Explainable machine-learning model for Real-Time Prediction of Intradialytic Hypotension (IDH)

Abstract Background: IDH is a common and serious complication of chronic hemodialysis that can lead to adverse long-term outcomes, including increased cardiovascular and all-cause mortality. Predicting IDH in advance may aid in establishing an intervention plan. Method: This study used 1173783 hemodialysis sessions from 837 patients and 211 features (including time-span varying and basic information features) as datasets. We constructed 21 features (including ion changes before and after dialysis) by clinical physicians from the raw datasets. We selected 89 features by grid search recursive feature elimination (RFE) based on extreme gradient boosting (XGBoost) (Gs-RFE-XGBoost) and XGBoost model to predict IDH. We used the Shapley value to achieve an explainable result for the 20 features. Besides, we used the quartile 2.5 to quartile 97.5 to get the reference range of the important continuous features. Result: We found the most important 20 features that are blood pressure related (including the datasets of 4 features and 4 construct features), platelet count, pathway type, pre-dialysis dry weight, dialyzer type, hematocrit, basophils, platelet distribution width, pre-dialysis uric acid, KT/V, parathyroid hormone (PTH), changes in calcium before and after dialysis, dialysis age. We get 0.8446 of avg recall, 0.9149 of avg F1, 0.9410 of avg PRAUC, 0.8823 of avg precision, and 0.8944 of avg MCC to predict the IDH in real-time. Conclusion: The Gs-RFE-XGBoost-Shapley methods can be applied as a practical, explained, and accurate method to predict IDH. The feature’s reference range can provide hints for the clinical physician to prevent the IDH of maintenance hemodialysis (MHD).