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

BACKGROUND AND OBJECTIVE

Accurate staging of keratoconus (KC) is crucial for timely intervention and improving patient quality of life. Unlike prior studies that relied on traditional base machine learning (ML) models, this paper proposes a more advanced two-stage ensemble learning model, designed to automate KC severity staging and track disease progression with improved performance.

METHOD

A clinical dataset collected from Pentacam corneal tomography devices serves as a comprehensive source of corneal data. Following extensive pre-processing, key Pentacam indices strongly correlated with KC severity staging are identified and clinically validated through a rigorous feature selection process. These selected indices are used to train, validate and optimize a two-stage ensemble learner that combines the strengths of four top-performing base ML models-Random Forest (RF), Gradient Boost (GB), Decision Tree (DT), and Support Vector Machine (SVM)-for KC severity staging. Three of these base learners are stacked to leverage their complementary strengths, with their predictions aggregated into a new feature matrix. This matrix is then passed as input to the fourth model, a meta-classifier, which generates the final KC staging results.

RESULTS

Experimental evaluation of the proposed ensemble learner achieved superior performance compared to previous studies. This approach achieved an overall validation accuracy of 99.41 %, a precision of 99.43 %, and a sensitivity of 99.41 %. The F1 and F2 scores were 99.42 % and 99.41 %, respectively. The classification quality, measured by Matthew's Correlation Coefficient, also attained a value of 0.993. Additionally, the model was evaluated on 100 previously unseen test samples, which were entirely excluded from training and cross-validation. It achieved an accuracy of 99 %, demonstrating exceptional consistency, robustness, and generalizability in distinguishing among the distinct stages of KC severity (0-4).

CONCLUSION

The proposed model, developed in collaboration with clinicians, provides a robust foundation for creating a reliable and practical diagnostic tool to detect KC severity stages, track disease progression over time, and evaluate the effectiveness of specific treatments.