Abstract: FR-PO0026
AnemiaCare HD: Achieving Protocol-Adherent Anemia Management in Hemodialysis Using a Deterministic Large Language Model (LLM) Framework
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
- Arriola Montenegro, Jose J, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States
- Ordaya Gonzales, Karina M, Universidad Cientifica del Sur, Lima, Lima, Peru
- Thongprayoon, Charat, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States
- Miao, Jing, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States
- Craici, Iasmina, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States
- Cheungpasitporn, Wisit, Mayo Clinic Division of Nephrology and Hypertension, Rochester, Minnesota, United States
Background
Optimal anemia management in ESKD requires strict adherence to complex, institution-specific protocols. General-purpose LLM, such as ChatGPT-4o, offer clinical decision support capabilities but often deviate from narrowly defined clinical thresholds due to their inherent flexibility. Consequently, strict protocol adherence remains a critical unmet need for clinical deployment of LLM-based tools in hemodialysis anemia management.
Methods
In March 2025, we deployed AnemiaCare HD, a protocol-constrained, domain-specific GPT-based assistant engineered to deliver reproducible, guideline-compliant recommendations. It employed a deterministic, instruction-guided architecture limited to 8,000 characters and required six mandatory clinical inputs—hemoglobin, hemoglobin change rate, TSAT, ferritin, current Aranesp dose, and hemoglobin trend—collected in a rigid, sequential format. Phase 1 (baseline) utilized a loosely specified prompt. Phase 2 incorporated a fully deterministic prompt refined based on Phase 1 outcome analysis, embedding explicit dosing algorithms, timing restrictions, proactive safety checks, and edge-case management strategies.
Results
In Phase 1, the loosely specified prompt achieved a protocol adherence rate of only 32% (96/300 cases), despite human-in-the-loop auditing successfully identifying errors in 68% (214/300) of cases. Persistent variability highlighted the necessity for the absolute and unambiguous translation of all clinical rules into structured instructions. Following deterministic prompt refinements, protocol adherence improved to 100% (300/300 cases), with complete elimination of timing violations, dosing miscalculations, and unsafe recommendations.
Conclusion
AnemiaCare HD demonstrates that deterministic prompt engineering based on rigid input sequencing,explicit decision matrices, and fixed output structures enables large language model systems to achieve full protocol fidelity in clinical decision support. Comprehensive encoding of domain-specific operational rules is essential to ensuring safety, auditability, and sustained clinical trust. Future directions include integration with electronic health records, optimization of dose calculation algorithms, and prospective validation studies in real-world dialysis environments.