Abstract: FR-PO426
Machine Learning Approach to Predict Post-Hemodialysis Blood Pressure in Children With ESKD
Session Information
- Pediatric Nephrology - I
November 04, 2022 | Location: Exhibit Hall, Orange County Convention Center‚ West Building
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1800 Pediatric Nephrology
Authors
- Bou Matar, Raed, Cleveland Clinic, Cleveland, Ohio, United States
- Bobrowski, Amy, Cleveland Clinic, Cleveland, Ohio, United States
Background
Hypertension in end-stage kidney disease (ESKD) is associated with increased cardiovascular morbidity. Blood pressure (BP) control on chronic hemodialysis (HD) is directly related to fluid removal targets and interdialytic fluid gains (IDWG). Clinicians rely on clinical judgement to set a prescribed dry weight (DW), taking into consideration complex trends in BP, IDWG and other clinical parameters.
Methods
A system to predict post-HD BP values was developed using machine learning (ML) models to assist clinicians in determining optimal DW. Our dataset included patient-specific trends in BP, IDWG, heart rate and weights collected from 2011 to 2021. To help improve the model's performance and scalability, input features were selected based on initial descriptive analysis. Input features included age, height, DW, post-HD weight, pre-HD weight, pre-HD heart rate, pre-HD BP and IDWG. Six models were fed the input features, trained and hyperparameters tuned using Sci-kit Learn and XGBoost python libraries. Model performance was assessed utilizing time series cross validation on a rolling basis (30-90 day training, 1 day testing). Tukey's HSD test was applied to compare mean absolute error (MAE) values between models after Box-Cox normalization.
Results
Children who underwent chronic HD for at least 3 months were included (49 patients, 14604 dialysis sessions). Support vector machines regression (SVR) with a linear kernel achieved better accuracy than K-nearest neighbor (p=0.013), extreme gradient boosting (p<0.0001) and SVR with RBF kernel (p<0.0001). However, the differences in MAE between SVR (linear kernel), random forest and linear regression were not statistically significant (Table).
Conclusion
Utilizing vital signs trends and other readily available parameters, ML models may be useful in predicting post-HD BP in children with ESKD. Predictions are intended to guide DW adjustment, supplementing clinical judgment.
| R | MAE | RMSE | |
| Support Vector Machines (Linear kernel) | 0.75 | 8.21 | 10.5 |
| Linear Regression | 0.73 | 8.41 | 10.8 |
| Random Forest | 0.73 | 8.46 | 10.8 |
| K-Nearest Neighbors | 0.73 | 8.53 | 10.9 |
| Support Vector Machines (RBF kernel) | 0.71 | 8.71 | 11.1 |
| Extreme Gradient Boosting (XGBoost) | 0.68 | 9.42 | 12.1 |
R: Pearson's Correlation Coefficient; MAE: Mean Absolute Error; RMSE: Root Mean Square Error