Language models reflect clinical evidence but fail to adapt it to patients
He, S.; Joseph, J. W.; Safari, P.; Goff, A.; Slusarz, P. J.; Mohamed, A.; Lord, S.; Goldstein, J. N.; Raja, A. S.
Show abstract
Background: Clinical language models must use evidence to produce the number and action required by a particular case. Whether numerical knowledge reliably becomes a correct clinical response is unclear.`n`nMethods: NUMBERS evaluated 16 model configurations on 1,300 questions linked to public clinical evidence. Linked experiments tested prevalence updating, patient-specific estimates and clinical actions. The direct-action experiment compared cutoff recall plus action selection with a supplied complete rule across 50 rules, five models and 7,500 calls.`n`nResults: Among diagnostic estimates outside the source-result tolerance, 77.1% remained within the evidence's central 80% predictive range. Models updated prevalence-dependent quantities correctly in 78.5% of comparisons with inputs and a calculation request, versus 33.7% with clinical wording. Supplying inputs and requesting calculation raised patient-level near-target answers from 29.7% to 78.8% across 2,340 pairs. Models selected an incorrect action despite stating a cutoff that implied the correct action in 413 of 3,750 recall-arm calls (11.01%; 95% confidence interval, 8.93 to 13.17). Supplying the complete rule raised action accuracy from 85.63% to 99.41%, an improvement of 13.79 percentage points (95% confidence interval, 11.65 to 15.95).`n`nConclusions: Models often produced evidence-consistent numbers but failed to adapt them to a case or act consistently with their own stated cutoff. Explicit inputs, calculations and complete rules substantially improved performance in controlled prompts.`n`nFunding: National Academy of Medicine, Agreement No. 2026A008797.
Matching journals
The top 3 journals account for 50% of the predicted probability mass.
Similar papers in this journal
- Beyond Metrics to Methods: A Scoping Review of Large Language Models for Detection of Social Drivers of Health in Clinical Notes 91%
- Collaborative Large Language Models for Automated Data Extraction in Living Systematic Reviews 90%
- Generative Large Language Models in Electronic Health Records for Patient Care Since 2023: A Systematic Review 90%
Similar papers in this journal
- From Tool to Teammate: A Randomized Controlled Trial of Clinician-AI Collaborative Workflows for Diagnosis 93%
- Interpretable Fine-tuned Large Language Models Facilitate Making Genetic Test Decisions for Rare Diseases 90%
- A Framework to Assess Clinical Safety and Hallucination Rates of LLMs for Medical Text Summarisation 90%
Similar papers in this journal
- irAE-GPT: Leveraging large language models to identify immune-related adverse events in electronic health records and clinical trial datasets 91%
- Consistent Performance of GPT-4o in Rare Disease Diagnosis Across Nine Languages and 4967 Cases 89%
- Transformer-based deep learning model for the diagnosis of suspected lung cancer in primary care based on electronic health record data 87%
Similar papers in this journal
- Low adherence to existing model reporting guidelines by commonly used clinical prediction models 91%
- Diagnostic Codes in AI prediction models and Label Leakage of Same-admission Clinical Outcomes 88%
- COVID-19 outcomes, risk factors and associations by race: a comprehensive analysis using electronic health records data in Michigan Medicine 87%
Similar papers in this journal
- Re-use of trial data in the first 10 years of the data-sharing policy of the Annals of Internal Medicine: a survey of published studies 89%
- Large-scale validation of the Prediction model Risk Of Bias ASsessment Tool (PROBAST) using a short form: high risk of bias models show poorer discrimination 89%
- Protocol for an observational study evaluating new approaches to modelling diagnostic information from large administrative hospital datasets 88%
"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.