Accuracy of a Commercial Large Language Model (ChatGPT) to Perform Disaster Triage of Simulated Patients Using the Simple Triage and Rapid Treatment (START) Protocol: Gage Repeatability and Reproducibility Study
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.
Bibliographic record
Abstract
BACKGROUND: The release of ChatGPT (OpenAI) in November 2022 drastically reduced the barrier to using artificial intelligence by allowing a simple web-based text interface to a large language model (LLM). One use case where ChatGPT could be useful is in triaging patients at the site of a disaster using the Simple Triage and Rapid Treatment (START) protocol. However, LLMs experience several common errors including hallucinations (also called confabulations) and prompt dependency. OBJECTIVE: This study addresses the research problem: "Can ChatGPT adequately triage simulated disaster patients using the START protocol?" by measuring three outcomes: repeatability, reproducibility, and accuracy. METHODS: Nine prompts were developed by 5 disaster medicine physicians. A Python script queried ChatGPT Version 4 for each prompt combined with 391 validated simulated patient vignettes. Ten repetitions of each combination were performed for a total of 35,190 simulated triages. A reference standard START triage code for each simulated case was assigned by 2 disaster medicine specialists (JMF and MV), with a third specialist (LC) added if the first two did not agree. Results were evaluated using a gage repeatability and reproducibility study (gage R and R). Repeatability was defined as variation due to repeated use of the same prompt. Reproducibility was defined as variation due to the use of different prompts on the same patient vignette. Accuracy was defined as agreement with the reference standard. RESULTS: Although 35,102 (99.7%) queries returned a valid START score, there was considerable variability. Repeatability (use of the same prompt repeatedly) was 14% of the overall variation. Reproducibility (use of different prompts) was 4.1% of the overall variation. The accuracy of ChatGPT for START was 63.9% with a 32.9% overtriage rate and a 3.1% undertriage rate. Accuracy varied by prompt with a maximum of 71.8% and a minimum of 46.7%. CONCLUSIONS: This study indicates that ChatGPT version 4 is insufficient to triage simulated disaster patients via the START protocol. It demonstrated suboptimal repeatability and reproducibility. The overall accuracy of triage was only 63.9%. Health care professionals are advised to exercise caution while using commercial LLMs for vital medical determinations, given that these tools may commonly produce inaccurate data, colloquially referred to as hallucinations or confabulations. Artificial intelligence-guided tools should undergo rigorous statistical evaluation-using methods such as gage R and R-before implementation into clinical settings.
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.
Direct model labels (unvalidated)
Per-model category and study-design labels from the labeling rounds. They are machine output, unvalidated, and the disagreement between models ships as data. No study design here is MEDLINE-validated yet.
| Model arm | Categories | Study design | Confidence |
|---|---|---|---|
| gemma | Metaresearch Domain: Reproducibility · Genre: Empirical About the Canadian research system: no · About a Canadian topic: no | Simulation or modeling | low |
| gpt | no category Domain: not available · Genre: Empirical About the Canadian research system: no · About a Canadian topic: no | Bench or experimental | low |
Full frame distilled prediction
Teacher imitationNot 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.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.031 | 0.010 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.001 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
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.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it