Answers are grounded in this paper and cite their sources.
RATIONALE Sepsis is the most common cause of clinical deterioration in hospitalized adults. General early warning scores have been shown to outperform sepsis-specific tools for predicting clinical deterioration in septic and non-septic patients. However, it is not known which scores can most accurately predict sepsis itself. Therefore, we sought to compare the predictive accuracy of two general early warning scores (the National Early Warning Score (NEWS) and eCART, an artificial intelligence/machine-learning (AI/ML) score) with two sepsis-specific tools (the Epic Sepsis Model (ESM) and Systemic Inflammatory Response System (SIRS)) for identifying sepsis before administration of antibiotics. METHODS We conducted a retrospective analysis of consecutive adult ward encounters across 7 hospitals between 8/3/2018 and 11/9/2023. ESM scores were prospectively calculated while the others were retrospectively calculated and time matched to the nearest ESM score. The primary outcome was sepsis, defined as receiving ≥4 consecutive days of antibiotics within 48 hours of a blood culture and organ dysfunction, occurring within 6 hours of a score. Observation level area under the receiver operator characteristics curves (AUROCs) were calculated and compared using 95% CIs. Encounter level efficiency curves were constructed to graph the percentage of the total population that would have ever triggered as a function of the sensitivity of the tool to ever trigger in septic patients. Sepsis encounters were censored after initiation of antibiotics. RESULTS The study included 417,031 encounters, of which 8,331 (2.0%) first met sepsis criteria on a med-surg ward. eCART had the highest AUC at the observation level (0.901 (CI 0.898, 0.905)), followed by ESM (0.783 (CI 0.778, 0.788)), NEWS (CI 0.722 (0.717, 0.727)) and SIRS (0.704 (CI 0.698, 0.708)) for predicting sepsis qualifying antibiotic initiation within 6 hours. At the encounter level, eCART demonstrated the highest efficiency (Figure). Matching thresholds on sensitivities in the 62.4%-64.7% range, SIRS≥2, ESM≥4, and NEWS≥5 would have triggered on a similar percentage of encounters (29.1%-32.8%), while eCART≥94 would have triggered on 18.6%, resulting in 59,350 fewer false positives compared to SIRS. CONCLUSION In head-to-head comparison, an AI/ML general early warning score outperformed two sepsis specific tools, resulting in fewer alerts to detect the same number of events. This study adds to the growing body of literature that general early warning scores may be more efficient than sepsis-specific tools in predicting both sepsis and deterioration in the hospital and supports the use of a single screening tool for both indications.
No takes yet. Share an insight, caveat, or question.
Edelson et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: