Autonomous artificial intelligence (AI) models for deciding treatment strategies are available but rarely applied prospectively in clinical settings. Here we present a prospective study of deploying daGOAT, an algorithm we have developed, as a conditional autonomous AI agent to prescribe a drug to prevent severe (grade 3−4) acute graft-versus-host disease (acute GvHD) following human leukocyte antigen (HLA)-mismatched haematopoietic cell transplantation (ClinicalTrials.gov, NCT05600855). During the enrollment period physicians invite 85% of eligible patients to participate and 88% of the invited patients agree. Among the 110 enrolled participants who receive HLA-haploidentical transplants, daGOAT predicts intermediate to high risk of severe acute GvHD in 57 participants between days +17 and +23 posttransplant and prescribes ruxolitinib in addition to the existing regimen to intensify immune suppression. The initial compliance with AI prescription is 98% (56/57), with dose and/or schedule deviating from the AI prescription within one month in a total of eight participants. In conclusion, we show that many physicians and patients are receptive to using conditional autonomous AI to prescribe a drug and that the decision for pharmaceutical intervention could be facilitated by autonomous AI. Autonomous artificial intelligence (AI) models to replace human decision-making in medical intervention need thorough testing. Here authors present the results of a clinical trial, NCT05600855, in which daGOAT, a conditional autonomous artificial intelligence agent successfully makes the decision whether to prescribe an immune suppressive drug to prevent severe acute graft-versus-host disease following HLA-mismatched haematopoietic cell transplantation.
Chen et al. (2025) studied this question.