Early Detection of Chronic Kidney Disease Using Deep Learning

Abstract

Chronic Kidney Disease (CKD) is a significant health issue on the global level and prompt diagnosis is crucial in preventing the occurrence of end-stage renal disease (ESRD). Improved methods of diagnosing CKD such as serum creatinine and glomerular filtration rate (GFR) cannot always detect the disease at early stages. The paper will address deep learning models, which include Convolutional Neural Networks (CNNs), fully connected Neural Networks (FCNNs) and Autoencoders compared to the traditional machine learning models, which involve Decision Trees and Support Vector Machines (SVMs) during the detection of CKD. These were compared using the UCI CKD data set of 400 patients and 24 clinical features. We discover that Deep Learning models are far more predictive accurate than the conventional models. Optimized CNN model reached a score of 91.0, 94.0 and a score of 0.97 on accuracy, recall and AUC respectively, which demonstrates that this model performs better in identifying CKD during the early stages. The Optimized FCNN model achieved the AUC of 0.93, recall of 89.5 meaning that it may have an application in the detection of CKD but it is a little less than CNN model. Relative to that, such old models of decision trees were known to perform at 80.0, recall of 75.5 and AUC of 0.85 and SVM models performed a bit better at 82.5, but still lagged behind deep learning models. These findings have shown that deep learning models and CNNs, specifically, are an encouraging solution to early CKD detection, and it is more sensitive and accurate compared to the traditional machine learning methods. However, the problems of data imbalance, model interpretation, and computational complexity are among the issues that should be taken into account to make it applicable to a larger clinical scale.