Assessing the Limitations of Large Language Models in Clinical Practice Guideline–Concordant Treatment Decision-Making on Real-World Data: Retrospective Study
Bibliographic record
Abstract
Background: Studies have shown that large language models (LLMs) are promising in therapeutic decision-making, with findings comparable to those of medical experts, but these studies used highly curated patient data. Objective: This study aimed to determine if LLMs can make guideline-concordant treatment decisions based on patient data as typically present in clinical practice (lengthy, unstructured medical text). Methods: We conducted a retrospective study of 80 patients with severe aortic stenosis who were scheduled for either surgical (SAVR; n=24) or transcatheter aortic valve replacement (TAVR; n=56) by our institutional heart team in 2022. Various LLMs (BioGPT, GPT-3.5, GPT-4, GPT-4 Turbo, GPT-4o, LLaMA-2, Mistral, PaLM 2, and DeepSeek-R1) were queried using either anonymized original medical reports or manually generated case summaries to determine the most guideline-concordant treatment. We measured agreement with the heart team using Cohen κ coefficients, reliability using intraclass correlation coefficients (ICCs), and fairness using the frequency bias index (FBI; FBI >1 indicated bias toward TAVR). Results: When presented with original medical reports, LLMs showed poor performance (Cohen κ coefficient: -0.47 to 0.22; ICC: 0.0-1.0; FBI: 0.95-1.51). The LLMs' performance improved substantially when case summaries were used as input and additional guideline knowledge was added to the prompt (Cohen κ coefficient: -0.02 to 0.63; ICC: 0.01-1.0; FBI: 0.46-1.23). Qualitative analysis revealed instances of hallucinations in all LLMs tested. Conclusions: Even advanced LLMs require extensively curated input for informed treatment decisions. Unreliable responses, bias, and hallucinations pose significant health risks and highlight the need for caution in applying LLMs to real-world clinical decision-making.
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 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 | 0.003 | 0.011 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 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 itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
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".