Abstract: TH-PO0516
Physiology-Driven Transformer Model to Predict Hemoglobin Response and Personalize Erythropoiesis-Stimulating Agent Dosing in Hemodialysis
Session Information
- Dialysis: Novel Therapeutics and Medication Management
November 06, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Zhao, Siwei, Sanderling Renal Services, Nashville, Tennessee, United States
- Luo, Yunfei, University of California San Diego, La Jolla, California, United States
- Rahman, Tauhidur, University of California San Diego, La Jolla, California, United States
- Dasgupta, Subhasis, University of California San Diego, La Jolla, California, United States
- Ahmadi, Armin, University of California San Diego, La Jolla, California, United States
- Tannenbaum, Jerome S., Sanderling Renal Services, Nashville, Tennessee, United States
- Ix, Joachim H., University of California San Diego, La Jolla, California, United States
- Malhotra, Rakesh, University of California San Diego, La Jolla, California, United States
Background
Anemia management in hemodialysis (HD) depends on individualized erythropoiesis-stimulating agent (ESA) dosing to achieve and maintain target hemoglobin (Hgb) levels. However, predicting individual Hgb responses remains a significant challenge due to nonlinear erythropoietic dynamics and patient heterogeneity. We developed a physiology-informed, transformer-based neural network to predict Hgb prediction and inform ESA dosing optimization.
Methods
We analyzed 728 adult hemodialysis patients from Sanderling clinics with ≥5 Hgb measurements. Clinical, dialysis, and laboratory data were structured into an event-based dataset, where each event included pre-dose Hgb, Mircera® and iron doses, and follow-up Hgb 7–21 days later. A transformer-based neural network was trained on sequences of five consecutive events, with a fixed red blood (RBC) kinetics module embedded to reflect Mircera® pharmacodynamics and constrain predictions. Model performance was evaluated using 5-fold cross-validation and mean absolute percentage error (MAPE). Simulations were conducted to model 40-day Hgb trajectories under hypothetical ESA doses (0–200 µg) in a representative patient.
Results
Mean age was 63 ± 14 years, with dialysis vintage 2.1 ± 1.8 years. Of these, 43% were female, 11% Black, and 40% had diabetes. Patients received 113.5 ± 56.4 mg/week of iron and 149.3 ± 91.6 µg/month of Mircera®. Mean Hgb was 10.1 ± 0.9 g/dL, transferrin saturation (TSAT) 31.0 ± 9.5%, and ferritin 703.4 ± 451.7 ng/mL. The physiology-informed transformer model predicted Hgb with a MAPE of ~6.2% (Figure 1-A). Simulated Hgb trajectories under varying Mircera® doses are shown in Figure 1-B.
Conclusion
By integrating pharmacodynamics with sequential clinical data, our transformer-based model provides a novel framework for individualized hemoglobin prediction. This approach may inform future efforts to optimize ESA dosing in hemodialysis care.
Funding
- NIDDK Support