Abstract: FR-PO0238
Retraining the Anemia Control Model (ACM) for the North American Dialysis Population: A Large-Scale In-Silico Trial
Session Information
- Anemia and Iron Metabolism
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Anemia and Iron Metabolism
- 200 Anemia and Iron Metabolism
Authors
- Garbelli, Mario, Renal Research Institute, Vaiano Cremasco, Italy
- Bellocchio, Francesco, Renal Research Institute, Vaiano Cremasco, Italy
- Barbieri, Carlo, Fresenius Medical Care Italia SpA, Palazzo Pignano, Lombardia, Italy
- Mari, Flavio, Fresenius Medical Care Italia SpA, Palazzo Pignano, Lombardia, Italy
- Usvyat, Len A., Renal Research Institute, Vaiano Cremasco, Italy
- Larkin, John W., Renal Research Institute, Vaiano Cremasco, Italy
- Jiao, Yue, Renal Research Institute, Vaiano Cremasco, Italy
- Wolf, Melanie, Renal Research Institute, Vaiano Cremasco, Italy
- Stuard, Stefano, Fresenius Medical Care Italia SpA, Palazzo Pignano, Lombardia, Italy
- Neri, Luca, Renal Research Institute, Vaiano Cremasco, Italy
Background
The ACM is a certified AI-based medical device designed to optimize anemia management in dialysis patients with CKD. Trained on over 900,000 hemoglobin (Hb) assessments from the European Nephrocare network, it has effectively guided therapy while reducing excessive drug use. To expand its global relevance, we are re-training the ACM using the ApolloDialDb, which includes clinical data from ~500,000 patients across diverse geographies and ethnic backgrounds. This effort supports the development of fair and generalizable decision-support tools. Our focus is on adapting the model to methoxy polyethylene glycol-epoetin beta (MPG-E), a drug more commonly used in the U.S., and revising ESA dosing strategies to reflect regional treatment targets (10–11 g/dL in the U.S. vs. 10–12 g/dL in Europe). We aim to validate the ACM’s performance and clinical utility in a heterogeneous North American dialysis population.
Methods
The ACM was trained using inputs such as laboratory results, hydration status, body weight, dialysis treatment data, demographics, and previous ESA/iron prescriptions. The model outputs a predicted Hb value to guide the minimal ESA dose required to achieve or maintain hemoglobin targets. Because real-world Hb cannot directly evaluate the model—given that an actual intervention is being made—we employed an in-silico trial. This approach applied the ACM-derived dose recommendations with a prediction error reflecting the model’s performance, enabling estimation of target attainment and potential ESA savings without a full clinical trial.
Results
We included 44,995 patients for a total of 2,735,733 patient-months. Each patient’s observed Hb served as a baseline, and we focused on those meeting the model’s eligibility criteria. Their observed Hb was compared with in silico predictions to assess how closely the model replicated real-world outcomes. Overall, the proportion of patients achieving the hemoglobin target increased from 63% to 70%, and the estimated ESA usage decreased by more than 56%.
Conclusion
We developed and tested an updated ACM using the North American cohort. Retraining improved performance and preserved fairness across demographics. In-silico validation showed effective hemoglobin control and reduced ESA use.