Adults undergoing hematopoietic cell transplantation often develop serious complications that cause rapid nutritional decline. We developed and evaluated an AI approach to standardize intravenous nutrition (total parenteral nutrition, TPN) during this vulnerable period. Using real-world records from Stanford Health Care (6402 transplants, 2008-2025), we analyzed 1473 adults who received TPN, totaling 27,447 patient-days, linking each day's clinical state to the next day's prescription. We created a library of 30 standardized TPN regimens and trained a model to recommend next-day dose adjustments based on laboratory data and the existing prescription. (Pearson r ≈ 0.71). We then assessed an AI policy learned from past care and found that the Reinforcement learning agent selected dose adjustments with a higher composite score than the existing clinical policy. These results show that AI-guided TPN is feasible and may enhance bedside decision-making for adult transplant care, warranting prospective evaluation.
Ariss et al. (Wed,) studied this question.