Machine-Learning Model to Predict the Intradialytic Hypotension Based on Clinical-Analytical Data

Abstract

Predicting whether patients will experience intradialytic hypotension (IDH) during hemodialysis (HD) is not an easy task. IDH is associated with multiple risk factors, meaning that traditional statistical models are unable to find the relationships that affect it. In this context, the use of models based on machine learning (ML) can allow the discovery of complex relationships, since they can solve problems without being explicitly programmed. In this work we developed, evaluated and identified an ML-based model that is capable of predicting at the beginning of the HD session whether a patient will suffer from IDH during its prolonged development. To develop the ML models, we used the hold-out and cross-validation methods; while, to evaluate the performance of the models we used the metrics F1-score, Matthews Correlation Coefficient, areas under the receiver operating characteristic (AUROC) and precision-recall curve (AUPRC). In this sense, we selected and used a reduced combination of variables from clinical records and blood analytics, which have proven to be decisive for the occurrence of IDH. The predictive results obtained through our work confirmed that the best ML model was based on the XGBoost model, achieving values of 0.969 and 0.945 for AUROC and AUPRC respectively. Therefore, our study suggests that the XGBoost model has a very high predictive capacity for the appearance of an IDH in HD patients and presents great versatility and flexibility in terms of supporting informed decision-making by medical staff.