Related Experiment Video
Updated: Dec 11, 2025
05:14
Comparison of Agreement and Accuracy using Binocular Wavefront Optometer with Autorefractor and Phoropter
Published on: September 16, 2025
464
Evaluating the Performance of Various Machine Learning Algorithms to Detect Subclinical Keratoconus
Ke Cao1,2, Karin Verspoor3, Srujana Sahebjada1,2
1Centre for Eye Research Australia, Royal Victorian Eye and Ear Hospital, Melbourne, Victoria, Australia.
Translational Vision Science & Technology
|August 21, 2020
Summary
Machine learning algorithms effectively detect subclinical keratoconus (KC) using routine eye exam data. This approach aids in early diagnosis, potentially preventing the need for corneal transplants.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Biomedical Engineering
Background:
- Keratoconus (KC) is a primary cause of corneal transplantation globally.
- Early detection of subclinical KC is crucial for management and preventing grafts.
- Diagnosing subclinical KC presents significant clinical challenges.
Purpose of the Study:
- To compare eight machine learning algorithms for differentiating subclinical KC from non-KC eyes.
- To identify optimal parameter combinations for accurate subclinical KC detection.
- To build predictive models using routinely collected clinical data.
Main Methods:
- Corneal and clinical parameters were collected from 49 subclinical KC and 39 control eyes using Oculus Pentacam.
- Eight machine learning algorithms were applied and evaluated for their performance.
- Parameter combinations were optimized to enhance model accuracy.
Main Results:
- Random forest, support vector machine, and k-nearest neighbors showed superior performance in detecting subclinical KC.
- The random forest model achieved an area under the curve of 0.97 with five parameters.
- High sensitivity (0.94) and specificity (0.90) were achieved by support vector machine and k-nearest neighbor models, respectively.
Conclusions:
- Machine learning algorithms can successfully identify subclinical KC.
- Minimal, routinely collected clinical parameters are sufficient for accurate detection.
- These models offer objective diagnostic assistance for subclinical KC.
