MétaCan
Menu
Back to cohort
Record W4319454867 · doi:10.1101/2023.02.04.23285478

ChatGPT for Clinical Vignette Generation, Revision, and Evaluation

2023· preprint· en· W4319454867 on OpenAlex

Why this work is in the frame

A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.

affAt least one author lists a Canadian institution in the pinned OpenAlex snapshot.

Bibliographic record

VenuemedRxiv · 2023
Typepreprint
Languageen
FieldMedicine
TopicClinical Reasoning and Diagnostic Skills
Canadian institutionsUniversity of Alberta HospitalUniversity of Alberta
Fundersnot available
KeywordsVignetteRewritingTriageSet (abstract data type)MedicinePerspective (graphical)LiteracyPsychologyPsychiatryArtificial intelligenceComputer scienceSocial psychologyProgramming language

Abstract

fetched live from OpenAlex

Abstract Objective To determine the capabilities of ChatGPT for rapidly generating, rewriting, and evaluating (via diagnostic and triage accuracy) sets of clinical vignettes. Design We explored the capabilities of ChatGPT for generating and rewriting vignettes. First, we gave it natural language prompts to generate 10 new sets of 10 vignettes, each set for a different common childhood illness. Next, we had it generate 10 sets of 10 vignettes given a set of symptoms from which to draw. We then had it rewrite 15 existing pediatric vignettes at different levels of health literacy. Fourth, we asked it to generate 10 vignettes written as a parent, and rewrite these vignettes as a physician, then at a grade 8 reading level, before rewriting them from the original parent’s perspective. Finally, we evaluated ChatGPT for diagnosis and triage for 45 clinical vignettes previously used for evaluating symptom checkers. Setting and participants ChatGPT, a publicly available, free chatbot. Main outcome measures Our main outcomes for de novo vignette generation were whether ChatGPT followed vignette creation instructions consistently, correctly, and listed reasonable symptoms for the disease being described. For generating vignettes from pre-existing symptom sets, we examined whether the symptom sets were used without introducing extra symptoms. Our main outcome for rewriting existing standardized vignettes to match patient demographics, and rewriting vignettes between styles, was whether symptoms were dropped or added outside the original vignette. Finally, our main outcomes examining diagnostic and triage accuracy on 45 standardized patient vignettes were whether the correct diagnosis was listed first, and if the correct triage recommendation was made. Results ChatGPT was able to quickly produce varied contexts and symptom profiles when writing vignettes based on an illness name, but overused some core disease symptoms. It was able to use given symptom lists as the basis for vignettes consistently, adding one additional (though appropriate) symptom from outside the list for one disease. Pediatric vignettes rewritten at different levels of health literacy showed more complex symptoms being dropped when writing at low health literacy in 87.5% of cases. While writing at high health literacy, it added a diagnosis to 80% of vignettes (91.7% correctly diagnosed). Symptoms were retained in 90% of cases when rewriting vignettes between viewpoints. When presented with 45 vignettes, ChatGPT identified illnesses with 75.6% (95% CI, 62.6% to 88.5%) first-pass diagnostic accuracy and 57.8% (95% CI, 42.9% to 72.7%) triage accuracy. Its use does require monitoring and has caveats, which we discuss. Conclusions ChatGPT was capable, with caveats and appropriate review, of generating, rewriting, and evaluating clinical vignettes.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

Full frame distilled prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.

metaresearch head score (Codex)0.006
metaresearch head score (Gemma)0.158
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesMetaresearch
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Observational · Consensus signal: Observational
GenreCandidate signal: Empirical · Consensus signal: Empirical
Teacher disagreement score0.405
Threshold uncertainty score0.849

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0060.158
Meta-epidemiology (narrow)0.0000.000
Meta-epidemiology (broad)0.0010.000
Bibliometrics0.0000.000
Science and technology studies0.0000.000
Scholarly communication0.0000.000
Open science0.0000.000
Research integrity0.0000.000
Insufficient payload (model declined to judge)0.0000.000

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.254
GPT teacher head0.492
Teacher spread0.238 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it