Predictive modeling of fluid status in hemodialysis: model development and internal validation using the MONitoring dialysis outcomes (MONDO) global database

作者：Yueh SH, Raimann J, Canaud B, Zhou M, Ye X, Mermelstein A, Kooman J, van der Sande F, Usvyat L, Kotanko P, Zhang H · 发表于：Renal failure · 年份：2026 · DOI：10.1080/0886022X.2026.2692105 · 研究领域：Renal Dialysis、Body Water、Kidney Failure, Chronic、Humans、Female、Middle Aged、Male、Electric Impedance、Dielectric Spectroscopy、Databases, Factual、Boosting Machine Learning Algorithms、Aged

BACKGROUND: Optimized fluid management is crucial in dialysis care because extracellular volume overload drives adverse cardiovascular outcomes. At the same time, comorbidities such as inflammation and protein energy wasting lead to decreased muscle mass and intracellular water. Accurate assessment of total body water (TBW) and its extracellular water (ECW) and intracellular water (ICW) compartments is therefore essential to guide ultrafiltration, evaluate dialysis adequacy, and monitor patient risk. METHOD: Using adult patients from the MONitoring Dialysis Outcomes (MONDO) 2012 cohort, we developed predictive models to estimate fluid volume compartments based on demographic data, laboratory values, treatment parameters, and multi-frequency whole-body bioimpedance spectroscopy (BIS) measurements. Clinical features were aggregated over an up-to-90-day look-back window, yielding 18,600 patients and 162,479 dialysis treatments. eXtreme Gradient Boosting (XGBoost) models were trained and tested using patient-level splits, with parallel models built either incorporating or excluding prior BIS measurements. RESULTS: Models including BIS data showed excellent accuracy (R2 > 0.85), models excluding BIS features achieved inferior performance (R2 = 0.73-0.81). In models using BIS inputs, recent bioimpedance changes dominated feature importance. Models without BIS data relied primarily on urea distribution volume, age, and height. CONCLUSION: These findings indicate that fluid volume...