Novel, high-performance machine learning model for detection of subclinical keratoconus

Abstract

Purpose : Keratoconus (KC) represents one of the leading causes for corneal transplantation. Early detection of KC is important due to its early age of onset and impact on quality of life of the patients. The aim of the study was to develop a simple, automatic machine learning model to detect subclinical KC from control (non-KC) using key parameters obtained from the entire Pentacam parameter set.

Methods : Complete Pentacam output (1692 parameters) and clinical data of 145 subclinical KC and 122 control eyes were collected and analysed. A random forest method was applied to these data to build models using different parameter combinations, and a best performing parameter set was identified. A final model was built using the key Pentacam parameter set, along with other clinical data, and validated using a separate test data.

Results : 3 novel Pentacam parameters were identified, consisting of eccentricity value at an angle of 30 degrees of the front cornea, eccentricity in the 9 mm diameter zone of the cornea and inferior versus superior corneal asymmetry, that were widely available in other commonly used imaging systems to achieve an accuracy of 94%, sensitivity of 97% and specificity of 91% that was comparable with previously reported models built using more parameters. The utility of this combination was further enhanced with the inclusion of clinical parameters (i.e. vision and refraction) to reach an accuracy of 97%, a sensitivity of 99%, and a specificity of 96% in detecting subclinical KC.

Conclusions : The use of the identified key Pentacam parameters along with important clinical parameters, achieved a high level of differentiation between subclinical KC and control eyes. The compact model proposed in this study could be incorporated into the diagnosis of subclinical keratoconus.

This is a 2021 ARVO Annual Meeting abstract.