Highly efficient stacking ensemble learning model for automated keratoconus screening.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Keratoconus (KC) screening heavily relies on traditional machine learning, with stacking ensemble methods being underexplored.
• This research introduces a novel stacking ensemble learning approach for enhanced automated KC screening.

Purpose of the Study:

• To develop and evaluate a stacking ensemble learning model for improved classification of non-KC (NKC), subclinical KC (SCKC), and clinical KC (CKC).
• To enhance the accuracy and efficiency of automated keratoconus detection using corneal imaging data.

Main Methods:

• Utilized a dataset of 2491 corneal cases (NKC, SCKC, CKC) with 79 Pentacam features.
• Implemented pre-processing and feature selection to identify key indices: steepest anterior point keratometry, surface variance, vertical asymmetry, height decentration, and height asymmetry.
• Developed a stacking ensemble model integrating tree-based classifiers (random forest, gradient boosting, decision trees) with an SVM meta-classifier.

Main Results:

• Feature selection reduced parameters by 93.67% and training time by over 85%.
• The model achieved 99.72% accuracy, precision, sensitivity, F1, and F2 scores, with a 0.995 MCC.
• Demonstrated robust performance on unseen data, accurately classifying all NKC and CKC cases with high generalizability.

Conclusions:

• The stacking ensemble approach effectively combines diverse models and Pentacam indices for reliable KC screening.
• Provides clinicians with an automated tool for early keratoconus detection and improved patient management.