Kidney Week

Abstract: FR-PO0011

Prediction of Intradialytic Hypotension Using Machine Learning in Hospitalized Patients Undergoing Intermittent Hemodialysis

Session Information

• Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Koyner, Henry, Cleveland Clinic, Cleveland, Ohio, United States

Henry Koyner

• Demirjian, Sevag, Cleveland Clinic, Cleveland, Ohio, United States

Sevag Demirjian, MD

Background

Intradialytic Hypotension (IDH) is a frequent and serious complication in hospitalized patients and is associated with morbidity and mortality. Prediction of IDH may enable providers to use interventions to minimize dialysis related complications. Given a set of biomarker data consisting of 2,775,940 data points, which were measured roughly every 15 minutes from 145,620 hemodialysis sessions, the goal of the project was to build a reliable predictor of hypotension using machine learning or other regression models.

Methods

We tested several machine learning models to predict IDH in critically ill patients based on clinical data measured at baseline and every 15 minutes during dialysis. After testing various thresholds, we defined IDH as a drop of 20 mmHg in systolic or a drop of 10 mmHg in diastolic blood pressure within 4 intervals (an hour) of a given reading. We developed and compared eight different prediction models, and also their collective ensemble.

Results

Each model was trained an tested on the 145,620 hemodialysis sessions in Python, then evaluated. XGBoost had the best performance accuracy of 70.4% and a precision of 50.3%, meaning more false positives. However, it had a recall of 71.2%. This led to an F-1 score of 0.589, an MCC of 0.382, and a Jaccard Index of 0.418 (see table).

Conclusion

We developed and compared models to predict IDH in hospitalized patients. The XGBoost based model outperformed the other machine learning methods tested and performed well in a single-center cohort. Future research should address generalizability in an external validation cohort, additional feature engineering, and developing models for other definitions of IDH.

Acknowledgment

I would like to acknowledge and thank Dr. Sevag Demirjian for providing me with the data set, as well as advice and direction for the project.

Model Performance (Sorted By F-1 Score)