Development of Boosting Algorithm Machine Learning Models for Diabetic Retinopathy Classification Using Symlet Wavelets

Abstract

This research aims to 1) develop and evaluate the performance of boosting machine learning models, and 2) classify retinal images with diabetic retinopathy using symlet wavelet transform feature extraction from 10 sub-bands. A total of 60 symlet features were obtained by extracting six statistical measures, namely mean, standard deviation, energy, entropy, skewness, and kurtosis, from each of the 10 sub-bands. The feature set was fed into three boosting machine learning algorithms, specifically XGBoost, LightGBM, and CatBoost, for classification model development and performance comparison. Model performance was assessed through 10-fold cross-validation and confusion matrix analysis to determine accuracy, sensitivity, specificity, and F1-score metrics for optimal model selection. Comparative performance evaluation demonstrated that the LightGBM model achieved the best overall performance across nearly all evaluation metrics, with an accuracy of 94.47%, sensitivity of 95.24%, specificity of 93.69%, and F1-score of 94.56%. These findings suggest that integrating symlet wavelet features with boosting-based machine learning algorithms provides a promising approach for diabetic retinopathy screening and diagnosis.