Thickness Speed Progression Index: Machine Learning Approach for Keratoconus Detection.

PURPOSE

To develop and validate a pachymetry-based machine learning (ML) index for differentiating keratoconus, keratoconus suspect, and normal corneas.

DESIGN

Development and validation of an ML diagnostic algorithm.

METHODS

This retrospective study included 349 eyes of 349 patients with normal, frank keratoconus (KC), and KC suspect (KCS) corneas. KCS corneas included topographically/tomographically normal (TNF) and borderline fellow eyes (TBF) of patients with asymmetric KC. Six parameters were derived from the corneal thickness progression map on the Galilei Dual Scheimpflug-Placido system and fed into a machine-learning algorithm to create the Thickness Speed Progression Index. The model was trained with 5-fold cross-validation using a random search over 7 different ML algorithms, and the best model and hyperparameters were selected.

RESULTS

A total of 133 normal eyes, 141 KC eyes, and 75 KCS eyes, subdivided into 34 TNF and 41 TBF eyes, were included. In experiment 1 (normal and KC), the best model (Random Forest) achieved an accuracy of 100% and area under the receiver operating characteristic (AUROC) of 1.00 for both normal and KC groups. In experiment 2 (normal, KCS, and KC), the model achieved an overall accuracy of 91%, and AUROC curves of 0.93, 0.83, and 0.99 in detecting normal, KCS, and KC corneas respectively. In experiment 3 (normal, TNF, TBF, and KC), the model achieved an accuracy of 87% with AUROC curves of 0.91, 0.60, 0.77, and 0.94 for normal, TNF, TBF, and KC corneas, respectively.

CONCLUSIONS

Using data solely based on pachymetry, ML algorithms such as the Thickness Speed Progression Index are able to discriminate normal corneas from KC and KCSs corneas with reasonable accuracy.