Related Experiment Video
Updated: Nov 3, 2025
Three Different Protocols of Corneal Collagen Crosslinking in Keratoconus: Conventional, Accelerated and Iontophoresis
Published on: November 12, 2015
Development of a classification system based on corneal biomechanical properties using artificial intelligence
Robert Herber1, Lutz E Pillunat2, Frederik Raiskup2
1Department of Ophthalmology, University Hospital Carl Gustav Carus, Universitätsklinikum Carl Gustav Carus an der Technischen Universität Dresden, Fetscherstraße 74, TU 01307, Dresden, Germany. Robert.Herber@uniklinikum-dresden.de.
Machine learning algorithms accurately predict keratoconus (KC) severity using dynamic Scheimpflug tonometry, distinguishing between healthy and various KC stages without keratometric data. The random forest model demonstrated superior accuracy in this classification task.
Area of Science:
- Ophthalmology
- Biomedical Engineering
- Artificial Intelligence
Background:
- Keratoconus (KC) is a progressive corneal disease affecting visual acuity.
- Accurate staging of KC is crucial for timely intervention and management.
- Dynamic Scheimpflug tonometry (CST) offers objective biomechanical measurements of the cornea.
Purpose of the Study:
- To develop and validate machine learning (ML) algorithms for differentiating corneal biomechanical properties in various stages of keratoconus (KC).
- To assess the ability of ML models to predict KC severity using dynamic Scheimpflug tonometry (CST) data.
- To evaluate the performance of ML models in classifying healthy eyes versus different topographical stages of KC.
Main Methods:
- A monocentric, cross-sectional pilot study included 318 keratoconic and 116 healthy eyes.
- Dynamic corneal response (DCR) and pachymetric parameters from CST were utilized to develop ML algorithms.
- Linear discriminant analysis (LDA) and random forest (RF) models were trained on 70% of data and validated on 30%.
Main Results:
- The RF model achieved high sensitivity/specificity for predicting healthy (91%/94%) and KC stages (mild 80%/90%, moderate 63%/87%, advanced 72%/95%).
- Overall accuracy for KC detection across all subgroups was 93% for both LDA and RF models.
- The RF model demonstrated superior overall accuracy (78%) compared to LDA (71%) in classifying KC severity.
Conclusions:
- The random forest (RF) model effectively predicts healthy eyes and various stages of keratoconus (KC) with high accuracy.
- Dynamic Scheimpflug tonometry (CST) can predict KC severity independently of keratometric data, offering clinical utility.
- ML models, particularly RF, show promise for objective KC staging and management.
