Exploring factors and models to predict post-dialysis volume overload status in maintenance hemodialysis patients based on pre-dialysis parameters.

OBJECTIVE

This study explored factors and models to predict post-dialysis volume overload status in maintenance hemodialysis patients (MHD) based on pre-dialysis parameters using machine learning.

MATERIALS AND

METHODS

Pre-dialysis clinical data, pre- and post-dialysis bioimpedance spectroscopy, and the ultrasound (US) assessment for the lung were involved. Intergroup comparisons, regression analysis, and the least absolute shrinkage and selection operator (LASSO) regularization algorithm were conducted to screen potential predictive factors. Seven machine learning algorithms (Random Forest, Naïve Bayes, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Gradient Boosting, and Neural Networks) were applied to construct prediction models.

RESULTS

This study included 120 MHD patients. The prevalence of post-dialysis volume overload status in participants was 31.67%. Regression analysis showed that age (p = 0.007), prescribed ultrafiltration volume (UFV)/weight ratio (p < 0.001), overhydration (OH) (p < 0.001), and pre-dialysis US B-lines (p = 0.015) were associated with post-dialysis volume overload status. After the LASSO regularization algorithm, prescribed UFV/weight ratio, OH, and pre-dialysis US-B lines were selected as the potential prediction factors for constructing prediction models. The best-performing model was the Random Forest with an area under the curve (AUC) of 0.96, accuracy of 91.67%, precision of 92.56%, recall of 91.67%, and F1 of 0.91.

CONCLUSION

Pre-dialysis parameters, including prescribed UFV/weight ratio, OH, and dialysis US-B lines, were predictive factors for post-dialysis volume overload status. The Random Forest model based on these parameters could predict the post-dialysis volume overload status with relative accuracy and may provide a helpful guide to optimal prescribed UFV.