Human-Supervised Generative AI-Assisted Development of a Myocardial Perfusion Scintigraphy Standard Operating Procedure
Oleg Obolenski, Panithaya Chareonthaitawee, Ralph A. Bundschuh, Patrick Martineau, Kuangyu Shi, Maria Holstensson, Florian C. Gaertner, Barbara Essler, Markus Essler, Riemer H.J.A. Slart, Milka MarinovaPurpose of the Report:
Standard operating procedures (SOPs) are essential for reproducible nuclear medicine workflows but require substantial expert time to draft and maintain. This single-center feasibility study evaluated whether a human-supervised generative artificial intelligence (AI) workflow could support the development of a guideline-concordant SOP for myocardial perfusion scintigraphy.
Materials and Methods:
An AI-assisted SOP draft was generated using a large language model based on structured prompting and guideline references from the European Association of Nuclear Medicine, the Society of Nuclear Medicine and Molecular Imaging/American Society of Nuclear Cardiology, and the German Society of Nuclear Medicine (DGN). The AI-assisted draft underwent Human-in-the-Loop validation by nuclear medicine experts and was compared with an established institutional SOP. Efficiency was assessed by drafting and validation time. Guideline alignment was explored using the Action Corridor framework. A multidisciplinary panel of 7 reviewers evaluated both SOP versions using 5-point Likert scales.
Results:
The AI-assisted workflow reduced total drafting and validation time compared with manual SOP development (1.4 vs. 6.5 h). Guideline alignment was high for both SOPs (AI-assisted SOP, 96.8%; institutional SOP, 93.2%). In the exploratory panel assessment, reviewer ratings tended to favor the AI-assisted SOP draft for structure, clarity, guideline representation, and overall impression. No safety-relevant procedural discrepancy was identified after expert validation.
Conclusions:
Human-supervised GenAI-assisted drafting may support efficient preparation of guideline-concordant SOP documents in nuclear cardiology. The findings should be interpreted as descriptive and hypothesis-generating because only one AI-assisted curated SOP and one institutional SOP were compared. Expert validation remains essential before clinical implementation.