Automated feature selection for early keratoconus screening optimization.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Keratoconus is a progressive eye condition requiring early detection for effective management.
• Current diagnostic methods can be enhanced by leveraging advanced computational techniques.
• Integrating optical coherence tomography (OCT) imaging with electronic health records (EHR) offers a rich data source for analysis.

Purpose of the Study:

• To develop and evaluate an automated feature selection (FS) method for optimizing machine learning (ML) models in early keratoconus screening.
• To identify the most relevant ocular parameters for distinguishing keratoconus stages.
• To improve the diagnostic accuracy and efficiency of keratoconus detection.

Main Methods:

• An automated feature selection (FS) method was applied to a dataset of 3162 observations from SS-1000 CASIA OCT and EHR.
• Analysis of Variance (ANOVA) was employed to identify the most relevant features from 448 analyzed parameters.
• The performance of K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Artificial Neural Networks (ANN) classifiers was evaluated.

Main Results:

• The automated FS method, selecting 50 features, significantly improved ML model performance.
• High classification accuracies were achieved: KNN (96.79%), SVM (98.95%), and ANN (95.64%) for distinguishing between 2 and 4 keratoconus classes.
• The method reduced computation time while enhancing diagnostic capabilities.

Conclusions:

• Automated feature selection is effective in optimizing ML models for early keratoconus screening.
• The identified features provide insights into ocular characteristics linked to keratoconus.
• This approach holds potential for advancing early diagnosis, risk prediction, and clinical management of keratoconus.