Predicting Diabetic Retinopathy and Nephropathy Complications Using Machine Learning Techniques

Abstract

Diabetes and its sequelae are major problems to worldwide healthcare systems especially Diabetic Retinopathy (DR) and Diabetic Nephropathy (DN). Thus, predictive models are needed for the early diagnosis and intervention. This is a typical problem of imbalanced data and complexity of features interaction in this work. The organized diabetes clinical datasets and APTOS 2019 retinal fundus images were used as publicly available datasets. The preprocessing involved the use of KNN imputer for missing values, outlier detection and handling, MinMax scaling and SMOTE oversampling to balance the data. The following several machine learning models were used for classification: Logistic Regression, Random Forest, XGBoost, LightGBM, CatBoost, Multi-Layer Perceptron, and hybrid ensemble models. Moreover, more complex models were constructed, such as StackingClassifier, as well as ensemble-of-ensemble and image classification models like ResNet50, DenseNet121, Xception, NasNetLarge and ensemble of Xception + DenseNet121. The assessment metrics used are: Accuracy, Precision, Recall, F1-Score, ROC-AUC, RMSE and LogLoss. DenseNet121 achieved the highest classification accuracy of 99.6% on the nephropathy challenge while StackingClassifier Oversampled and LightGBM Oversampled achieved 99.9% and 99.6% accuracy, respectively, for the nephropathy challenge. To obtain the model interpretability, the explainable AI models LIME, SHAP and Grad-CAM were used and the prediction deployment was provided to the user through an explainable interface based on Flask framework, which allowed to obtain accurate, transparent and actionable clinical decision support.