Prediction of chronic kidney disease risk using multimodal data

Chronic kidney disease (CKD) is a widespread public health problem and often leads to kidney failure which needs hemodialysis or even kidney transplantation. Undoubtedly, prediction of the risk of CKD among healthy people is highly desirable and very meaningful. However, most studies in this field used logistic regression (LR) and produced results with limited accuracy. Also, these studies ignored unstructured data which contained useful information. To improve CKD prediction, in this study, we built a novel multimodal data model that integrated Bidirectional Encoder Representations from Transformers with Light Gradient Boosting Machine (termed MD-BERT-LGBM model hereafter), and applied it to a group of 3295 participants for CKD prediction study. We collected medical data for over three months from each participant. We compared this novel integrated framework with three conventional models: the LR, LGBM, and Multimodal Disease Risk Prediction algorithm based on Convolutional Neural Networks (CNN-MDRP). The experimental results show that the new MD-BERT-LGBM model outperformed all the three conventional models in terms of accuracy, recall, and Area Under the ROC curve (AUC), which are 78.12%, 75.65%, and 85.15%, respectively. This result demonstrates the potential of this proposed method in the clinical application of CKD prediction and prevention.