Keratoconus Severity Classification Using Features Selection and Machine Learning Algorithms.

Area of Science:

• Ophthalmology
• Biomedical Engineering
• Data Science

Background:

• Keratoconus is a noninflammatory corneal disease causing progressive vision impairment, often starting in adolescence.
• Early detection is difficult due to the absence of pain, hindering timely treatment.
• Machine and deep learning offer potential for early keratoconus detection and management.

Purpose of the Study:

• To identify the most relevant parameters for keratoconus classification using machine learning.
• To evaluate different feature selection algorithms and classifiers for keratoconus detection.
• To improve early diagnosis accuracy for prompt patient treatment.

Main Methods:

• Analysis of 446 parameters from 3162 observations in the Harvard Dataverse keratoconus dataset.
• Application of 11 different feature selection algorithms, including sequential forward selection (SFS).
• Utilizing classifiers such as random forest (RF) for performance evaluation.

Main Results:

• The sequential forward selection (SFS) method identified a subset of 10 key variables.
• The random forest (RF) classifier achieved 98% accuracy for 2 keratoconus classes and 95% for 4 classes.
• The selected subset of variables using SFS yielded classification accuracy comparable to using all original features.

Conclusions:

• Machine learning, particularly SFS with RF, effectively identifies critical parameters for keratoconus classification.
• This approach enhances early detection accuracy, facilitating timely therapeutic interventions.
• The study demonstrates the utility of data-driven methods in diagnosing complex eye conditions like keratoconus.