SCH: INT: Collaborative Research: S.E.P.S.I.S.: Sepsis Early Prediction Support Implementation System

Sepsis, infection plus systemic manifestations of infection, is the leading cause of in-hospital mortality. About 700,000 people die annually in US hospitals and 16% of them were diagnosed with sepsis (including a high prevalence of severe sepsis with major complication). In addition to being deadly, sepsis is the most expensive condition associated with in-hospital stay, resulting in a 75% longer stay than any other condition. The total burden of sepsis to the US healthcare system is estimated to be $20.3 billion, most of which is paid by Medicare and Medicaid. In fact, in June 2015 the Centers for Medicare & Medicaid Services (CMS) reported that sepsis accounted for over $7 billion in Medicare payments (second only to major joint replacement), a close to 10% increase from the previous year. This pervasive drain on health care resources is due, in part, to difficulties in diagnosis and delayed treatment. For example, every one hour delay in treatment of severe sepsis/shock with antibiotics decreases a patient's survival probability by 10%. Many of these deaths could have been averted or postponed if a better system of care was in place. The goal of this research is to overcome these barriers by integrating electronic health records (EHR) and clinical expertise to provide an evidence-based framework to diagnose and accurately risk-stratify patients within the sepsis spectrum, and develop and validate intervention policies that inform sepsis treatment decisions. The project to bring together health care providers, researchers, educators, and students to add value to patient care by integrating machine learning, decision analytical models, human factors analysis, as well as system and process modeling to advance scientific knowledge, predict sepsis, and prevent sepsis-related health deterioration. In addition to the societal impact that clinical translation of these findings may bring, the project will provide engineering and computer science students and health services researchers with cross-disciplinary educational experience.

The proposed research will apply engineering and computer science methodologies to analyze patient level EHR across two large scale health care facilities, Mayo Clinic Rochester and Christiana Care Health System and to inform clinical decision making for sepsis. The multi-institutional, interdisciplinary collaboration will enable the development of health care solutions for sepsis by describing and accurately risk-stratifying hospitalized patients, and developing decision analytical models to personalize and inform diagnostic and treatment decisions considering patient outcomes and response implications. The Sepsis Early Prediction Support Implementation System (S.E.P.S.I.S.) project aims will be to: 1) Develop data-driven models to classify patients according to their clinical progression to diagnose sepsis and predict risk of deterioration, thus informing therapeutic actions. 2) Develop personalized intervention policies for patients within the sepsis spectrum. 3) Develop decision support systems (DSS) for personalized interventions focusing on resource implications and usability within a real hospital setting. The team will 1) identify important factors that uncover patient profiles based on Bayesian exponential family principal components analysis; 2) develop hidden Markov models (HMMs) and input-output HMMs to identify clusters of patients with similar progression patterns within the sepsis spectrum; 3) provide an analytical framework to support sepsis staging in clinical practice using bilevel optimization. They will 1) predict short- and long-term individual patient outcomes using multivariate statistical models and simulation; 2) develop semi-Markov decision process and partially observable semi-Markov decision process models to identify timing of therapeutic actions and diagnostic tests. Furthermore, the team will 1) predict demand for resources and develop and validate a hybrid mixed integer programming and queueing model to optimize system level allocations; 2) utilize human factors analysis and usability testing to assess the implementation of the DSS.