Abstract
Background: Sepsis is a life-threatening condition requiring timely intervention. Machine learning (ML) models have shown promise for early detection, but their performance in emergency department (ED) triage settings remains understudied. This study aimed to develop and validate an ML-based sepsis prediction model using routinely collected triage data.Methods: We conducted a retrospective cohort study using electronic health records from a tertiary care ED between January 2020 and December 2022. Adult patients (≥18 years) with suspected infection were included. The primary outcome was sepsis within 24 hours of triage, defined per Sepsis-3 criteria. We trained four ML algorithms (logistic regression, random forest, gradient boosting, and neural network) on 70% of the data (n=12,845) and tested on 30% (n=5,506). Model performance was assessed using area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, and positive predictive value.Results: The gradient boosting model achieved the highest AUROC of 0.89 (95% CI: 0.87–0.91) on the test set, with sensitivity of 0.82 and specificity of 0.84. Key predictors included heart rate, respiratory rate, temperature, white blood cell count, and lactate level. The model outperformed the systemic inflammatory response syndrome (SIRS) criteria (AUROC 0.68) and qSOFA (AUROC 0.71). In a simulated deployment, the ML alert would have triggered a median of 2.3 hours earlier than clinical recognition.Conclusions: ML models using triage data can accurately predict sepsis within 24 hours, potentially enabling earlier intervention. Prospective validation and implementation studies are warranted to assess clinical impact and workflow integration.