Kidney Week

Abstract: SA-PO0085

Effect of Prediction Horizons on Machine Learning Mortality Models in Patients on Hemodialysis

Session Information

• Artificial Intelligence: Kidney Imaging, Pathology, and Precision Phenotyping
  October 24, 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

Radek Holik, MS

• Yueh, Sheng-Han, Renal Research Institute, New York, New York, United States

Sheng-Han Yueh, MS

• Bergling, Karin, Renal Research Institute, New York, New York, United States

Karin Bergling, MD, PhD

• Nandorine Ban, Andrea, Renal Research Institute, New York, New York, United States

Andrea Nandorine Ban, MS, MSc

• Usvyat, Len A., Renal Research Institute, New York, New York, United States

Len A. Usvyat, PhD, FASN

• Zhang, Hanjie, Renal Research Institute, New York, New York, United States

Hanjie Zhang, PhD

Background

Mortality risk among hemodialysis patients remains high. This highlights the need for accurate risk stratification tools. Machine learning approaches have shown promise in leveraging longitudinal clinical data for mortality prediction. However, it is unclear how the choice of prediction horizon affects model discrimination and classification performance. This study evaluates the effect of prediction window length on model discrimination and classification performance when comparing 90- and 180-day horizons.

Methods

We conducted a retrospective cohort study of 46,425 US patients who underwent hemodialysis between January 2021 and July 2023. This data included 145 variables related to demographics, treatment and laboratories. The features were summarized as 90-day averages and updated using a 30-day rolling window. Data was split into training, validation, and test sets. We addressed the issue of class imbalance by applying weighting. Two separate eXtreme Gradient Boosting models (XGBoost) were trained to predict 90-day and 180-day mortality. We evaluated model performance on a held-out test set using the following metrics: area under the curve (AUC), precision–recall area under the curve (PR AUC), precision, sensitivity, specificity, F1 score, and screen-positive rates.

Results

In the test set (69,753 prediction windows from 6,964 patients), mortality prevalence was 7.77% for 180-day and 3.63% for 90-day prediction. The AUC was slightly higher for the 90-day model (0.811 vs. 0.797). The sensitivity and specificity also increased slightly (from 0.515 to 0.542 and from 0.870 to 0.879, respectively), while the screen-positive rate decreased (from 16.0% to 13.7%). In contrast, precision and F1 score were reduced, consistent with the lower event prevalence. PR AUC was similarly lower for the 90-day model (0.195 vs 0.303). A summary of performance metrics is provided in Table 1.

Conclusion

Short-term (90-day) prediction demonstrated slightly improved AUC and higher sensitivity and specificity, while longer-term (180-day) prediction identified a greater proportion of patients at risk. These findings indicate that prediction horizons are associated with differences in model performance characteristics.

Acknowledgment

Artificial intelligence tools were used in the preparation of this abstract. The authors reviewed and approved the final content.

Funding

• Commercial Support – Fresenius Medical Care