Improving sepsis care with AI-based clinical decision support

Project Summary Sepsis is a syndrome characterizied by a dysregulated host immune response to an infection that leads to organ dysfunction. Because sepsis is among the leading causes of death among hospitalized patients and accounts for substantial harms, costs, and loss of quality of life, many efforts have been made to improve

sepsis care. The mainstay of treatment is timely recognition and prompt initiation of broad-spectrum an- timicrobial therapy. However, identification of sepsis is fraught with uncertainty in busy and complex clinical environments and treatment delays are common in the emergency department, hospital ward, and intensive

care unit. As a result, the use of artificial intelligence (AI) and machine learning (ML) methods to provide timely clinical decision support (CDS) has good face validity to improve care. Despite hundreds of published papers on predictive sepsis systems, there is very little evidence that such systems actually improve care

processes or patient outcomes. Therefore, this proposal outlines three important knowledge gaps at the in- tersection of AI/ML methods and clinical care that have so far hindered the development of successful sepsis CDS systems. First, the optimal outcome (i.e., training label) on which to develop predictive systems for sep-

sis is unknown. Current sepsis definitions were designed primarily to standardize clinical trial enrollment and epidemiologic surveillance rather than to support bedside treatment decisions. Second, although sepsis is currently defined by changes in organ function from baseline, the optimal approach to capture time-varying

changes in clinical parameters remains unknown. Many AI/ML methods are uniquely suited to learning such important patterns in the data but their use in predicting sepsis remains under-explored. Third, there are significant differences in patient outcomes and clinical presentation between community- and hospital-onset

sepsis. However, how these differences might affect predictive accuracy, estimates of variable importance, timing and use of predictive alerts, among other important considerations for CDS development, remains un- known. Thus, this proposal seeks to answer these fundamental questions to overcome key knowledge gaps

and realize the promise of AI/ML methods for improving sepsis care. Broadly speaking, we will consider sev- eral state-of-the-art approaches to answer these questions, including the use of i) informatics methods such as active learning to facilitate efficient and large-scale clinician review of patient data, ii) advanced causal

inference methods such as target trial emulation to compare the clinical effects of treatment according to different sepsis definitions, and iii) AI/ML methods such as convolutional neural networks and denoising autoencoders to determine the optimal representations of complex and time-varying clinical features. An-

swering these questions will pave the way for the development of AI/ML CDS systems that are relevant at the bedside, scalable across a diverse range of care contexts and patient populations, and most importantly that improve clinical care and patient outcomes.