Construction and application of machine learning models for predicting intradialytic hypotension

Abstract

Introduction: Intradialytic hypotension (IDH) remains a prevalent complication of hemodialysis, which is associated with adverse outcomes for patients. This study seeks to harness machine learning to construct predictive models for IDH based on multiple definitions.

Methods: In this study, a comprehensive approach was employed, leveraging a dataset comprising 26,690 hemodialysis (HD) sessions for training and testing cohort, with an additional 12,293 HD sessions serving as a temporal validation cohort. Five definitions of IDH were employed, and models for each IDH definition were constructed using ten machine learning algorithms. Subsequently, model interpretation was facilitated. Feature simplification ensued, leading to the creation and evaluation of a streamlined machine learning model. Both the most effective machine learning model and its simplified counterpart underwent temporal validation.

Results: Across the five distinct definitions of IDH, the CatBoost model demonstrated superior predictive prowess, generally yielding the highest receiver operating characteristic - area under the curve (ROC-AUC) (Definition 1-5: 0.859, 0.864, 0.880, 0.848, 0.845). Noteworthy is the persistent inclusion of certain features within the top 20 across all definitions, including left ventricular mass index (LVMI), etc. Leveraging these features, we developed robust machine learning models that exhibited good performance (ROC-AUC for Definition 1-5: 0.866, 0.858, 0.874, 0.843, 0.838). Both the leading original machine learning model and the refined simplified machine learning model demonstrated robust performance on a temporal validation set.

Conclusion: Machine learning emerged as a reliable tool for predicting IDH in HD patients. Notably, LVMI emerged as a crucial feature for effectively predicting IDH. The simplified models are accessible on the provided website.

Fig 1. ROC curves of 10 machine learning models for the 5 definitions of IDH.

A–E show the ROC curves and ROC-AUC of 10 machine learning models for the 5 definitions of IDH, respectively. ‘Defn1’, ‘Defn2’, ‘Defn3’, ‘Defn4’, and ‘Defn5’ represent the 5 definitions of IDH, respectively. KNN, k-nearest neighbor; LR, Logistic Regression; DT, Decision Tree; ET, Extremely randomized Tree; RF, Random Forest; GBDT, Gradient Boosting Decision Tree; LGBM, Light Gradient Boosting Machine; XGBoost, Extreme Gradient Boosting; AdaBoost, Adaptive Boosting. ROC, Receiver Operating Characteristic Curve; AUC, Area Under Curve.

Fig 2. SHAP summary plots of CatBoost models for the 5 definitions of IDH.

A–E show the SHAP summary plots of CatBoost models for the 5 definitions of IDH, respectively. The top 20 features in order of importance are shown in the plots, and the horizontal coordinates in the plots are the SHAP values. A dot is created for the SHAP value of each feature for each HD session, with red indicating high feature values and blue indicating low feature values. The more positive a point’s SHAP value is, the more positively it affects the prediction, and conversely, the more negative a point’s SHAP value is, the more negatively it affects the prediction.

Fig 3. ROC curves of the simplified CatBoost models for the 5 definitions of IDH.

A–E show the ROC curves and ROC-AUC of the simplified CatBoost models for the 5 definitions of IDH, respectively. ‘Defn1’, ‘Defn2’, ‘Defn3’, ‘Defn4’, and ‘Defn5’ represent the 5 definitions of IDH, respectively. ‘3f’, ‘7f’, ‘8f’, and ‘11f’ represent the simplified CatBoost models with 3 features, 7 features, 8 features, and 11 features, respectively. ‘44f’ represents the original CatBoost model (with 44 features). ROC, Receiver Operating Characteristic Curve; AUC, Area Under Curve.

Fig 4. Model validation (ROC curves) of the original CatBoost models and the simplified CatBoost models for the 5 definitions of IDH.

A and B show the model validation (ROC curves and ROC-AUC) results of the original CatBoost models and the simplified CatBoost models for the 5 definitions of IDH, respectively. ‘Defn1’, ‘Defn2’, ‘Defn3’, ‘Defn4’, and ‘Defn5’ represent the 5 definitions of IDH, respectively. ‘11f’ represents the simplified CatBoost model (with 11 features), and ‘44f’ represents the original CatBoost model (with 44 features). ROC, Receiver Operating Characteristic Curve; AUC, Area Under Curve.

Fig 5. Input-Output Demonstration of the IDH Prediction System.