Two-stage ensemble learning framework for automated classification of keratoconus severity

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Accurate keratoconus (KC) staging is vital for patient care.
• Traditional machine learning (ML) models have limitations in KC staging.
• This study proposes an advanced two-stage ensemble learning model for automated KC staging.

Purpose of the Study:

• To develop and validate a novel two-stage ensemble learning model for automated KC severity staging.
• To improve the accuracy and reliability of KC staging compared to existing methods.
• To provide a tool for tracking KC progression and treatment effectiveness.

Main Methods:

• Utilized a clinical dataset from Pentacam corneal tomography.
• Selected key Pentacam indices strongly correlated with KC severity through rigorous feature selection.
• Developed a two-stage ensemble learner combining Random Forest, Gradient Boost, Decision Tree, and Support Vector Machine models.
• Employed stacking for three base learners and a meta-classifier for final staging.

Main Results:

• The proposed model achieved superior performance with 99.41% validation accuracy, 99.43% precision, and 99.41% sensitivity.
• F1 and F2 scores were 99.42% and 99.41%, respectively, with a Matthew's Correlation Coefficient of 0.993.
• The model demonstrated exceptional consistency and generalizability, achieving 99% accuracy on unseen test data for stages 0-4.

Conclusions:

• The developed model offers a robust foundation for a reliable diagnostic tool for KC severity.
• It can aid in detecting KC stages, monitoring disease progression, and evaluating treatment efficacy.
• Collaboration with clinicians ensures the model's practical applicability in patient care.