Xu Xiao, Xu Zhiyuan, Ma Tiantian, Li Shaomei, Pei Huayi, Zhao Jinghong, Zhang Ying, Xiong Zibo, Liao Yumei, Li Ying, Lin Qiongzhen, Hu Wenbo, Li Yulin, Zheng Zhaoxia, Duan Liping, Fu Gang, Guo Shanshan, Zhang Beiru, Yu Rui, Sun Fuyun, Ma Xiaoying, Hao Li, Liu Guiling, Zhao Zhanzheng, Xiao Jing, Shen Yulan, Zhang Yong, Du Xuanyi, Ji Tianrong, Wang Caili, Deng Lirong, Yue Yingli, Chen Shanshan, Ma Zhigang, Li Yingping, Zuo Li, Zhao Huiping, Zhang Xianchao, Wang Xuejian, Liu Yirong, Gao Xinying, Chen Xiaoli, Li Hongyi, Du Shutong, Zhao Cui, Xu Zhonggao, Zhang Li, Chen Hongyu, Li Li, Wang Lihua, Yan Yan, Ma Yingchun, Wei Yuanyuan, Zhou Jingwei, Li Yan, Dong Jie, Niu Kai, He Zhiqiang
Renal Division, Department of Medicine, Peking University First Hospital; Institute of Nephrology, Peking University; Key Laboratory of Renal Disease, Ministry of Health; Key Laboratory of Renal Disease, Ministry of Education; Beijing, China.
Key Laboratory of Universal Wireless Communications, Ministry of Education, Beijing University of Posts and Telecommunications, Beijing, China.
Clin Kidney J. 2024 Aug 29;17(9):sfae242. doi: 10.1093/ckj/sfae242. eCollection 2024 Sep.
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.
尽管在不同国家和地区的腹膜透析（PD）人群中已证实越来越多的心血管危险因素，但临床医生仍难以准确且个体化地预测近期死亡情况。我们旨在开发并验证基于机器学习的模型，以预测近期全因死亡和心血管死亡。在7539例PD患者中开发机器学习模型，通过5折交叉验证的五次随机洗牌将其随机分为训练集和内部测试集，以预测3个月内的心血管死亡和全因死亡。我们选择客观收集的指标，如患者人口统计学特征、临床特征、实验室数据和透析相关变量来为模型提供信息，并使用一系列常见的性能指标评估预测性能，如敏感性、阳性预测值、受试者工作特征曲线下面积（AUROC）和精确召回率曲线下面积。在测试集中，CVDformer模型预测近期全因死亡和心血管死亡时，AUROC分别为0.8767（0.8129，0.9045）和0.9026（0.8404，0.9352），精确召回率曲线下面积分别为0.9338（0.8134，0.9453）和0.9073（0.8412，0.9164）。CVDformer模型在预测我们PD人群3个月内的全因死亡和心血管死亡方面具有较高的敏感性和阳性预测值。未来有必要进行进一步校准。
