A Predictive Model for Real-time Prediction of Intradialytic Hypotension Based on Machine Learning Algorithms
Keywords: Hemodialysis, intradialytic hypotension, machine learning, predicting model, XGBoost
Abstract: Objective. Develop a machine learning-based model to predict IDH using pre-dialysis features. And to continuously predict IDH within the next hour during the dialysis session by incorporating real-time monitoring data. This approach helps in timely intervention, potentially reducing IDH rates and improving clinical outcomes for patients.
Methods. Collected maintenance hemodialysis (MHD) patients from October 1, 2021, to July 31, 2022, and divide them into development and validation datasets based on the treatment time point of May 1, 2022. IDH is defined as follows: (1) Nadir90: intradialytic systolic blood pressure (SBP) < 90 mmHg; (2) Fall20Nadir90: intradialytic SBP < 90 mmHg and a drop of ≥ 20 mmHg from pre-dialysis SBP. Analyzed the model’s predictive performance trained with various machine learning (ML) classification algorithms using k-fold cross-validation, evaluated by plotting the receiver operating characteristic curve (ROC) and precision-recall curve (PRC), calculating the area under the ROC (AUROC) and PRC (AUPRC), and computing the true positive rate (TPR), and true negative rate (TNR).The XGBoost algorithm was used to identify the important features required for the warning models.
Results. Data from 644 patients were analyzed, contributing 61,823 HD sessions with 302,942 intradialytic SBP measurements. IDH occurred in 2,659 (4.3%) HD sessions (Nadir90), in 1,706 (2.76%) sessions (Fall20Nadir90). Among various models compared, XGBoost achieved the best performance for predicting IDH before HD session (TPR: 0.6, TNR: 0.99, AUROC: 0.955, AUPRC: 0.686). Key predictive features included historical minimum SBP, average of historical minimum SBP, current SBP, diastolic blood pressure (DBP), IDH incidence rate, interdialytic weight change rate, prescribed dialysis duration, and dialysis vintage. The real-time model for predicting IDH within the next hour showed a TPR of 0.89, TNR of 0.92, AUROC of 0.959, and AUPRC of 0.38, with additional important features being mean arterial pressure (MAP), dialysis time, and ultrafiltration (UF) changes.
Conclusion. The XGBoost model has a high predictive capability for IDH during an ongoing HD session, assisting healthcare providers in assessing IDH risk and making timely decisions.
Submission Number: 25
Loading