Published research · npj Digital Medicine

MAPLES in practice: AI‑assisted OSCE assessment

A prospective UTSW study brings multimodal AI into an objective structured clinical examination (OSCE), with targeted human review.

Research update

Published 7 October 2026 Article in Press

Prospective deployment of multimodal AI grading for medical student OSCEs

Huong-Tra Ngo and Ameer Hamza Shakur are co-first authors, with Andrew R. Jamieson as the corresponding and senior author.

Part of npj Digital Medicine’s Transforming Medical Education through Artificial Intelligence collection.

Assessment across notes, audio and video.

In Fall 2025, MAPLES was deployed prospectively at UT Southwestern in a primarily formative OSCE. Written post-encounter notes, audio and video supported assessment of clinical documentation, communication and physical examination performance.

Second-year medical students
222
OSCE stations per student
5
Retained item-level scores
72,907

A focused role for human review

Low-scoring learners were routed for independent educator rescoring, with physician adjudication of selected AI–human disagreements. The study counted 5,616 actual human grading passes against 72,907 passes in a modeled, single-pass manual comparator: 92.3% fewer grading passes. This comparison counts grading passes; it does not measure time or cost savings.

From deployment to ongoing oversight.

The UTSW program expanded in stages. Explore how the assessment sources and platform operation evolved, alongside the governance cycle that supports oversight.

A staged rollout at UTSW

How the workflow evolved

The timeline includes earlier and later program milestones. The quantitative results above describe the Fall 2025 study only.

Fall 2023

Start with clinical notes

AI-assisted grading of post-encounter notes.

Assessment sources

  • Notes
AI-assisted assessment Rubric-driven workflow
Platform operator Engineer Research-led platform operation

Fall 2024

Continue note assessment

Post-encounter note grading continues across another deployment cycle.

Assessment sources

  • Notes
AI-assisted assessment Rubric-driven workflow
Platform operator Engineer Research-led platform operation

Spring 2025

Add transcripts

Notes are joined by transcript-based assessment of communication and physical examination.

Assessment sources

  • Notes
  • Transcript
AI-assisted assessment Rubric-driven workflow
Platform operator Engineer Research-led platform operation

Fall 2025 Published study cohort

Bring three modalities together

Notes, native audio and video are coordinated in one rubric-driven workflow.

Assessment sources

  • Notes
  • Native audio
  • Video
AI-assisted assessment Rubric-driven workflow
Platform operator Engineer Research-led platform operation

Spring 2026

Move to educator operation

Educators operate the notes, native-audio and video workflow after brief training.

Assessment sources

  • Notes
  • Native audio
  • Video
AI-assisted assessment Rubric-driven workflow
Platform operator Physician Educator-led platform operation

In the published study, each rubric item was scored using its designated source. Platform operation is distinct from the faculty’s assessment and oversight roles.

Governance across deployment cycles

A loop of human oversight

  1. Define the assessment protocol and supporting evidence.

  2. The supervisory committee reviews readiness and proposed changes.

  3. Run the approved workflow and monitor processing and data quality.

  4. Conduct human review, adjudication and quality checks.

  5. Bring operational, agreement and adjudication findings back to the committee.

  6. Use retrospective analysis to inform the next proposed protocol.

Faculty retain authority over learner review and educational decisions.

An original interactive explanation of the UTSW deployment history and governance process reported by Ngo, Shakur et al. (2026).

View original published Figure 1The source timeline and oversight cycle, reproduced unchanged.
Figure 1. Panel a shows notes in Fall 2023 and 2024, added transcripts in Spring 2025, and notes, native audio, and video in Fall 2025 and Spring 2026; the operator changes from engineer to physician in Spring 2026. Panel b cycles through proposed protocol, supervisory committee sign-off, prospective deployment, human review and adjudication, reporting, and retrospective analysis.

Panel a · The deployment timeline The progression from written notes to transcript, native-audio and video assessment.

Panel b · The oversight cycle Protocol design, committee sign-off, deployment, human review, reporting and retrospective analysis.

The figure places the Fall 2025 study within the program’s wider deployment and oversight timeline.

Figure 1 from Ngo, Shakur et al., npj Digital Medicine (2026), reproduced unchanged. © The Author(s) 2026. CC BY-NC-ND 4.0. View the full-size published figure

See the workflow

Ameer Hamza Shakur walks through rubric upload, AI-assisted scoring, evidence review and export of faculty-reviewed results. The demonstration shows the platform workflow; the paper reports the prospective study.

Watch the MAPLES walkthrough

Explore a validation partnership

How well does this approach transfer to another institution’s learners, stations and rubrics? Explore a focused pilot, a human-review workflow or an external-validation study with the team.

Start a research conversation
Shareable study summary Open the visual overview for colleagues and educators.
Visual summary of the prospective UTSW MAPLES study. The setting, grading-pass comparison and limitations are described in the study overview above.
A plain-language summary of the published UTSW study and its human-review workflow. Open the full-size summary graphic.
Back to Blog

Related Posts

View All Posts »
Wayfinder: AI-Assisted Rubric Authoring in MAPLES

Wayfinder: AI-Assisted Rubric Authoring in MAPLES

Watch Minhan Park and Licheng Yi demonstrate Wayfinder: create a structured rubric from a short description, enhance it with accept/reject suggestions, transform it for a new scenario, and continue in chat.