Predicting Diabetic Retinopathy And Nephropathy Complications Using Machine Learning Techniques

Abstract

Diabetes and its consequences, especially DR and DN are the major problems to the healthcare systems worldwide, demanding reliable prediction models for early diagnosis and intervention. Imbalanced datasets and complicated feature interactions are poorly handled by traditional methods. The Diabetes clinical data and APTOS 2019 retinal fundus photos were organized and publicly available datasets. Preprocessing included KNN imputer for missing values, outlier detection and handling, MinMax scaling and SMOTE oversampling to balance the data. In terms of classification several ML Algorithms were developed, which included various versions of LR, RF, XGBoost, LightGBM, CatBoost, Multi-Layer Perceptron and hybrid ensembles thereof. Also, advanced models such as StackingClassifier and Ensemble-of-Ensembles were constructed, as well as image classification models: ResNet50, DenseNet121, Xception, NasNetLarge and Xception + DenseNet121. Evaluation metrics used were Accuracy, Precision, Recall, F1-Score, ROC-AUC, RMSE and LogLoss. DenseNet121 was the best model in terms of classification with 99.6% Accuracy, while StackingClassifier Oversampled and LightGBM OverSampled were the best models in terms of nephropathy and retinopathy classification with 99.9% and 99.6% respectively. We used explainable AI approaches like LIME, SHAP and Grad-CAM to interpret the model and a user-friendly prediction interface with Flask to enable accurate, transparent and actionable clinical decision assistance.