Scholarlink

<p>Prediction of in-hospital death in chronic kidney disease (CKD) cohorts could aid medical resource allocation to reduce mortality. The relative contribution of dynamic data based on electronic health record (EHR) for prediction of CKD patients in-hospital death remains understudied. In this retrospective cohort study based on EHR data from the First Center of Chinese PLA General Hospital and the Sixth Center of Chinese PLA General Hospital, a long short-term memory (LSTM) model and a logistic regression (LR) feature engineer model based on sequential data, and a LR baseline model based on static data were developed to predict in-hospital death among CKD patients. The LSTM model, the LR feature engineer model, and the LR baseline model were compared. And the LSTM model was evaluated with different length of window size. Among 16,109 CKD hospitalized patients from two hospitals, 1196 (7.42%) in-hospital death were observed. The LSTM model was associated with an AUC of 0.978 (0.966–0.991) in the internal test set, and an AUC of 0.927 (0.909–0.946) in the external test set. The LSTM model showed better performance than the LR baseline model and the logistic regression feature engineer model (<i>p</i> &lt; 0.001). The LSTM model with 6-day sequential data could excel the performance of the LR feature engineer using the whole sequence data. In this study, these results suggested that the LSTM model using time sequential laboratory indicators contributed to earlier and more accurate prediction for in-hospital mortality in the CKD cohort.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Electronic health record dynamic features for in-hospital death among chronic kidney disease patients

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月24日

查看全文
  • Jianan Li,
  • Jing Qi,
  • Xiuzheng Yue,
  • Qin Zhong,
  • Chongyou Rao,
  • Qiuyang Li,
  • Kunlun He

摘要

Prediction of in-hospital death in chronic kidney disease (CKD) cohorts could aid medical resource allocation to reduce mortality. The relative contribution of dynamic data based on electronic health record (EHR) for prediction of CKD patients in-hospital death remains understudied. In this retrospective cohort study based on EHR data from the First Center of Chinese PLA General Hospital and the Sixth Center of Chinese PLA General Hospital, a long short-term memory (LSTM) model and a logistic regression (LR) feature engineer model based on sequential data, and a LR baseline model based on static data were developed to predict in-hospital death among CKD patients. The LSTM model, the LR feature engineer model, and the LR baseline model were compared. And the LSTM model was evaluated with different length of window size. Among 16,109 CKD hospitalized patients from two hospitals, 1196 (7.42%) in-hospital death were observed. The LSTM model was associated with an AUC of 0.978 (0.966–0.991) in the internal test set, and an AUC of 0.927 (0.909–0.946) in the external test set. The LSTM model showed better performance than the LR baseline model and the logistic regression feature engineer model (p < 0.001). The LSTM model with 6-day sequential data could excel the performance of the LR feature engineer using the whole sequence data. In this study, these results suggested that the LSTM model using time sequential laboratory indicators contributed to earlier and more accurate prediction for in-hospital mortality in the CKD cohort.