Machine Learning for Identification of Short-Term All-Cause and Cardiovascular Deaths among Patients Undergoing Peritoneal Dialysis Patients
Xiao Xu,Zhiyuan Xu,Tiantian Ma,Shaomei Li,Huayi Pei,Jinghong Zhao,Ying Zhang,Zibo Xiong,Yumei Liao,Ying Li,Qiongzhen Lin,Wenbo Hu,Yulin Li,Zhaoxia Zheng,Liping Duan,Gang Fu,Shanshan Guo,Beiru Zhang,Rui Yu,Fuyun Sun,Xiaoying Ma,Li Hao,Guiling Liu,Zhanzheng Zhao,Jing Xiao,Yulan Shen,Yong Zhang,Xuanyi Du,Tianrong Ji,Caili Wang,Lirong Deng,Yingli Yue,Shanshan Chen,Zhigang Ma,Yingping Li,Li Zuo,Huiping Zhao,Xianchao Zhang,Xuejian Wang,Yirong Liu,Xinying Gao,Xiaoli Chen,Hongyi Li,Shutong Du,Cui Zhao,Zhonggao Xu,Li Zhang,Hongyu Chen,Li,Lihua Wang,Yan,Yingchun Ma,Yuanyuan Wei,Jingwei Zhou,Yan Li,Jie Dong,Kai Niu,Zhiqiang He
DOI: https://doi.org/10.1093/ckj/sfae242
2024-01-01
Clinical Kidney Journal
Abstract:Although more and more cardiovascular risk factors have been verified in peritoneal dialysis (PD) populations in different countries and regions, it is still difficult for clinicians to accurately and individually predict death in the near future. We aimed to develop and validate machine learning-based models to predict near-term all-cause and cardiovascular death. Machine learning models were developed among 7539 PD patients, which were randomly divided into a training set and an internal test set by five random shuffles of 5-fold cross-validation, to predict the cardiovascular death and all-cause death in 3 months. We chose objectively collected markers such as patient demographics, clinical characteristics, laboratory data, and dialysis-related variables to inform the models and assessed the predictive performance using a range of common performance metrics, such as sensitivity, positive predictive values, the area under the receiver operating curve (AUROC), and the area under the precision recall curve. In the test set, the CVDformer models had a AUROC of 0.8767 (0.8129, 0.9045) and 0.9026 (0.8404, 0.9352) and area under the precision recall curve of 0.9338 (0.8134,0.9453) and 0.9073 (0.8412, 0.9164) in predicting near-term all-cause death and cardiovascular death, respectively. The CVDformer models had high sensitivity and positive predictive values for predicting all-cause and cardiovascular deaths in 3 months in our PD population. Further calibration is warranted in the future.