Abstract
Purpose :
The current standard for diagnosis of keratoconus (KC) is based on corneal topography, clinical signs, pachymetry and best corrected visual acuity. Corneal biomechanical properties are altered by KC but are not widely used as screening. The purpose of this study is to design and develop a supervised Machine Learning (ML) model which is able to detect KC corneas from healthy ones, using biomechanical data from Ocular Response Analyzer (ORA) and Corneal Visualization Scheimpflug Technology (Corvis ST).
Methods :
Corneal biomechanical data from eyes with diagnosis of KC (n = 78) and healthy (n = 76) were used to train and test a ML model. Naïve Bayes (NB) and Logistic Regression (LR) were used as supervised learners. Principal Components Analysis (PCA) was used to derive Principal Components (PC) from ORA and Corvis ST data. The amount of PC was set as the lowest number that explained at least 80% of variance. Using Orange Data Mining, four different models were created, one for each type of data: ORA, Corvis ST, PCA-ORA and PCA-Corvis ST. Each of these models was trained and tested twice: including and excluding Central Corneal Thickness (CCT). Each model was trained and tested using both NB and LR, for a total of 16 sub-models.
Results :
The use of 10-folds cross validation allowed the assessment of models’ performances without the necessity of a separate test-dataset. Performances of each single model were evaluated according to Area Under the Receiving Operative Characteristics (AUC) and Classification Accuracy (CA) values. LR hyperparameters (Regularization and Cost) were tuned to maximise the values of AUC and CA. All the models created showed good values in terms of AUC (between 0.781 and 0.942) and CA (between 0.734 and 0.864). All five sub-models with best performances used clinical data, four used Corvis ST data (one ORA data), three included CCT within the measurement (two did not) and two used LR (three NB). The best model was given by the combination of Corvis ST data and LR, not including the CCT in the training dataset (AUC = 0.942, CA = 0.864).
Conclusions :
The efficacy of the best model can be considered very good in detecting KC from healthy. The model that used data from Corvis ST outperformed models that used data from ORA. CCT, despite being considered important in KC diagnosis, was not the most important predictor of these models; PCA did not improve the performances.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.