Syndromic Analysis of Sepsis Cohorts Using Large Language Models
Authors' conclusion
This cohort study found that an LLM accurately extracted presenting signs and symptoms from admission notes that clustered into syndromes differentially correlated with infection sources, multidrug-resistant infections, and mortality.
Study context
- Population
- Hospitalised adults with possible infection (blood culture drawn and IV antibiotics within 24 hours of arrival), 5 Massachusetts hospitals, June 2015–August 2022
- Participants
- 104,248 adults
- Exposure
- Associations of LLM-extracted presenting signs and symptoms (syndromes) with infection sources, multidrug-resistant infections and mortality
- Outcome
- Positive cultures for MRSA, positive cultures for MDRGN organisms, and in-hospital mortality
Methodological detailsmulticentre, large language model extraction, unsupervised clustering, multivariable logistic regression, Data from 5 Massachusetts hospitals within 1 health care system, June 1, 2015–August 1, 2022
- Multicentre
- Large language model extraction
- Unsupervised clustering
- Multivariable logistic regression
- Data from 5 Massachusetts hospitals within 1 health care system, June 1, 2015–August 1, 2022
Main finding
As reported in the abstract
- LLM labels vs physician reviewer: accuracy
- 99.3% (95% CI, 99.2%-99.3%)
- LLM labels: balanced accuracy
- 84.6% (95% CI, 83.5%-85.8%)
- LLM labels: PPV
- 68.4% (95% CI, 66.0%-70.7%)
- LLM labels: sensitivity
- 69.7% (95% CI, 67.3%-72.0%)
- LLM labels: specificity
- 99.6% (95% CI, 99.6%-99.6%)
- Skin and soft tissue symptoms and methicillin-resistant Staphylococcus aureus (MRSA) culture positivity, aOR
- 1.73 (95% CI, 1.49-2.00)
- Gastrointestinal symptoms and MRSA culture positivity, aOR
- 0.63 (95% CI, 0.54-0.73)
- Urinary tract symptoms and MRSA culture positivity, aOR
- 0.34 (95% CI, 0.22-0.50)
- Cardiopulmonary symptoms and increased mortality, aOR
- 1.30 (95% CI, 1.17-1.45)
Observational data: associations, not effects.
Further findings
LLM labelling coverage
- Patients whose notes were labelled by the LLM
- 93 674 of 94 913 (98.7%)
Original publication
Syndromic Analysis of Sepsis Cohorts Using Large Language Models. JAMA Netw Open. Published 1 October 2025. doi:10.
Composed by AI from the published abstract and checked against it. How Briefs are made →