OpenConsult, an Open-Source Fully Local AI Scribe With Live Decision Support: Development and Formative Evaluation
Background. Ambient clinical AI — software that listens to a consultation and drafts the clinical note, often called an AI scribe — is now a flagship use of artificial intelligence in healthcare. Most tools delivering it are proprietary and cloud-based: consultation audio leaves the building, errors cannot be independently characterised, and the clinician who signs the note answers for output nobody outside the vendor can examine. Open, fully local alternatives that a clinician or researcher can inspect, run and modify remain rare. Objective. This paper reports OpenConsult, an open-source AI scribe with live clinical decision support that runs entirely on one machine in the consulting room and is inspectable end to end, and describes its formative evaluation. Methods. Evaluation was formative and used synthetic consultations only (scripted and acted). Automated tests ran fifteen scripted consultations with pre-written marking schemes, checking transcription quality gates, the urgency alarm’s behaviour across emergency and routine scripts, and whether every cited claim in a note points to a real transcript line. Five acted consultations followed in a real consulting room, and the AI assistants that built the system ran an adversarial security audit under the developer’s direction. The work is reported against the applicable DECIDE-AI items, one stage before live clinical evaluation. Results. OpenConsult is a fully local, open-source consultation assistant, built by a practising general practitioner directing AI coding assistants, running on a single consumer graphics card. It transcribes the consultation live and separates the speakers, drafts a structured clinical note with every sentence linked to its supporting transcript words, runs live decision support whose urgency alarm is shown only to the doctor, drafts referral letters, and offers a supervised spoken history-taking mode. In generated notes, 301 of 302 citations pointed to real transcript lines — an existence check, with clinical faithfulness unmeasured pending independent rating; the one uncited claim was flagged for manual checking. The urgency alarm fired in all eight scripted emergencies and stayed silent in six of seven routine consultations. Five acted consultations surfaced five defects the automated checks had missed, four of them serious; all were fixed, each pinned by a regression test. The security audit returned twelve findings, including one critical vulnerability, fixed and verified before any public exposure. Conclusions. OpenConsult is a working, auditable proof of concept — not yet a validated clinical tool — released for research and education only, with real patient care out of scope. It is intended as an open research instrument: it records whether the reviewing clinician opened the cited evidence before approving, how long review takes, and every revision of its decision support, making human review of ambient clinical AI measurable. Independent, prospective evaluation is required before any step toward care.
Authors
- HENRY W.W. POTTS (ORCID: https://orcid.org/0000-0002-6200-8804)
- M W I Herath (ORCID: https://orcid.org/0009-0008-6367-1588)
Institutions
- University College London (GB)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11771422/v1
- Primary Topic
- Artificial Intelligence in Healthcare and Education
- Type
- preprint