Diagnosis of Subclinical Keratoconus Based on Machine Learning Techniques

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Machine Learning

Background:

• Keratoconus is a progressive corneal disease causing vision loss.
• Early detection is crucial for managing keratoconus progression.
• Subclinical keratoconus (SCKC) presents subtle signs requiring advanced diagnostic methods.

Purpose of the Study:

• To employ the random forest machine learning technique for classifying and predicting subclinical keratoconus.
• To evaluate the efficacy of Pentacam and Corvis metrics in SCKC detection.
• To identify key biomechanical and topographic variables for SCKC identification.

Main Methods:

• Retrospective cross-sectional study of 81 eyes (61 healthy, 20 SCKC).
• Utilized Pentacam topographic and Corvis biomechanical data.
• Applied decision tree and random forest algorithms for classification and variable ranking.

Main Results:

• Random forest model achieved 89% overall accuracy.
• Successfully predicted subclinical keratoconus with 93% specificity.
• Identified stiffness parameter A1 (SP A1) as the most critical variable for SCKC classification.

Conclusions:

• The random forest model is effective for classifying subclinical keratoconus.
• SP A1, derived from corneal biomechanics, is a primary indicator for SCKC.
• A2 time is another significant variable in identifying early keratoconus stages.