MO761: Real-Time Prediction of Intradialytic Hypotension using Machine Learning and Cloud Computing Infrastructure. (3rd May 2022)

• Record Type: Journal Article
• Authors: Zhang, Hanjie; Chaudhuri, Sheetal; Pickering, Aaron; Usvyat, Len A; Larkin, John; Waguespack, Peter; Kuang, Zuwen; Kooman, Jeroen; Maddux, Frank; Kotanko, Peter
• Is Part Of: Nephrology dialysis transplantation. Volume 37(2022)Supplement 3
• Publication Date: 2022-05-03
• Subjects: Nephrology -- Periodicals; Hemodialysis -- Periodicals; Kidneys -- Transplantation -- Periodicals; Hemodialysis; Kidneys -- Transplantation; Nephrology; Periodicals; 616.61
• DOI: 10.1093/ndt/gfac079.040
• Languages: English
• ISSNs: 0931-0509

Abstract

Abstract: BACKGROUND AND AIMS: Intradialytic hypotension (IDH) is a frequent complication of chronic hemodialysis that could lead to serious clinical outcomes and can be difficult to predict. We used machine learning (ML) and cloud computing infrastructure to develop and evaluate a model that enables real-time IDH prediction. METHOD: Data were analyzed from six dialysis clinics (Fresenius Kidney Care, Waltham, MA, United States) between January 2020 and November 2020. Data included demographic, clinical, treatment and laboratory were obtained from the electronic health record. The intradialytic data such as systolic blood pressure (SBP), diastolic blood pressure (DBP), heart rate and blood flow rate were transferred in real time to Amazon Web Services (AWS, Amazon Web Services, Inc., Seattle, WA, United States) using secure Internet of Things software. Data were randomly divided into model training (80%) and testing (20%) sets. IDH was defined as SBP less than 90 mmHg. A ML tool known as an extreme gradient boosting algorithm (XGBoost) was used. Each time when an intradialytic SBP was measured, the model, which was hosted in AWS, would enable real-time prediction (in less than a minute) of patients who were at risk of IDH within the next 15–75 min. Area under the receiver operating characteristic curve (AUROC) was used to measure the model's performance. Shapley values were calculated to derive variables that were considered highly predictive of IDH. RESULTS: We obtained data from 42 546 hemodialysis sessions involving 693 patients and 355 693 intradialytic observations. IDH occurred in 17% of the treatments. Our model reasonably predicted IDH, achieving an AUROC of 0.89. Top three predictors of IDH within the next 15–75 min were last intradialytic SBP, IDH rate of the last 10 treatments and mean nadir SBP of the last 10 treatments (Fig. 1 ). CONCLUSION: Our study clearly shows the possibility of predicting IDH within the next 15–75 min using ML and cloud computing infrastructure, it suggests that the tool has the potential to assist clinicians to intervene proactively in patients at risk for IDH in real time.