Learning from clinical health data for real-time decision support in emergency department care of sepsis

Abstract

The emergency department (ED) is the first point of contact with clinicians for most patients with acute illnesses. Early identification along with appropriate interventions (including procedures, medications, and triaging to an appropriate level of care) in the ED can be critical drivers of good outcomes, particularly in the care of patients with sepsis. Although sepsis is a leading cause of in-hospital mortality, it can be difficult to identify on presentation, and debate continues about the best practices in certain aspects of managing sepsis patients. In this thesis, we applied machine learning-based analyses to better understand the ED course of patients with sepsis and to build systems that can operate at the bedside to aid clinicians in the care of sepsis, including both detection of sepsis at the earliest possible stages and management of deteriorating cardiovascular function and hemodynamic status.

We extracted data using automated methods as well as manual chart review in a selection of two years' worth of ED visits to Massachusetts General Hospital. Clustering blood pressure trajectories showed that only 20% of 765 sepsis patients showed sustained responses to fluid bolus therapy, while 25% of patients requiring escalated hemodynamic support via vasopressor therapy had very low blood pressure for at least two hours before escalation from fluid to vasopressor administration. Subsequently, we showed that a simple logistic regression model with only six basic elements of patient data can distinguish between patients who required vasopressors and those whose hemodynamic function recovered with fluid therapy alone with area under the receiver operating characteristic curve (AUC) of 0.91 (95% CI: 0.88-0.94) at a final decision time.

A predictive version of the model could detect advance need for vasopressors within six hours with an AUC of 0.82 (95% CI: 0.80-0.83) and could retain performance in acutely hypotensive patients at an AUC of 0.77 (95% CI: 0.74-0.90). We also developed a model to detect the presence of sepsis at triage and throughout the ED stay, combining vital signs, presenting symptoms, and baseline risk factors to discriminate between 1,663 sepsis and non-sepsis acutely ill patients at triage with an AUC of 0.88 (95% CI: 0.86-0.90) and over the course of the whole ED stay with an AUC of 0.92 (95% CI: 0.91-0.94), improving significantly over existing sepsis screening tools such as qSOFA (triage AUC of 0.61). We designed these models to minimize user input needs so as to integrate into clinical workflows without extensive demands on clinicians interacting with the electronic medical record system or a bedside monitor.

These models provide a feasible way to build clinical decision support tools that can operate in real-time in the ED to improve sepsis care from the very first point of contact with a potential sepsis patient.