFICIAL INTELLIGENCE (XGBOOST).

PREDICTION OF PERIMETRIC PROGRESSION IN OPEN ANGLE GLAUCOMA PATIENTS BASED ON HIGH-SPEED CORNEAL IMAGING (CORVIS ST) MEASUREMENTS USING ARTIFICIAL INTELLIGENCE (XGBOOST).

Published 2023 - 27th ESCRS Winter Meeting

Reference: O8 | Type: Free Paper Submission | DOI: 10.82333/kw23-wz41

Authors: Marta Isabel Martínez-Sánchez* 1 , Gema Bolívar 2 , Daniel Heinisch 3 , Alberto Castaño 3 , Carlos Alonso 3 , Purificación Escámez 2 , Ingrid Rosado 2 , Miguel A. Teus 2

1Hospital Universitario Infanta Leonor,Madrid,Spain, 2Hospital Universitario Príncipe de Asturias,Alcalá de Henares,Spain, 3Recog Analytics,Madrid,Spain

Purpose

The aim of our research was to create an artificial intelligente (AI) model that may predict the risk of visual field (VF) progression in patients with open-angle glaucoma using Dynamic Corneal Response (DCR) Corvis ST parameters registered before and one, three or six months after initiating therapy with prostaglandin analogues (PGs). Different AI models were tested to achieve optimal predictions as measured by their accuracy and area under the curve (AUC, referring to ROC curve). The predictive variables were analyzed to assess which are relevant for final prediction of VF progression.

 

Setting

Hospital Universitario Príncipe de Asturias Universidad de Alcalá, Alcalá de Henares, Madrid, Spain.

Methods

A dataset of 65 eyes was analyzed. The predicted variable (VF progression) was determined in two ways, using both ethe event analysis and the AGIS criteria, collected over 6 years of follow-up. The AI training pipeline to predict the consolidated target consisted of an initial outlier removal process, feature selection and training using the leave-one-out (LOO) cross validation. Techniques based on decision trees using boosting such as the SGBoost (Xtreme Gradient Boosting) algorithm were used. The objective was to optimize AUC, and the results were compared for different AI algorithms.

Results

The most accurate data predictors were those registered at 1 month follow up, predicting glaucomatous progression with an accuracy of 86.2%. The main variables involved in the prediction were the length of the arc at the moment of maximum concavity (B1-HC dArc length), the time from the onset of the deformation to the maximum concavity (B1-HC Deflection Time), and the length of the corneal deflection at the time of maximum concavity (B1-HC Deflection Length). This was consistent across different algorithms, and we also found that results did not improve including additional variables such as corneal hysteresis or IOP values.

Conclusion

The results show that glaucomatous VF progression can be predicted with relatively high accuracy using Corvis ST measurements registered one month after PGs analogues treatment. It seems that the moment of maximum concavity of the cornea seems to be crucial for estimating the risk of progression. The early identification of potential “fast progressors” could have an evident clinical benefit.