Protocol-Constrained Artificial Intelligence (AI) Enhances Tacrolimus Dosing Accuracy in Kidney Transplant Care

Abstract

Background: Tacrolimus, a cornerstone immunosuppressant in kidney transplantation, presents dosing challenges due to interpatient variability. While large language models such as ChatGPT-4o hold potential for clinical decision support, adherence to clinical guidelines remains inconsistent. TacroDose AI, a protocol-constrained GPT-based assistant, was developed to align tacrolimus dosing recommendations with Mayo Clinic’s guidelines, aiming to improve dosing accuracy and reduce sentinel errors. Methods: A dataset of 300 simulated tacrolimus dosing cases with varying trough levels was utilized to assess the performance of TacroDose AI. Following the identification of inappropriate dose adjustments, the model was refined to TacroAI 2.0 by implementing stricter calculation protocols, enhanced dose range definitions, and emphasis on total daily dose calculations. Statistical analyses, including paired t-tests and McNemar’s test, evaluated changes in accuracy and error rates between versions. Results: The targeted refinements in TacroAI 2.0 led to a marked improvement in dosing accuracy, increasing from 73.8% to 94.4% (p < 0.001), with total errors reduced from 75 to 16 and sentinel errors from 24 to 1. Dose adjustment errors decreased from 60 to 10, underscoring substantial improvements in guideline adherence. Residual errors persisted in percentage calculations for specific trough levels, primarily affecting cases with trough levels of 10-12 ng/mL, suggesting a need for further algorithmic refinements. Conclusion: The protocol-constrained TacroAI 2.0 significantly enhanced tacrolimus dosing accuracy in kidney transplant care, effectively aligning AI-driven recommendations with clinical guidelines and reducing sentinel errors. Future refinements should focus on optimizing percentage calculations for specific trough levels to mitigate residual errors. Prospective clinical validation and integration into existing EHR systems are recommended to facilitate broader clinical implementation.

Ähnliche Arbeiten