ESCRS - PP23.17 - Atomization Of The Keratoconus Diagnosis Processes And Of The Decision-Making Algorithm For Its Treatment

Atomization Of The Keratoconus Diagnosis Processes And Of The Decision-Making Algorithm For Its Treatment

Published 2022 - 40th Congress of the ESCRS

Reference: PP23.17 | Type: Free paper | DOI: 10.82333/cdjk-1h11

Authors: Lyubov Axenova* 1 , Sergey Sakhnov 2 , Kirill Axenov 3 , Victoria Myasnikova 1

1scientific department,S. Fyodorov EYE MICROSURGERY Federal State Institution,Krasnodar,Russian Federation, 2Director,S. Fyodorov EYE MICROSURGERY Federal State Institution,Krasnodar,Russian Federation, 3secure communication systems,St. Petersburg State University of Telecommunications,Saint-Petersburg,Russian Federation

Purpose

To develop a keratoconus diagnostic algorithm with machine learning methods and Pentacam measurements and make an algorithm for treatment tactic selection.

Setting

Retrospective study conducted at the S Fyodorov Eye Microsurgery Federal State Institution Moscow, Russia and its brunches Saint-Petersburg, Cheboksary, Krasnodar.

Methods

In this work we used classical machine learning pipeline. First we prepared database by combination, deleting duplicates, filtering and normalization data, obtained from the Pentacam. Then we labeled data according to the Amsler Krumeih algorithme. To avoid dimension explosion we used RFE and РСА. For data classification QDA method has been chosen. Decision making algorithm for treatment tactic selection was presented in the form of a rule-based algorithm and complied with the recommendations of doctors. Deployment of the algorithm in the form of a web application was carried out with secure access from anywhere using containers and a local server for storing data

Results

AUC values were highest for stage 4 (1.00, class 5) and for normal eyes (0.98, class 0). For eyes with 2 (class 3) and 3 (class 4) stages AUC was 0.97. The accuracy of determining stage 1 (class 2) was 0.96 AUC. The lowest AUC had prekeratoconus (class 1), the staging result for this eye group was 0.95. Accuracy of the decisionmaking algorithm will be measured in clinical practice using web interface.

Conclusions

The Roc analysis of the obtained machine learning algorithm showed high accuracy which indicates the possibility of its application for keratoconus diagnostic.