Abstract:Chronic kidney disease (CKD) patients have high risks of end-stage kidney disease (ESKD) and pre-ESKD death. Therefore, accurately predicting these outcomes is useful among CKD patients, especially in those who are at high risk. Thus, we evaluated whether a machine-learning system can predict accurately these risks in CKD patients and attempted its application by developing a Web-based risk-prediction system. We developed 16 risk-prediction machine-learning models using Random Forest (RF), Gradient Boosting Decision Tree, and eXtreme Gradient Boosting with 22 variables or selected variables for the prediction of the primary outcome (ESKD or death) on the basis of repeatedly measured data of CKD patients (n = 3,714; repeatedly measured data, n = 66,981) in their electronic-medical records. The performances of the models were evaluated using data from a cohort study of CKD patients carried out over 3 years (n = 26,906). One RF model with 22 variables and another RF model with 8 variables of time-series data showed high accuracies of the prediction of the outcomes and were selected for use in a risk-prediction system. In the validation, the 22- and 8-variable RF models showed high C-statistics for the prediction of the outcomes: 0.932 (95% CI 0.916, 0.948) and 0.93 (0.915, 0.945), respectively. Cox proportional hazards models using splines showed a highly significant relationship between the high probability and high risk of an outcome ( p <0.0001). Moreover, the risks of patients with high probabilities were higher than those with low probabilities: 22-variable model, hazard ratio of 104.9 (95% CI 70.81, 155.3); 8-variable model, 90.9 (95% CI 62.29, 132.7). Then, a Web-based risk-prediction system was actually developed for the implementation of the models in clinical practice. This study showed that a machine-learning-based Web system is a useful tool for the risk prediction and treatment of CKD patients. Chronic kidney disease (CKD) patients have high risks of end-stage kidney disease (ESKD) and pre-ESKD death. Although the development of a new artificial intelligence (AI) model is expected to be useful for screening CKD patients at high risk, none of the models developed so far have yet been put into practical use. There are many reasons for the difficulties in developing and applying AI models to clinical settings: (1) AI models often have too many variables, some of which are uncommon, to input in busy clinical settings. (2) The accuracy of AI models to predict risks is not always high when evaluating nontarget patients' prognoses. (3) The codes of AI models do not always match electronic medical records of different hospitals. Therefore, in this study, AI models with a small number of commonly used variables were developed. Then, their performances were rigorously evaluated on subgroups of CKD patients classified on the basis of kidney function, age, and diabetes mellitus. Moreover, we developed and opened a Web-based risk-prediction system using the selected AI models that showed the highest accuracies for clinical practice: http://160.16.88.112:8000/. Our study fills the gap between AI development in research and AI in clinical practice.

Protocol to process follow-up electronic medical records of peritoneal dialysis patients to train AI models

Mortality Prediction with Adaptive Feature Importance Recalibration for Peritoneal Dialysis Patients: a deep-learning-based study on a real-world longitudinal follow-up dataset

Artificial intelligence and machine learning in peritoneal dialysis: a systematic review of clinical outcomes and predictive modeling

Enhancing End Stage Renal Disease Outcome Prediction: A Multi-Sourced Data-Driven Approach

Machine learning algorithms for the prediction of adverse prognosis in patients undergoing peritoneal dialysis

Implementation of artificial intelligence Chatbot in peritoneal dialysis nursing care: Experience from a Taiwan medical center

Machine Learning for the Prediction of First-Year Mortality in Incident Hemodialysis Patients

A Hemodialysis Mortality Prediction Model Based on Active Contrastive Learning.

Prediction of mortality in hemodialysis patients based on autoencoders

Use of deep learning to develop continuous-risk models for adverse event prediction from electronic health records

Precision Dialysis: Leveraging Big Data and Artificial Intelligence

Artificial Intelligence in Peritoneal Dialysis: General Overview

Forecasting acute kidney injury and resource utilization in ICU patients using longitudinal, multimodal models

Machine-learning-based Web system for the prediction of chronic kidney disease progression and mortality

Machine Learning for Identification of Short-Term All-Cause and Cardiovascular Deaths among Patients Undergoing Peritoneal Dialysis Patients

[Application of recurrent neural network in prognosis of peritoneal dialysis]

Predicting in-hospital outcomes of patients with acute kidney injury

NephroCAGE—German-Canadian Consortium on AI for Improved Kidney Transplantation Outcome: Protocol for an Algorithm Development and Validation Study

Electrocardiography-based Artificial Intelligence Algorithms Aid in Prediction of Long-term Mortality After Kidney Transplantation

In-hospital mortality, readmission, and prolonged length of stay risk prediction leveraging historical electronic patient records

Long-Term Efficacy of an AI-Based Health Coaching Mobile App in Slowing the Progression of Non-Dialysis Dependent CKD: A Retrospective Cohort Study