MétaCan
Menu
Back to cohort
Record W4414887759 · doi:10.2196/73941

Application of Large Language Models in Complex Clinical Cases: Cross-Sectional Evaluation Study

2025· article· en· W4414887759 on OpenAlexvenueno aff
Yuanheng Huang, Guozhen Yang, H D Chen, Weibin Wu, Yonghui Wu, Jiannan Xu, Jian Zhang

Bibliographic record

VenueJMIR Medical Informatics · 2025
Typearticle
Languageen
FieldMedicine
TopicArtificial Intelligence in Healthcare and Education
Canadian institutionsnot available
Fundersnot available
KeywordsBoosting (machine learning)Language modelProcess (computing)Modeling language

Abstract

fetched live from OpenAlex

Background: Large language models (LLMs) have made significant advancements in natural language processing (NLP) and are gradually showing potential for application in the medical field. However, LLMs still face challenges in medicine. Objective: This study aims to evaluate the efficiency, accuracy, and cost of LLMs in handling complex medical cases and to assess their potential and applicability as tools for clinical decision support. Methods: We selected cases from the database of the Department of Cardiothoracic Surgery, the Third Affiliated Hospital of Sun Yat-sen University (2021-2024), and conducted a multidimensional preliminary evaluation of the latest LLMs in clinical decision-making for complex cases. The evaluation included measuring the time taken for the LLMs to generate decision recommendations, Likert scores, and calculating decision costs to assess the execution efficiency, accuracy, and cost-effectiveness of the models. Results: A total of 80 complex cases were included in this study, and the performance of multiple LLMs in clinical decision-making was evaluated. Experts required 33.60 minutes on average (95% CI 32.57-34.63), far longer than any LLM. GPTo1 (0.71, 95% CI 0.67-0.74), GPT4o (0.88, 95% CI 0.83-0.92), and Deepseek (0.94, 95% CI 0.90-0.96) all finished under a minute without statistical differences. Although Kimi, Gemini, LLaMa3-8B, and LLaMa3-70B took 1.02-3.20 minutes, they were still faster than experts. In terms of decision accuracy, Deepseek-R1 had the highest accuracy (mean Likert score=4.19), with no significant difference compared to GPTo1 (P=.699), and both performed significantly better than GPT4o, Kimi, Gemini, LLaMa3-70B, and LLaMa3-8B (P<.001). Deepseek-R1 and GPTo1 demonstrated the lowest hallucination rates-6/80 (8%) and 5/80 (6%), respectively-significantly outperforming GPT-4o (7/80, 9%), Kimi (10/80, 12%), and the Gemini and LLaMa3 models, which exhibited substantially higher rates ranging from 13/80 (16%) to 25/80 (31%). Regarding decision costs, all LLMs showed significantly lower costs than the Multidisciplinary Team, with open-source models such as Deepseek-R1 offering a zero direct cost advantage. Conclusions: GPTo1 and Deepseek-R1 show strong clinical potential, boosting efficiency, maintaining accuracy, and reducing costs. GPT4o and Kimi performed moderately, indicating suitability for broader clinical tasks. Further research is needed to validate LLaMa3 series and Gemini in clinical decision.

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.

How this classification was reachedexpand

Full frame machine prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.

metaresearch head score (Codex)0.023
metaresearch head score (Gemma)0.071
Version: metacan-v3-hybrid-931329e0061cValidation 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 score0.023
Threshold uncertainty score0.120

Distilled classifier scores by category (both heads)

CategoryCodexGemma
Metaresearch0.0230.071
Meta-epidemiology (narrow)0.0010.001
Meta-epidemiology (broad)0.0010.002
Bibliometrics0.0020.001
Science and technology studies0.0010.001
Scholarly communication0.0010.002
Open science0.0010.001
Research integrity0.0010.001
Insufficient payload (model declined to judge)0.0020.001

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 head0.292
GPT teacher head0.597
Teacher spread0.305 · 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

Classification

machine, unvalidated

Machine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.

The models applied no category: nothing in the taxonomy fit this work.
Study designObservational
Domainnot available
GenreEmpirical

How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".

Quick stats

Citations3
Published2025
Admission routes1
Has abstractyes

Explore more

Same venueJMIR Medical InformaticsSame topicArtificial Intelligence in Healthcare and EducationFrench-language works237,207