Physician-Supervised GPT Assistance for Communicating Cardiopulmonary Exercise Test Results

Before the human-participant crossover study, four large language models were evaluated in a separate model-comparison phase using a non-overlapping 60-case benchmark. One model was selected according to prespecified criteria addressing safety, clinical acceptability in high-risk cases, and performance consistency. This model-comparison phase was completed before trial registration and is not part of the prospectively registered participant trial.

Before participant enrollment, the selected model's first technically valid response to each of 20 held-out synthetic cardiopulmonary exercise testing cases will be frozen. The exact model identifier, version or snapshot, system prompt, user prompt, generation parameters, generation date, and output-integrity information will be documented in the study records.

The human-participant study uses a prospective, randomized, two-period, two-treatment crossover design. Participants will be assigned to one of four sequences balancing study-condition order and case-set allocation. Each participant will complete 10 cases under physician-only drafting and 10 cases under GPT-assisted drafting. The two periods will be separated by 7 plus or minus 2 days.

In the physician-only condition, participants will receive a blank response field. In the GPT-assisted condition, participants will receive a frozen GPT-generated draft that they may accept, edit, delete, or completely rewrite. The physician remains responsible for the submitted final response. Each case has a maximum completion time of 8 minutes. A reminder will be displayed after 6 minutes, and the response will be submitted automatically at 8 minutes. Participants will receive a fixed 3-minute rest after every five cases. Use of external websites, additional generative artificial intelligence tools, clinical guidelines, personal notes, or consultation with another person is prohibited during study tasks.

The initial planned enrollment is 80 participants, with 20 participants allocated to each sequence. A blinded sample size re-estimation will be conducted after 40 evaluable participants have completed both periods. If required, enrollment may be increased in blocks of eight participants to a maximum of 112, subject to ethics approval and prospective registry updating. All study cases are synthetic. No real patient records will be used. Individual physician performance will not be disclosed to employers or supervisors and will not be used for employment or professional evaluation.

Inclusion criteria

• Licensed physicians or institutionally recognized clinical trainees.
• Able to read and write Chinese.
• Experience in, or relevant training involving, cardiopulmonary exercise testing, cardiology, respiratory medicine, rehabilitation medicine, or related clinical interpretation within the previous 24 months.
• Able to complete the computer-based study procedures.
• Provision of written informed consent.

Exclusion criteria

• Participation in the development of the study case bank, reference answers, scoring manual, or frozen GPT-generated drafts.
• Previous exposure to any Phase 2 study case.
• Inability to complete the study tasks independently using the study computer interface.
• A direct financial conflict of interest related to the evaluated artificial intelligence system.
• Withdrawal of consent before completion of the assigned study procedures.