MétaCan
Menu
Back to cohort

Accuracy of an Artificial Intelligence Chatbot’s Interpretation of Clinical Ophthalmic Images

2024· article· en· W4394760737 on OpenAlex
Andrew Mihalache, Ryan S. Huang, Marko M. Popovic, Nikhil S. Patil, Bhadra U. Pandya, Reut Shor, Austin Pereira, Jason M. Kwok, Peng Yan, David T. Wong, Peter J. Kertes, Rajeev H. Muni

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.
aboutThe title or abstract carries a Canadian signal from the geographic lexicon.

Bibliographic record

VenueJAMA Ophthalmology · 2024
Typearticle
Languageen
FieldMedicine
TopicArtificial Intelligence in Healthcare and Education
Canadian institutionsSunnybrook Health Science CentreSt. Michael's HospitalMcMaster UniversityUniversity of Toronto
Fundersnot available
KeywordsChatbotMedicineArtificial intelligenceMedical physicsOphthalmologyOptometryComputer science

Abstract

fetched live from OpenAlex

Importance: Ophthalmology is reliant on effective interpretation of multimodal imaging to ensure diagnostic accuracy. The new ability of ChatGPT-4 (OpenAI) to interpret ophthalmic images has not yet been explored. Objective: To evaluate the performance of the novel release of an artificial intelligence chatbot that is capable of processing imaging data. Design, Setting, and Participants: This cross-sectional study used a publicly available dataset of ophthalmic cases from OCTCases, a medical education platform based out of the Department of Ophthalmology and Vision Sciences at the University of Toronto, with accompanying clinical multimodal imaging and multiple-choice questions. Across 137 available cases, 136 contained multiple-choice questions (99%). Exposures: The chatbot answered questions requiring multimodal input from October 16 to October 23, 2023. Main Outcomes and Measures: The primary outcome was the accuracy of the chatbot in answering multiple-choice questions pertaining to image recognition in ophthalmic cases, measured as the proportion of correct responses. χ2 Tests were conducted to compare the proportion of correct responses across different ophthalmic subspecialties. Results: A total of 429 multiple-choice questions from 136 ophthalmic cases and 448 images were included in the analysis. The chatbot answered 299 of multiple-choice questions correctly across all cases (70%). The chatbot's performance was better on retina questions than neuro-ophthalmology questions (77% vs 58%; difference = 18%; 95% CI, 7.5%-29.4%; χ21 = 11.4; P < .001). The chatbot achieved a better performance on nonimage-based questions compared with image-based questions (82% vs 65%; difference = 17%; 95% CI, 7.8%-25.1%; χ21 = 12.2; P < .001).The chatbot performed best on questions in the retina category (77% correct) and poorest in the neuro-ophthalmology category (58% correct). The chatbot demonstrated intermediate performance on questions from the ocular oncology (72% correct), pediatric ophthalmology (68% correct), uveitis (67% correct), and glaucoma (61% correct) categories. Conclusions and Relevance: In this study, the recent version of the chatbot accurately responded to approximately two-thirds of multiple-choice questions pertaining to ophthalmic cases based on imaging interpretation. The multimodal chatbot performed better on questions that did not rely on the interpretation of imaging modalities. As the use of multimodal chatbots becomes increasingly widespread, it is imperative to stress their appropriate integration within medical contexts.

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)0.001
metaresearch head score (Gemma)0.003
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesnone
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Other design · Consensus signal: none
GenreCandidate signal: Empirical · Consensus signal: Empirical
Teacher disagreement score0.972
Threshold uncertainty score0.831

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0010.003
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)0.0010.000

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 head0.257
GPT teacher head0.535
Teacher spread0.278 · 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