Skip Navigation
Skip to contents

대한신장학회


간행물 검색

현재 페이지 경로
  • HOME
  • 간행물
  • 간행물 검색
논문분류 춘계학술대회 초록집
제목 Machine Learning Models for Predicting Intradialytic Hypotension
저자 Hyung Woo Kim, Seok-Jae Heo, Chung-Mo Nam, Beom Seok Kim
출판정보 2021; 2021(1):
키워드
초록 Objective: All hemodialysis sessions from September 2006 to November 2019 in Severance Hospital were screened. Intradialytic hypotension was defined as nadir systolic blood pressure (BP) <90 mmHg (definition 1) or decrease in systolic BP ≥20 mmHg compared to the initial BP (definition 2). Demographics, comorbidities, hypertensive medication, laboratory values, electrocardiogram parmeters, vascular access, dialyzer, ultrafiltration rate, blood flow, and dialysate flow were used as features. In addition, several features from previous dialysis which contains historical information were created. The area under the receiver operating characteristic curves (AUROC), the area under the precision-recall curves (AUPRC) were observed to compare machine learning models including logistic regression, lasso regression, random forest, XGBoost, deep neural network, and scoring model. Methods: A total of 159,078 hemodialysis sessions were finally collected after excluding sessions having missing variables. The number of sessions in which intradialytic hypotension occurred was 5,978 sessions (3.8%) and 68,668 sessions (43.2%) according to definition 1 and 2, respectively. Compared with other models, the deep neural network model had the best performance to predict nadir SBP <90 mmHg (overall AUROC, 0.921) and XGBoost model had the best performance to predict decrease in SBP ≥20 mmHg compared to the initial BP (overall AUROC, 0.778). Results: Machine learning and deep learning methods can present the risk of intradialytic hypotension in advance before hemodialysis session start. Conclusions: Objective: Intradialytic hypotension occurs frequently and is an important prognostic factor in patients with chronic hemodialysis. However, it is difficult to predict intradialytic hypotension with classical statistical model because there are many different factors in intradialytic hypotension. Therefore, we aimed to develop the prediction model for intradialytic hypotension using machine learning methods. Methods: All hemodialysis sessions from September 2006 to November 2019 in Severance Hospital were screened. Intradialytic hypotension was defined as nadir systolic blood pressure (BP) <90 mmHg (definition 1) or decrease in systolic BP ≥20 mmHg compared to the initial BP (definition 2). Demographics, comorbidities, hypertensive medication, laboratory values, electrocardiogram parmeters, vascular access, dialyzer, ultrafiltration rate, blood flow, and dialysate flow were used as features. In addition, several features from previous dialysis which contains historical information were created. The area under the receiver operating characteristic curves (AUROC), the area under the precision-recall curves (AUPRC) were observed to compare machine learning models including logistic regression, lasso regression, random forest, XGBoost, deep neural network, and scoring model. Results: A total of 159,078 hemodialysis sessions were finally collected after excluding sessions having missing variables. The number of sessions in which intradialytic hypotension occurred was 5,978 sessions (3.8%) and 68,668 sessions (43.2%) according to definition 1 and 2, respectively. Compared with other models, the deep neural network model had the best performance to predict nadir SBP <90 mmHg (overall AUROC, 0.921) and XGBoost model had the best performance to predict decrease in SBP ≥20 mmHg compared to the initial BP (overall AUROC, 0.778). Conclusions: Machine learning and deep learning methods can present the risk of intradialytic hypotension in advance before hemodialysis session start.
원문(PDF) PDF 원문보기
위로가기