Prediction Model of Intradialytic Hypertension in Hemodialysis Patients Based on Machine Learning

• Published in: Journal of medical systems Vol. 49; no. 1; p. 112
• Main Authors: Wang, Yu, Zhou, Hongming, Guo, Qi, Wang, Kang, Luo, Yehua, Luan, Shaodong, Tang, Donge, Dong, Shuangyong, Yin, Lianghong, Dai, Yong
• Format: Journal Article
• Language: English
• Published: New York Springer US 11.09.2025; Springer Nature B.V
• Subjects: Adult; Aged; Algorithms; Artificial intelligence; Biomarkers; Blood pressure; Clinical decision making; Datasets; Dialysis; Electrolytes; End-stage renal disease; Female; Health Informatics; Health Sciences; Hemodialysis; Humans; Hypertension; Hypertension - diagnosis; Hypertension - etiology; Hypotension; Kidney diseases; Kidney Failure, Chronic - therapy; Learning algorithms; Machine Learning; Male; Medicine; Medicine & Public Health; Middle Aged; Missing data; Neurological complications; Parameters; Patients; Peritoneal dialysis; Prediction models; Predictive analytics; Real time; Recall; Renal Dialysis - adverse effects; Risk Assessment - methods; Risk groups; Statistics for Life Sciences; Support Vector Machine; Support vector machines; Venous access; China; End-stage renal disease; Hemodialysis; Light gradient boosting machine; Machine learning
• ISSN: 1573-689X; 0148-5598; 1573-689X
• DOI: 10.1007/s10916-025-02237-5

The escalating global burden of chronic kidney disease (CKD), particularly end-stage renal disease (ESRD), has intensified reliance on hemodialysis (HD), imposing substantial financial and operational burdens on healthcare systems and patients. Intradialytic hypertension (IDH), a critical complication during HD, is associated with life-threatening cardiovascular and neurological sequelae if unmanaged. This study aims to develop a machine learning (ML)-driven early-alert system for IDH risk prediction by integrating demographic profiles and dialysis session records, enabling clinicians to preemptively identify high-risk patients and prioritize targeted monitoring. Two clinical prediction models (IDH-1 and IDH-2) were developed using Light Gradient Boosting Machine (LGBM), Support Vector Machine (SVM), and TabNet algorithms. IDH-1 estimates immediate hypertension risk by analyzing pre-dialysis vital signs and longitudinal treatment patterns, whereas IDH-2 predicts subsequent session risks by synthesizing real-time dialysis parameters with historical biomarkers. Model performance was rigorously validated using standardized metrics, including AUC-ROC, sensitivity, accuracy, and F1 score, to ensure clinical applicability. 185,125 HD sessions as training set and 71,427 sessions as testing set were used in this study. For IDH-1, the LGBM model demonstrated superior discriminative capacity (AUC: 0.87; recall: 0.73; F1 score: 0.36), outperforming SVM and TabNet. Similarly, LGBM achieved the highest performance for IDH-2 (AUC: 0.74; recall: 0.56; F1 score: 0.26). Most significant parameters in IDH-1 Predictor with LGBM were pre-dialysis diastolic pressures, historical mean arterial pressure, and historical average IDH episodes. For the IDH-2 model with LGBM, historical average IDH episodes and post-dialysis systolic pressures were most important parameters. This study provides two kinds of superior discriminative capacity LGBM model for IDH predicting. The proposed models offer a scalable framework for personalized risk stratification, potentially mitigating adverse outcomes in hemodialysis populations.