Development Of A Machine Learning Algorithm For The Diagnosis Of Keratoconus Based On Raw Data From Dual Scheimpflug Corneal Tomography

To develop and test a machine learning algorithm for the diagnosis of Keratoconus based on raw data from Dual Scheimpflug Corneal Tomography (Galilei, Ziemer Ophthalmic Systems AG, Port, Switzerland)

Cross-sectional multicentre study, conducted at Complexo Hospital de Clínicas da Universidade Federal do Paraná, Curitiba, Brazil, and Ocularis Oftalmologia Avançada, Curitiba, Brazil.

The dataset of this two-centre study included for the training group 676 eyes that underwent corneal Dual Scheimpflug Tomography (284 eyes with different stages of keratoconus and 392 eyes in the control group). In order to balance the dataset and to improve minority class samples Synthetic Minority Oversampling Technique (SMOTE) was applied, which increased the training sample size to 769 examinations. For the testing group 60 eyes were further included randomly (30 controls and 30 keratoconic eyes). Examinations were extracted raw, each one consisting of 15 .csv tables with 4 columns and 18000 rows. The model was trained using an eXtreme Gradient Boosting algorithm (XGBoost) and a threshold of 0.1 was set.

A total of 736 eyes (829 after SMOTE) were gathered for the dataset. The classification model achieved an area under the receiver operating characteristics (AUROC) of 0.967, with a sensitivity of 93.3% and a specificity of 100%. The mean prediction time of the model to process each examination was 1.4 milliseconds.

The automated classification model based on raw data from Dual Scheimpflug corneal tomography showed excellent performance on distinguishing different stages of keratoconus from healthy corneas. The use of raw data as an alternative to device generated maps provided time-efficient decisions with minimal computational complexity and hardware requirements, which could facilitate the integration into corneal imaging equipment and further ease the dissemination of the technology.