Voa Health의 AI 기록 보조 도구가 외래 진료 음성을 초안 기록으로 변환하는 과정을 다중 병원에서 AI 사용 여부를 비교하여 의사의 웰빙, 문서화 부담 및 환자 경험을 평가하는 무작위 임상 시험. (SOAR)

This study is a randomized controlled trial designed to assess the impact of an ambient artificial-intelligence (AI) scribe on physician well-being, documentation workload, and patient experience in routine outpatient care.

The intervention consists of using the Voa Health ambient AI scribe during clinical encounters. The system records the audio of the consultation and generates a structured draft clinical note in real time, aligned with specialty-specific templates that reflect the routine workflow of each clinic (for example, different templates for general cardiology, heart failure, dyslipidemia, etc.). At the end of the visit, the physician reviews, edits, and signs the draft in the electronic medical record (EMR), remaining fully responsible for the accuracy and completeness of the documentation. In the control condition, physicians conduct consultations and document encounters using their usual methods without AI support. For study purposes, audio may still be recorded in the control arm, but no AI-generated note is displayed or used by the clinician.

The unit of randomization is the individual consultation. For participating physicians, eligible visits are automatically allocated to one of two parallel arms: (1) usual documentation without AI and (2) documentation assisted by the ambient AI scribe. Randomization is designed to preserve the existing organization of each clinic and to avoid interference with scheduling or patient flow. Clinical care, diagnostic and therapeutic decisions, and follow-up procedures are not dictated by the protocol and follow usual practice; the only experimental element is the use (or non-use) of the AI scribe for documentation and the collection of audio and questionnaires.

Adult patients seen in participating outpatient clinics, and their physicians, are invited to take part. After informed consent, the entire consultation is audio-recorded. Immediately after each visit, both patient and physician are asked to complete brief, structured questionnaires that capture the main outcomes of interest. To keep data collection feasible in a busy ambulatory setting, the instruments were built from subsets of items derived from internationally used scales, while keeping the number of questions per consultation small.

For physicians, items are drawn from the Professional Fulfillment Index (PFI), the Mini-Z 2.0 survey, and the 4-item Physician Task Load / NASA-TLX. These items assess professional fulfillment and burnout (physical and emotional exhaustion), perceived sufficiency of time for documentation, work in the EMR outside direct patient contact, perceived documentation burden, and temporal demand of the visit. Additional study-specific items evaluate the perceived quality and completeness of the final note, time required to edit the AI-generated draft, confidence that key clinical details were captured, occurrence of potential AI "hallucinations" (information not actually stated in the visit), and the perceived impact of documentation on attention to the patient.

For patients, questionnaires use items derived from the Consultation and Relational Empathy (CARE) Measure, the Patient Satisfaction Questionnaire Short-Form (PSQ-18), and the Communication Assessment Tool (CAT). These items cover domains such as active listening, understanding of patient concerns, clarity of explanations, adequacy of time spent with the physician, perceived empathy, overall satisfaction with care, and understanding of diagnosis and treatment. In the AI arm, one additional item specifically asks whether the use of AI during the consultation helped, did not change, or hindered the clarity of communication with the physician.

The primary outcomes are physician-reported well-being and perceived documentation workload when using the ambient AI scribe compared with usual documentation. Key secondary outcomes include patient-reported experience and satisfaction, physician-rated quality and completeness of notes, time required for documentation and for editing AI-generated drafts, and the frequency and clinical relevance of AI-related documentation errors or hallucinations. All outcomes are measured at the level of the individual consultation, immediately after each visit.

The trial is initially conducted in multiple outpatient clinics of Hospital de Clínicas of the Federal University of Paraná (UFPR), Brazil, across different medical specialties. In each service, structured note templates are developed in collaboration with local clinical leaders so that the AI-generated drafts reflect the real-world flow of that specialty without changing the standard of care. Medical students and residents trained in the protocol may support consent and questionnaire administration under supervision of attending physicians, to ensure consistent and feasible data collection.

Data are stored in secure, access-controlled servers, with linkage between audio recordings, questionnaires, and EMR notes managed through coded identifiers. A data monitoring committee, independent from the development team of the AI scribe, periodically reviews aggregated data for protocol adherence, data quality, and any safety concerns related to the use of AI in documentation (for example, systematic documentation errors that could potentially affect patient care). Because the intervention is limited to documentation support and clinicians remain responsible for all clinical decisions and for finalizing the notes, the overall risk of participation is considered minimal.

The protocol allows for future expansion to other outpatient services and collaborating centers that adopt the same randomization, data collection procedures, and outcome definitions. The results are expected to provide pragmatic evidence on how ambient AI scribes can be implemented safely and effectively in real-world clinical practice, particularly regarding their impact on physician well-being, documentation workload, and the patient's experience of the consultation.