Keratoconus Diagnostic and Treatment Algorithms Based on Machine-Learning Methods

Abstract

The accurate diagnosis of keratoconus, especially in its early stages of development, allows one to utilise timely and proper treatment strategies for slowing the progression of the disease and provide visual rehabilitation. Various keratometry indices and classifications for quantifying the severity of keratoconus have been developed. Today, many of them involve the use of the latest methods of computer processing and data analysis. The main purpose of this work was to develop a machine-learning-based algorithm to precisely determine the stage of keratoconus, allowing optimal management of patients with this disease. A multicentre retrospective study was carried out to obtain a database of patients with keratoconus and to use machine-learning techniques such as principal component analysis and clustering. The created program allows for us to distinguish between a normal state; preclinical keratoconus; and stages 1, 2, 3 and 4 of the disease, with an accuracy in terms of the AUC of 0.95 to 1.00 based on keratotopographer readings, relative to the adapted Amsler-Krumeich algorithm. The predicted stage and additional diagnostic criteria were then used to create a standardised keratoconus management algorithm. We also developed a web-based interface for the algorithm, providing us the opportunity to use the software in a clinical environment.

Keywords: classification; data visualisation; diagnostics; keratoconus; keratotomography; keratotopography; machine learning; treatment.

Figure 1

Study design.

Figure 2

Distribution of the train data on the 2-D plane after the PCA method and coloured according to the stages of keratoconus. Blue—normal; dark purple—preclinical keratoconus; light purple—stage 1; pink—stage 2; orange—stage 3; yellow—stage 4.

Figure 3

(A). Distribution of the test data on the 2-D plane after the PCA method and coloured according to the adopted AK algorithm stages. Blue—normal; dark purple—preclinical keratoconus; light purple—stage 1; pink—stage 2; orange—stage 3; yellow—stage 4. (B). Distribution of the test data after the QDA method and coloured according to the predicted stages of keratoconus.

Figure 4

Result of ROC analysis and calculation of AUC (area) for prediction of keratoconus stages from test data relative to the adapted AK algorithm.

Figure 5

Graphical interface of the software for determining the stage of keratoconus, as well as the indications for surgical intervention: (A)—field for manual input of parameters to determine stage of keratoconus; (B)—field for manual input of parameters to determine patient management tactics; (C)—graphical representation of model including data distribution after PCA and QDA fit (coloured points) as well as new patient data (red points) after PCA and QDA predictions; (D)—treatment algorithm result.