Development and validation of predictive models for the hemodialysis adequacy index (Kt/V) using multiple linear...

Kt/V is a widely used indicator for assessing hemodialysis (HD) adequacy, but conventional measurement methods require blood sampling and are impractical for large-scale or real-time use. This study aimed to develop and validate predictive models for Kt/V using multiple linear regression (MLR) and machine learning algorithms, enabling clinicians to estimate dialysis adequacy before a session begins using simple, non-invasive parameters.

A total of 1,017 eligible patients were enrolled from seven HD centers in Hohhot City, China, and divided into training (n = 677), internal (n = 220), and external validation (n = 120) sets. From 26 candidate variables, eight clinically meaningful derived features were constructed, and a multi-algorithm fusion strategy integrating F test, Mutual Information, Random Forest importance, XGBoost importance, LASSO coefficients, Permutation Importance, Recursive Feature Elimination, and Boruta was employed for feature selection. Five final features (body surface area, pre-dialysis weight, standardized blood flow, sex, and circulating blood volume) were selected. Eight point-prediction models (MLR, XGBoost, LightGBM, Random Forest, Gradient Boosting, KNN, SVM, and BPNN), a simple average ensemble of the three top-performing models, and two probabilistic prediction methods (LightGBM Quantile Regression and NGBoost) were developed. Hyperparameter tuning was performed using GridSearchCV with 5-fold cross-validation. Model performance was evaluated using R², RMSE, and MAE on internal and external validation sets.

On external validation, the SVM model achieved the highest R² (0.5931) and lowest RMSE (0.1633) and MAE (0.1214) among individual models, followed by Gradient Boosting (R² = 0.5748, RMSE = 0.1670, MAE = 0.1276) and XGBoost (R² = 0.5681, RMSE = 0.1683, MAE = 0.1288). The ensemble model achieved the best overall performance (R² = 0.5911, RMSE = 0.1637, MAE = 0.1240). Calibration assessment revealed all models were well-calibrated without requiring post-hoc adjustment. For probabilistic prediction, NGBoost provided well-calibrated 80% prediction intervals (PICP = 82.27%) with narrower intervals than LightGBM Quantile Regression. SHAP analysis identified BSA, standardized blood flow, and gender as the most influential predictors.

Machine learning models, particularly SVM, the ensemble model, and NGBoost, can provide accurate, interpretable, and uncertainty-aware predictions of Kt/V using non-invasive, readily available parameters. The development of a local standalone executable application further facilitates clinical translation, enabling pre-dialysis prediction and real-time prescription adjustment. Prospective validation in diverse populations is warranted to confirm generalizability.