Comparative Analysis of Machine Learning Techniques for Diabetic Retinopathy Detection: A Case Study on Messidor Dataset

Identification and classification of diabetic retinopathy (DR) is essential before vision loss might be avoided. This work uses attributes from the 1200 image Messidor dataset to apply machine learning models for automatic DR detection. The dataset was balanced by Synthetic Minority Oversampling Technique (SMOTE) to fix class imbalance after feature standardizing. Random Forest, Support Vector Machine (SVM), Logistic Regression, and K-Nearest Neighbors (KNN) were trained and fine-tune utilizing grid search using 5-fold cross-valuation in order to maximize hyperparameters. SVM achieved the best performance with an F1-score of 0.88, demonstrating high accuracy, precision, and recall. Logistic Regression closely followed with an F1-score of 0.87, showcasing robust classification capabilities. Random Forest also performed well with an accuracy of 81.78% and a weighted F1-score of 0.82. However, KNN struggled significantly, particularly with misclassifications in Class 0, resulting in a lower accuracy of 68.42% and a weighted F1-score of 0.58. These results highlight the effectiveness of SVM and Logistic Regression for accurate classification tasks. The lowest performance of KNN shows how little training data it has for minority classes and how sensitive it is too high-dimensional environments. Especially SVM and Logistic Regression, this study emphasizes the possibilities of machine learning models for exact and automatic DR detection.