血液透析
医学
概化理论
血压
Boosting(机器学习)
重症监护医学
透析
机器学习
人工智能
透明度(行为)
计算机科学
变量(数学)
预测建模
数据挖掘
鉴定(生物学)
内科学
特征选择
作者
Wun-Yi Huang,Cheng-Jui Lin,Yu-Xiang Zheng,Hung-Ting Chen,Fang‐Ju Sun,Chih‐Jen Wu,Kun‐Pin Wu
出处
期刊:Ndt Plus
[Oxford University Press]
日期:2025-11-14
卷期号:18 (12): sfaf349-sfaf349
摘要
Background: Intradialytic hypotension and hypertension significantly increase cardiovascular risk in hemodialysis patients, highlighting the need for effective monitoring systems. By predicting systolic blood pressure (SBP) values, monitoring capabilities can be improved beyond traditional event classification. Despite this potential advantage, current predictive models report a consistent error range (11-13 mmHg) and operate as black boxes. This study aims to develop a transparent SBP prediction model and evaluate its generalizability. Methods: This retrospective study analyzed data from 524 patients across metropolitan and suburban hospital branches (2016-19), collecting hemodialysis parameters, vital signs and predialytic measurements. The Explainable Boosting Machine (EBM) was used as the primary model to predict SBP changes. Cross-branch validation was implemented to evaluate model generalizability and perform feature selection. The EBM model was then compared with five other common machine learning methods. Results: EBM with 16 dynamic features achieved competitive performance compared with other machine learning methods (mean absolute error: 10.57-11.33 mmHg, Pearson's r: 0.80-0.83) in cross-validation between branches. A waterfall plot visualized the model's additive scoring process, showing how each feature quantitatively contributes to the final SBP prediction. Error analysis of the models revealed strong positive correlations (Pearson's r: 0.81-0.87) between patients' blood pressure variability and prediction errors. Conclusions: The EBM models demonstrated excellent cross-branch generalizability while providing feature-level transparency that reflects clinical understanding of dialysis physiology, unlike black-box models that require additional post-hoc explanation methods such as SHapley Additive exPlanations (SHAP). These findings provide a foundation for developing patient-specific solutions for individuals with highly variable blood pressure patterns.
科研通智能强力驱动
Strongly Powered by AbleSci AI