Abstract: FR-PO0007
Trajectory-Based Machine Learning Reveals Progressive Mortality Risk Escalation Before Death in Patients on 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
- Holik, Radek, Renal Research Institute, New York, New York, United States
- Yueh, Sheng-Han, Renal Research Institute, New York, New York, United States
- Bergling, Karin, Renal Research Institute, New York, New York, United States
- Nandorine Ban, Andrea, Renal Research Institute, New York, New York, United States
- Usvyat, Len A., Renal Research Institute, New York, New York, United States
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
Background
Machine learning models can estimate mortality risk in hemodialysis patients, but how predicted risk evolves over time prior to death remains poorly characterized. Understanding temporal risk trajectories may help identify periods of accelerating clinical deterioration. This study evaluated longitudinal patterns of predicted mortality risk preceding death.
Methods
We retrospectively analyzed longitudinal clinical data from 46,425 US patients undergoing hemodialysis between January 2021 and July 2023, including 468,245 prediction windows. An eXtreme Gradient Boosting (XGBoost) model was trained to predict 90-day mortality using demographic, treatment, and laboratory variables summarized over rolling 90-day windows updated every 30 days. The dataset was randomly split into three subsets at the patient level: training (70%), validation (15%), and test (15%). Similar mortality prevalence was observed across all subsets. Predicted probabilities were generated for the test set. Among patients who died, predictions were aligned by days before death, with day 0 representing the death date. Mean predicted probabilities were summarized in 14-day bins to generate the average trajectory curve.
Results
Predicted mortality risk progressively increased as patients approached death (Figure 1). Risk trajectories showed more rapid escalation during the final 90 days preceding death. Early fluctuations more than 1000 days before death likely reflected fewer available observations.
Conclusion
Trajectory-based machine learning analysis demonstrated progressive escalation of predicted mortality risk prior to death in hemodialysis patients. These findings suggest that longitudinal risk trajectories may help characterize temporal patterns of worsening risk over time.
Acknowledgment
Artificial intelligence tools were used in the preparation of this abstract. The authors reviewed and approved the final content.
Figure 1: Predicted mortality probability trajectories aligned by days preceding death in hemodialysis patients. The dots represent individual predictions. The red line represents the 14-day moving average of predicted mortality risk.
Funding
- Commercial Support – Fresenius Medical Care