Kidney Week

Abstract: FR-PO0014

Artificial Intelligence-Based Model for Personalized Erythropoiesis-Stimulating Agent Dosing in Patients on Hemodialysis Who Have Anemia

Session Information

• Artificial Intelligence and Digital Health at the Bedside
  November 07, 2025 | Location: Exhibit Hall, 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

• Chae, Seung Yun, The Catholic University of Korea College of Medicine, Seocho-gu, Korea (the Republic of)

Seung Yun Chae, MD, PhD

• Lee, Hanbi, The Catholic University of Korea College of Medicine, Seocho-gu, Korea (the Republic of)

Hanbi Lee, MD

• Eum, Sang Hun, The Catholic University of Korea College of Medicine, Seocho-gu, Korea (the Republic of)

Sang Hun Eum, MD

• Shin, Seok Joon, The Catholic University of Korea College of Medicine, Seocho-gu, Korea (the Republic of)

Seok Joon Shin, MD, PhD

• Yoon, Hye Eun, The Catholic University of Korea College of Medicine, Seocho-gu, Korea (the Republic of)

Hye Eun Yoon, MD, PhD

Background

Erythropoiesis-stimulating agents (ESAs) are key to managing anemia in hemodialysis patients, but traditional, experience-based dosing often falls short. This study developed AI-based models to optimize ESA dosing using detailed clinical data.

Methods

We retrospectively analyzed 148,381 monthly hemodialysis sessions from 9,738 patients at 8 university hospitals between 2009 and 2022. Two models were compared for predicting hemoglobin levels and ESA dose: an ML model using 25 variables (3 demographic factors [age, smoking history, height], 15 laboratory variables, and 7 dialysis-related measures) with 3-step time lag, and a DL-GRU+ATTENTION model using 16 variables (2 demographic factors [gender, height], 11 laboratory, and 3 dialysis-related measures) with the same time integration.

Results

For hemoglobin prediction, the XGBoost model achieved an MSE of 0.33, RMSE of 0.58, MAE of 0.36, and R² of 0.67, while the DL-GRU+ATTENTION model produced similar error values but demonstrated a higher R² of 0.75, indicating better explanatory power. In predicting ESA doses, machine learning models showed less accuracy, with MSEs ranging from 382 to 431, RMSEs from 10.77 to 20.00, MAEs from 6.36 to 10.62, and low R² values between 0.003 and 0.09. In contrast, the DL-GRU+ATTENTION model showed markedly improved performance, achieving an MSE of 16.36, RMSE of 4.05, MAE of 1.33, and an R² of 0.96, reflecting its strong predictive precision.

Conclusion

The DL-GRU+ATTENTION model demonstrated superior performance over conventional ML approaches in predicting both hemoglobin levels and appropriate ESA dosing. This suggests that deep learning-based methods may offer a more reliable framework for implementing precision medicine in anemia management among hemodialysis patients, with the potential to enhance treatment outcomes.

Digital Object Identifier (DOI)

• doi: 10.1681/ASN.2025rgwsdwty