Time Burden of Physician Edits in AI-Assisted Patient Messages
Poursoltan and colleagues examined physicians' edits to AI-generated patient messages and the time cost involved in this process.
A study conducted by Poursoltan and colleagues investigated the edits made by physicians to AI-generated patient messages and the associated time burden. It was found that clinical interpretation messages had the highest time burden per message, while appointment arrangements constituted the largest cumulative burden.
Scope of the Research
Poursoltan and colleagues detailedly examined the modifications made by physicians on AI-assisted patient messages and the resulting time cost.
Methodology and Dataset
The researchers developed an expert-validated taxonomy consisting of 15 editing categories using a large language model and evaluated the frequency, scope, and relationship of these edits with response time. The initial dataset included 14,350 AI-assisted responses belonging to 1,131 physicians.
Retrospective Cohort Study
Patient messages at UC San Diego Health between April 2024 and August 2025 were analyzed. A total of 13,632 responses belonging to 1,110 physicians were included in the study.
Most Frequent Edits
The analyses showed that the most common edits were scheduling or rescheduling appointments at a rate of 38.5%. This was followed by providing lifestyle or non-pharmacological guidance at 18.2%, and offering empathy or emotional support at 16.1%.
Impact on Response Times
It was determined that edits related to radiology interpretations prolonged the response time by 70.1%. Additionally, diagnostic clarification was associated with a 63.9% longer response time, and laboratory interpretation with a 60.8% longer response time.
Future Priorities
The findings contribute to identifying potential priorities for improving AI draft systems in order to reduce the workload of physicians.