Answers are grounded in this paper and cite their sources.
Retrospective cohort study evaluates machine learning for diagnosing cardiogenic shock in patients without hypotension, suggesting promising results.
Introduction Cardiogenic shock (CS) is challenging to identify, particularly in patients without hypotension. Retrospective identification of CS is important for accurate documentation and coding, with implications for national quality metrics and research. Methods We developed a retrospective cohort of patients admitted to the XXX cardiac intensive care unit between January 1, 2015 and December 31, 2024. Patients were manually reviewed by a cardiologist to determine whether CS without hypotension was present. 20% of cases were dual-reviewed by a second cardiologist. Using clinical notes from each hospitalization, a LLM was queried to determine whether each patient met criteria for CS without hypotension. Reliability and discriminative testing were performed to evaluate LLM performance against cardiologist-adjudicated labels. Results 551 patients met our inclusion criteria, of which 249 patients (45.2%) had CS without hypotension and 302 patients (54.8%) did not. Inter-cardiologist agreement was very high (Κ=0.925). The LLM demonstrated high internal reliability (r=0.959) and discrimination (F1 score 0.880 [95% CI: 0.849-0.909], AUROC 0.890 [0.866-0.916)], sensitivity 0.916 [0.879-0.948], specificity 0.864 [0.827-0.905]). LLM-cardiologist agreement was high (Κ=0.774) but did not achieve inter-cardiologist levels of agreement. Discussion The LLM achieved high reliability and discriminative accuracy in identifying CS without hypotension. Further research is needed to determine performance for other forms of CS, in alternative clinical settings, and for prospective CS prediction.
No takes yet. Share an insight, caveat, or question.
Ketcham et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: