Detecting keratoconus severity from corneal data of different populations with machine learning

Abstract

Purpose : To develop a machine learning model to detect keratoconus severity from corneal topography, elevation and pachymetry parameters and to assess the generalizability of the model using an independent dataset from a different population.

Methods : We developed a machine learning model to detect different keratoconus severity levels from corneal parameters and compared it to four other machine learning models. The base model was trained and tested using 5881 cases in Brazil We then evaluated the model using a dataset with 1351 cases. We included only 50 raw corneal parameters and excluded all parameters that were generated by Pentacam to assess keratoconus. We utilized a 5-fold cross validation of the area under the receiver operating characteristic curve (AUC) to evaluate machine learning models.

Results : A total of 1726 eyes from the Brazil dataset were diagnosed as normal and 4155 were abnormal based on the Pentacam KCN index. A total of 400 eyes from Japan dataset were normal and 951 eyes diagnosed as KCN based on the Pentacam KCN index. The Random Forest classifier achieved the best AUC of 0.96 compared to four other classifiers once tested on the eyes from the Brazil dataset (Fig. 1). This model also achieved a high AUC of 0.87 once independently tested on all eyes from the Japan dataset (Fig. 2, left panel). Most of the confusions in the model tested on Japan dataset was between normal eyes and eyes with the mild stage of keratoconus (Fig 2, right panel).

Conclusions : The proposed machine learning algorithm provided a highly specific and sensitive model that can detect normal and four stages of keratoconus, with only raw Pentacam corneal parameters, without the machine generated keratoconus indices. The proposed model was generalizable to other cohorts from different races and has the potential to be used in cornea research and clinical practice to automatically detect keratoconus severity from corneal parameters.

This is a 2021 ARVO Annual Meeting abstract.

 

ROC curves of different machine learning classifiers tested on eyes from the Brazil dataset (for healthy class)

 

Validation of the Random Forest classifier using independent eyes from the Japan dataset. Left) the ROC curves, and Right) the confusion matrix.