CKD.Net: A novel deep learning hybrid model for effective, real-time, automated screening tool towards prediction of multi stages of CKD along with eGFR and creatinine

Clinical tests have long been considered appropriate in diagnosing chronic kidney disease (CKD) because of their noninvasiveness, simplicity, and cost. Timely detection and management of CKD are the most effective methods to address the expanding global burden induced by CKD. We adopted an S-MTL (Supervised Multi-task Learning) approach and combined SimpleRNN (Simple Recurrent Neural Network) and MLP (Multi-Layer Perception) to develop a hybrid model-CKD.Net to predict five CKD stages. This hybrid neural network architecture was trained on massive clinical datasets with heterogeneous 27 features to predict kidney function. We employed various data augmentation strategies to balance the five CKD stage datasets and meticulously utilized the hyper -parameter to minimize the loss and validation loss to reduce overfitting and hence increase model generalization. Performance comparisons of CKD.Net were evaluated using Accuracy, Precision, Recall, and F1-score while comparing the performance with that of generic SimpleRNN and MLP models. CKD.Net demonstrated superior classification accuracy ranging from 99.2 to 99.8 percent in predicting the five classes. Furthermore, CKD.Net was utilized to predict eGFR (estimated glomerular filtration rate) and creatinine by evaluating the confidence level using Pearson correlation values. Subsequently, key risk factors of CKD were identified, and their clinical significance was discussed. CKD.Net web application was developed to automate the prediction of CKD disease. To the best of our knowledge, the CKD.Net model is the first essential step toward predicting multi-stages of kidney disease as an effective, real-time, automated screening tool. CKD.Net allows noninvasive measurement of kidney function, which is a crucial objective of artificial intelligence powered by functional automation in clinical practice.