Evaluating Logistic Regression and SVM for Enhanced Diabetic Retinopathy Screening

Diabetic retinopathy (DR) remains a leading cause of visual impairment and blindness in working-age individuals globally. Prompt diagnosis and treatment are paramount to circumvent lasting damage; however, the traditional diagnostic approach hinges upon human observation of retinal images, a labor-intensive and subjective method. This paper presents a comparative study of two machine learning algorithms, namely Logistic Regression and Support Vector Machine (SVM), for automatically detecting diabetic retinopathy using retinal image analysis. For data preparation, the APTOS 2019 Blindness Detection dataset is utilized for analysis through a systematic pretreatment pipeline that includes scaling, Gaussian filtering, data augmentation, normalization, and tensor transformation. The logistic Regression model uses softmax for multiclass classification, and is very interpretable, achieving an accuracy of 94.90% precision of 94.22%, and strong recall and F1 scores. On the other hand, the SVM model, which is trained in a one-vs-rest approach with a hinge loss objective, achieves significantly lower performance metrics, thus showing more vulnerability to image noise and computational difficulties for parameter adjustment. Comparative analysis indicates some benefits of Logistic Regression, such as simplicity, decreasing overfitting, and being better suited in terms of model tailoring and smoothing integration in clinical practices.