Early sepsis prediction in the emergency department using machine learning

Early recognition of sepsis is critical for timely intervention. Early intervention (eg. with antibiotics) can reduce the risk on organ injury and eventually death. Commonly used clinical scores, such as qSOFA and NEWS, have limited predictive performance to predict organ failure and mortality in patients with early sepsis. Moreover, they require to be estimated by a human operator and lack the capacity to use diverse input sources. Machine learning (ML) models may improve risk stratification by integrating a broader range of parameters.

This study is a post-hoc analysis of prospectively collected data from 1431 patients presenting to the ED with suspected sepsis. A total of 149 clinical parameters, which were routinely collected as part of clinical care during ED stay, were extracted from electronic health records and bedside measurements. Machine learning (ML) models were developed in 10-fold cross validation to predict organ injury and 30-day mortality. Model performance was validated afterwards and compared to qSOFA and NEWS.

The ML models achieved an AUROC (area under the receiver operator curve) of 0.798, 95 % CI [0.775, 0.821] for organ injury and 00.808, 95 % CI [0.777, 0.839] for 30-day mortality, outperforming qSOFA (AUROC 0.557-0.577) and NEWS (AUROC 0.576-0.610). Serum creatinine, SOFA score, ISAR-HP score, and clinical impression contributed to most to the model's score.

Machine learning models using routinely available ED data were associated with improved predictive performance of organ injury and 30-day mortality in early sepsis compared to conventional scoring systems, highlighting their potential for clinical decision support.