AI Medical Scribe for NHS Lothian

Background

Clinical documentation is one of the most time-intensive parts of a clinician’s day. In primary care and outpatient settings, GPs and specialists must accurately record consultation details, assign clinical codes, and produce structured letters while maintaining patient-facing focus. Commercial AI scribe tools have begun entering this space, but they often lack integration with NHS coding standards and run on external infrastructure, raising legitimate information governance concerns for Health Boards.

NHS Lothian identified an opportunity to develop a locally controlled alternative: an AI medical scribe built on open-source foundations, running on NHS infrastructure, and aligned with NHS data governance standards from the ground up.

Miniml was engaged to deliver the technical development under a structured four-milestone programme.

The Challenge

The core requirements were:

• Accuracy for clinical language — general-purpose speech recognition performs poorly on drug names, clinical terminology, and specialist vocabulary.
• Structured outputs — transcripts alone are not useful; clinicians need SOAP notes, GP consultation templates, and OPD clinic letters.
• SNOMED CT integration — unlike commercial tools, the system needed to surface and map clinical codes aligned with NHS Lothian’s own recommended READ code set.
• Information governance compliance — all development on de-identified, synthetic data only; no integration with live EPR systems during this phase.
• Clinician usability — the tool had to fit clinical workflows, not require clinicians to adapt to it.

What We Built

Enhanced Clinical Speech Recognition

We fine-tuned OpenAI’s Whisper model on domain-specific medical speech data, incorporating simulated local accents alongside clinical terminology, drug names, and symptom language. Transcription accuracy improved significantly, with Word Error Rate dropping from 0.1219 to 0.0809 on the United-Syn-Med medical speech benchmark — a 34% reduction in errors.

A concrete example of the improvement: where the baseline model transcribed a drug name as “Yahtzee is a prescription medication”, the enhanced model correctly identified it as “YOCIT is a prescription medication.”

Specialty-Configurable Clinical Note Generation

The system generates structured consultation notes in multiple formats, including standard SOAP and extended GP Consultation templates. Clinicians can configure specialty-specific profiles that tailor note structure, background clinical context, and relevant terminology. This was demonstrated across specialties including Mental Health, Diabetes, and Learning Disability, with psychiatry teams noting particular potential for high-demand services such as ADHD assessments.

Note verbosity is configurable, allowing clinicians to choose between brief summaries and comprehensive documentation depending on workflow needs.

SNOMED CT Code Association

The system automatically identifies clinical concepts within generated notes and associates them with relevant SNOMED CT codes. The interface applies a three-tier highlighting system:

• Standard SNOMED CT codes
• NHS Lothian-recommended READ codes with SNOMED CT mapping
• Specialty-specific relevant codes

Clinicians retain full control: suggested codes can be reviewed, added to, modified, or removed before any documentation is finalised.

OPD Clinic Letter Generation

Structured consultation notes can be automatically transformed into outpatient department clinic letter format, reducing the manual effort involved in post-consultation documentation.

Integrated User Guide

A built-in user guide is embedded within the application interface, designed to support clinical adoption without requiring separate training materials.

Delivery

All four project milestones were completed on schedule over a four-month period. The programme included regular fortnightly meetings between the technical and clinical teams, iterative feedback cycles with NHS clinicians, and ongoing alignment with NHS Lothian’s information governance and eHealth requirements.

Development was conducted entirely in secure, offline environments using de-identified synthetic data. No real patient data was used at any stage of this phase.

Key Outcomes

What Clinicians Said

Early testing with NHS clinicians highlighted the tool’s potential to reduce documentation burden across primary care. Psychiatry colleagues specifically noted the value in high-demand, documentation-heavy services like ADHD assessment pathways, where consultation throughput is constrained partly by administrative load.

Why This Matters

AI scribes are already entering NHS settings through commercial channels. This project demonstrates that NHS organisations can develop locally controlled alternatives that are purpose-built for their infrastructure, coding standards, and governance requirements without compromising on capability. Critically, the SNOMED CT integration capability that commercial tools do not currently offer has direct implications for data quality, resource allocation, and patient care pathways across the NHS.

Developed in partnership with NHS Lothian and Health Innovation South East Scotland (HISES). All development conducted on de-identified synthetic data in compliance with NHS information governance standards.