Dialysis adequacy predictions using a machine learning method.

Area of Science:

• Nephrology
• Biomedical Engineering
• Data Science

Background:

• Dialysis adequacy is a critical survival indicator for patients undergoing chronic hemodialysis.
• Traditional methods of assessing dialysis adequacy via blood samples present significant practical challenges and disadvantages.
• There is a need for non-invasive, continuous monitoring methods to evaluate dialysis effectiveness.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) models in predicting dialysis adequacy.
• To compare the performance of various ML and deep learning models against traditional linear regression.
• To utilize continuously measured data from hemodialysis machines for prediction.

Main Methods:

• Collected data from 1333 hemodialysis sessions across 61 patients.
• Utilized continuously measured demographic and clinical parameters from hemodialysis machines (240 measurements per session).
• Compared Random Forest, Extreme Gradient Boosting (XGBoost), Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and multivariable linear regression models using fivefold cross-validation.

Main Results:

• The XGBoost model demonstrated superior performance with the lowest Mean Absolute Percentage Error (MAPE) of 2.500 and Root Mean Square Error (RMSE) of 2.906, and highest Spearman's rank correlation coefficient (Corr) of 0.873.
• Deep learning models (CNN and GRU) showed comparable performance to XGBoost.
• Multivariable linear regression models exhibited the lowest performance, indicating the advantage of advanced ML techniques.

Conclusions:

• Machine learning models, particularly XGBoost, can accurately predict hemodialysis adequacy.
• Continuous data streams from hemodialysis machines are valuable inputs for ML-based adequacy prediction.
• This approach offers a promising, less invasive alternative for monitoring dialysis effectiveness and patient outcomes.