Flagship · The Diversity Engine
In our landmark study in JMIR AI, our team found that DALL-E and Midjourney significantly under-represent darker skin tones when generating medical imagery. The bias was significant (P < .001) and held across hundreds of generated images.
Our method overcomes this bias such that there is no statistically significant difference between AI-generated images and the real-life distribution of skin tones. This approach is now the basis of the SimPatient Diversity Engine.
Source paper
Ensuring appropriate representation in AI-generated medical imagery: a methodological approach to address skin tone bias
O’Malley AS, Veenhuizen MA, Ahmed A. · JMIR AI 2024;3:e58275
Read in JMIR AIThe Diversity Engine has since been extended to gender, age, and ethnicity dimensions, and is calibrated against Scotland 2022 Census benchmarks in addition to US demographic data.
Headline finding
The Diversity Engine reduced the demographic bias of standard generative models from statistically large to near-representative.
AI-generated medical images analysed
13Citations on Google Scholar
Our research
Authored by the St Andrews team. Published in BMC Medical Education, the JMIR family, PLOS ONE, Advances in Simulation, Simulation in Healthcare and more.
Ahmed A, Kerr E, O'Malley AS
BMC Medical Education · 2025
Quality-assured AI-authored exam questions performed no differently from human-authored questions when tested on 142 St Andrews students.
Read in BMC Medical EducationM'gadzah S, O'Malley AS
JMIR Formative Research · 2025
Complex prompting raised GPT-4 diagnostic accuracy from 60.4% to 90.1% (p < .001), with the largest gains on conditions prevalent in low- and middle-income countries.
Read in JMIR Formative ResearchO'Malley AS, Veenhuizen MA, Ahmed A
JMIR AI · 2024
Standard generative models under-represented darker skin tones (P < .001 vs US demographics). A custom model reduced the gap to P = .04. Foundation of the Diversity Engine.
Read in JMIR AIWetzlmair-Kephart LC, O'Malley A, O'Carroll V
PLOS ONE · 2025
248 questionnaire participants and 23 interviews across UK medical schools revealed a clear gap in teleconsultation training and an appetite to close it.
Read in PLOS ONEWetzlmair L, O'Carroll V, O'Malley AS, Murray SW
The Clinical Teacher · 2022
Systematic review of 14 studies (JBI methodology). Teleconsultation education increases student knowledge, confidence and satisfaction; further high-quality research and educator guidance is warranted.
Read in The Clinical TeacherVeenhuizen MA, O'Malley AS
Medicine & Health (AIMEC 2024 proceedings) · 2024
Conference paper at the 1st International Conference on AI in Medical Education (AIMEC 2024). Extended the diversity work to gender representation in clinical LLMs.
Read in Medicine & Health (AIMEC 2024 proceedings)Veenhuizen MA, O'Malley AS
Advances in Simulation · 2025
Compared AI-generated simulated patient cohorts against census benchmarks, extending the representation work from imagery to whole patient populations.
Read in Advances in SimulationO'Malley A, Veenhuizen M
Journal of the Academy of Medical Educators · 2026
Introduces a structured PDSA cycle for prompt engineering, giving educators a repeatable method for building reliable AI teaching tools.
Read in Journal of the Academy of Medical EducatorsO'Malley A
Simulation in Healthcare · 2026
A reflective account of deploying SimPatient to meet a real teaching-capacity challenge in undergraduate medical education.
Read in Simulation in HealthcareConference papers · 2025
Solving educational capacity challenges with an AI-powered patient simulator
O'Malley AS, Duggal S, Gordon I, Murad S, Wang X
SESAM 2025 · Society for Simulation in Europe, Valencia
Enhancing undergraduate clinical communication teaching and learning through AI simulation
Duggal S, Hughes A, O'Malley AS, Wang X, Murad S, Zachariou M, Gordon I
UKCCC 2025 · UK Council of Clinical Communications
Currently under review
Large language models display human-like social desirability biases in health screening questionnaires
O'Malley AS
Under review at Computers in Medicine
Deployed at scale
SimPatient is being rolled out across the University of St Andrews School of Medicine and the Scottish Graduate-Entry Medicine (ScotGEM) programme. Early data shows strong learner engagement.
Primary deployment
University of St Andrews
School of Medicine
Medical students
active on the platform
Live in the curriculum
this academic year
SimPatient is integrated into the medical-school curriculum at St Andrews. The School of Medicine’s Education Division is co-evaluating learning gain through the academic year.
Programme partner
ScotGEM
Scottish Graduate-Entry Medicine
Trains doctors for remote
and rural Scotland
Scenarios mapped to the
Medical Licensing Assessment
The St Andrews lead of the ScotGEM programme (Dr Andrew O’Malley, Senior Lecturer in the School of Medicine) is using SimPatient to scale consultation practice for students training to serve rural and remote communities across Scotland.
The frameworks we ship against
SimPatient’s rubric system isn’t invented in-house. It implements the published frameworks medical schools and GMC accreditation already use, with citations on every page.
Calgary-Cambridge
Clinical communication
Calgary-Cambridge model
The five-stage clinical-communication model: initiating, gathering information, physical examination, explanation & planning, and closing. Twenty named sub-skills across the consultation.
Kurtz SM, Silverman J, Draper J. Teaching and Learning Communication Skills in Medicine (2nd ed.). Radcliffe, 2005.
In product: Built-in rubric in SimPatient. Every consultation can be graded against the 20 sub-skills with section scores and transcript citations.
CRI-HT
Clinical reasoning
Clinical Reasoning Indicators for History Taking (CRI-HT)
Eight indicators rated 1–5 covering lead-taking, recognising relevant information, symptom specification, pathophysiological thinking, logical questioning, checking with the patient, summarising, and overall data quality.
Fürstenberg S, et al. Med Teach. 2020;42(8):914–921.
In product: Built-in rubric in SimPatient. Used by programmes that grade clinical reasoning explicitly alongside communication.
LCSAS
Validated assessment
Liverpool Communication Skills Assessment Scale (LCSAS)
A validated assessment scale developed at the University of Liverpool, with calibrated rating-scale labels and per-anchor descriptors for high-stakes communication-skills assessment.
University of Liverpool Communication Skills Assessment Scale.
In product: LCSAS-style anchors are first-class in SimPatient's custom-rubric editor, for programmes running validated, calibrated rubrics.
GMC MLA
UK regulatory
GMC Medical Licensing Assessment
The General Medical Council's Medical Licensing Assessment content map. The regulatory framework all UK medical schools must align to from 2024–25 onward.
General Medical Council, UK.
In product: Every SimPatient scenario maps to GMC MLA categories. Twelve curriculum categories aligned at launch.
Scottish Doctor
Scottish curriculum
Scottish Doctor learning outcomes
The agreed Scottish-medical-schools curriculum framework defining learning outcomes for undergraduate medical education in Scotland.
Scottish Deans' Medical Curriculum Group.
In product: AI-generated feedback maps performance to Scottish Doctor outcomes, alongside GMC MLA.
The path in
Primary
A 30-minute call with a clinician on our team. We’ll show the wizard, run a live consultation in your preferred mode, and walk through how rubric grading would map to your existing curriculum.
Book a demoLower-stakes
Run a 4-week pilot with a single cohort. We’ll help you set up your org, import your existing marking scheme, and share a written report at the end measuring usage, learner sentiment, and rubric performance.
Request a pilotAll pricing is per-organisation and includes unlimited learners. Get tailored pricing on the call.
