Implementation and Evaluations of Previously Developed Novel Early Warning System to Detect and Treat Sepsis

Sepsis represents a significant burden to the healthcare system. National predictions estimate 751,000 cases of severe sepsis per annum which will increase at a rate of 1.5%. Sepsis accounts for >$23 billion in aggregate hospital costs across all payers and represents nearly 4% of all hospital stays. Six percent of all deaths in the US can be attributed to sepsis. Protocol driven care bundles improve clinical outcomes but require early and accurate detection of sepsis. Unfortunately, identifying sepsis early remains elusive even for experienced clinicians leading to diagnostic uncertainty.

To improve diagnostic consensus, a task force in 2016 agreed upon a new sepsis definition. The task force also included a new risk stratification tool to improve early identification, the quick Sepsis-related Organ Failure Assessment (qSOFA) model, which was more accurate than the older Systemic Inflammatory Response Syndrome (SIRS) in predicting adverse clinical outcomes. However, due to the reliance of end organ dysfunction, the new definition has been criticized for its detection of sepsis late in the clinical course. Clinical decision support tools based on predictive analytics can provide actionable information and improve diagnostic accuracy particularly in sepsis.

Several early warning tools have been described in the published literature based upon predictive analytics and large datasets. One example is the National Early Warning Score (NEWS), which was developed to discriminate patients at risk of cardiac arrest, unplanned intensive care admission, or death. Scores such as NEWS are typically broad in scope and not designed to specifically target sepsis. They are also conceptually simple, as they use only a small number of variables and compare them to normal ranges to generate a composite score. In assigning independent scores to each variable and using only the most recent value, they both ignore complex relationships between the variables and their evolution in time.

In previous work, our group developed a framework to model multivariate time series using multitask Gaussian processes, accounting for the high uncertainty, frequent missing values, and irregular sampling rates typically associated with real clinical data can be read in our prior work. Our machine learning approach is superior to other sepsis detection models that use traditional analytics and machine learning techniques. A custom web application, Sepsis Watch, presents the risk score along with relevant patient information and prompts the user to further evaluate the patient and begin treatment, if appropriate. The Sepsis Watch system is now being implemented by clinical operations at Duke University Hospital.

Our study employs a sequential roll-out study design in the Emergency Department at Duke University Hospital. Our study will involve pods A, B, C, and the Resuscitation Bay. The operational project is not being implemented on the psychiatry wing, fast track, triage or any inpatient encounters. The operational project and thus our study period is based upon a two-phase roll out: