Abstract: FR-PO0012
A Physiology-Driven Framework for Predicting Hemoglobin Response and Personalizing Erythropoiesis-Stimulating Agent Dosing in 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
- Zhao, Siwei, Sanderling Renal Services, Nashville, Tennessee, 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
- Luo, Yunfei, 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 accurate dosing of erythropoiesis-stimulating agents (ESAs), but hemoglobin (Hgb) response remains difficult to predict because it reflects delayed and highly variable physiological processes. To address this, we developed a physiology-driven framework centered on red blood cell kinetics and Mircera® pharmacodynamics. In this approach, the physiology model is the primary component, while the AI model serves only as a secondary tool to capture residual temporal patterns not fully explained by the mechanistic structure.
Methods
We studied 728 adult HD patients from Sanderling clinics who had at least 5 Hgb measurements while receiving intravenous Mircera® and iron. Clinical data, laboratory values, and treatment records were organized into an event-based dataset. Each event linked the pre-dose Hgb, ESA and iron exposure, and patient-level factors to a subsequent Hgb measurement obtained 7 to 21 days later. The core model incorporated RBC kinetics to represent delayed erythropoietic response, RBC production, and turnover under Mircera® treatment. A transformer-based model was included only as an auxiliary layer to improve prediction of post-dose Hgb across sequential treatment events. Model performance was evaluated using cross-validation and mean absolute percentage error (MAPE), and simulation studies were used to explore Hgb response under alternative dose scenarios.
Results
Among 728 HD patients, mean age was 63.1±14.0 years, dialysis vintage was 2.1±1.8 years, 43% were female, 11.3% were Black, and 40% had diabetes. Patients received 113.5±56.4 mg/week iron and 149.3±91.6 mcg/month ESA, with mean hemoglobin 10.1±0.9 g/dL, TSAT 31.0±9.5%, and ferritin 703.4±451.7 ng/mL. Using this event-based dataset, the physiology-driven modeling framework predicted post-dose hemoglobin with approximately 6.2% mean absolute percentage error (MAPE). Model predictions showed good agreement with observed hemoglobin trajectories over time. In a representative patient, the model captured delayed Hgb responses to Mircera® dosing and reproduced observed longitudinal trends, including both stable and dynamic phases of Hgb variation. (Firgure 1)
Conclusion
Our findings support a hybrid mechanistic–AI framework for predicting Hgb dynamics in HD, enabling more physiologically grounded and potentially individualized ESA dosing strategies.
Funding
- NIDDK Support