A Hybrid LSTM-Autoencoder Framework for Accurate Prediction of Diabetic Kidney Disease

Abstract

Diabetic Kidney Disease (DKD), a complication of diabetes, leads to the gradual decline of renal function. Screening individuals for DKD is essential, as a timely intervention has been shown to improve patient outcomes. Although early detection of DKD can prevent its progression, systematic screening is not universally feasible, which can result in missed or delayed diagnoses. Deep learning (DL) models have shown promise in the medical field, providing promising results. Many researchers have proposed DL models for DKD classification, but achieving reliable accuracy has still been challenging due to the presence of noise and unwanted features in medical data. To address this issue, the paper proposes a hybrid DL model by integrating Long Short-Term Memory (LSTM) into an Autoencoder (AE) architecture, called LSTM-AE, for effective DKD prediction. The combination of these two models effectively identifies important features, resulting in accurate classification. The data used to analyze the DL model's performance was collected from the UCI repository. The data was affected by various issues, and several preprocessing steps were performed to clean the data. This preprocessing also contributed to achieving effective outcomes. The proposed model was compared with three popular DL models: Convolutional Neural Network (CNN), LSTM, and AE. The LSTM-AE achieved the highest accuracy of 99% in DKD prediction, while the other models produced accuracies ranging from 94% to 97%. The proposed model was also compared with existing models from recent studies, and in all experimental outcomes, the LSTM-AE outperformed the others. The results demonstrate that the proposed model is reliable for DKD prediction and can be deployed in real-time practice.