Machine Learning-Based Identification of Risk Factors of Keratoconus Progression Using Raw Corneal Tomography Data

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Data Science

Background:

• Keratoconus progression is challenging to predict using standard clinical methods.
• Early identification of disease progression is crucial for timely intervention, such as corneal cross-linking.

Purpose of the Study:

• To identify early indicators of keratoconus progression using machine learning (ML) on Pentacam data.
• To evaluate the efficacy of ML models in predicting keratoconus deterioration.

Main Methods:

• A retrospective analysis of 11,760 Pentacam tomography tests from keratoconus patients was performed.
• Machine learning models, specifically a boosted decision tree, were trained using cross-validation.
• Data were categorized into stable and unstable (referred for cross-linking) eyes.

Main Results:

• Training ML models on single tomography tests yielded poor predictive accuracy (AUC 0.59).
• Training on serial tomography tests significantly improved predictive ability (AUC 0.75), with notable parameters including age, central keratoconus index, Rs B, and D10 mm pachy.
• The model demonstrated improved precision (0.32), recall (0.67), and F1 score (0.43) when using serial data.

Conclusions:

• Machine learning algorithms trained on multiple Pentacam tests can effectively evaluate keratoconus deterioration.
• Novel parameters identified by the ML model offer insights beyond current clinical practices for assessing progression.
• These findings can support clinicians in making informed decisions regarding cross-linking referrals.