MétaCan
Menu
Back to cohort

DeepSeek-R1 vs OpenAI o1 for Ophthalmic Diagnoses and Management Plans

2025· article· en· W4413982530 on OpenAlex
David Mikhail, Andrew Farah, Jason Milad, Andrew Mihalache, Daniel Milad, Fares Antaki, Michael Balas, Marko M. Popovic, Rajeev H. Muni, Pearse A. Keane, Renaud Duval

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

VenueJAMA Ophthalmology · 2025
Typearticle
Languageen
FieldMedicine
TopicArtificial Intelligence in Healthcare and Education
Canadian institutionsUniversité de MontréalSt. Michael's HospitalHôpital Maisonneuve-RosemontCentre Hospitalier de l’Université de MontréalMcGill UniversityUniversity of WaterlooUniversity of Toronto
Fundersnot available
KeywordsMcNemar's testMedicineMedical diagnosisClinical PracticeMedical physicsFamily medicinePathologyStatistics

Abstract

fetched live from OpenAlex

Importance: Large language models (LLMs) are increasingly being explored in clinical decision-making, but few studies have evaluated their performance on complex ophthalmology cases from clinical practice settings. Understanding whether open-weight, reasoning-enhanced LLMs can outperform proprietary models has implications for clinical utility and accessibility. Objective: To evaluate the diagnostic accuracy, management decision-making, and cost of DeepSeek-R1 vs OpenAI o1 across diverse ophthalmic subspecialties. Design, Setting, and Participants: This was a cross-sectional evaluation conducted using standardized prompts and model configurations. Clinical cases were sourced from JAMA Ophthalmology's Clinical Challenge articles, containing complex cases from clinical practice settings. Each case included an open-ended diagnostic question and a multiple-choice next-step decision. All cases were included without exclusions, and no human participants were involved. Data were analyzed from March 13 to March 30, 2025. Exposures: DeepSeek-R1and OpenAI o1 were evaluated using the Plan-and-Solve Plus (PS+) prompt engineering method. Main Outcomes and Measures: Primary outcomes were diagnostic accuracy and next-step decision-making accuracy, defined as the proportion of correct responses. Token cost analyses were performed to estimate expenses. Intermodel agreement was evaluated using Cohen κ, and McNemar test was used to compare performance. Results: A total of 422 clinical cases were included, spanning 10 subspecialties. DeepSeek-R1 achieved a higher diagnostic accuracy of 70.4% (297 of 422 cases) compared with 63.0% (266 of 422 cases) for OpenAI o1, a 7.3% difference (95% CI, 1.0%-13.7%; P = .02). For next-step decisions, DeepSeek-R1 was correct in 82.7% of cases (349 of 422 cases) vs OpenAI o1's accuracy of 75.8% (320 of 422 cases), a 6.9% difference (95% CI, 1.4%-12.3%; P = .01). Intermodel agreement was moderate (κ = 0.422; 95% CI, 0.375-0.469; P < .001). DeepSeek-R1 offered lower costs per query than OpenAI o1, with savings exceeding 66-fold (up to 98.5%) during off-peak pricing. Conclusions and Relevance: DeepSeek-R1 outperformed OpenAI o1 in diagnosis and management across subspecialties while lowering operating costs, supporting the potential of open-weight, reinforcement learning-augmented LLMs as scalable and cost-saving tools for clinical decision support. Further investigations should evaluate safety guardrails and assess performance of self-hosted adaptations of DeepSeek-R1 with domain-specific ophthalmic expertise to optimize clinical utility.

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)0.000
metaresearch head score (Gemma)0.000
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesnone
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Observational · Consensus signal: Observational
GenreCandidate signal: Empirical · Consensus signal: Empirical
Teacher disagreement score0.220
Threshold uncertainty score0.547

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0000.000
Meta-epidemiology (narrow)0.0000.000
Meta-epidemiology (broad)0.0000.000
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 head0.110
GPT teacher head0.448
Teacher spread0.338 · 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