Wednesday, October 26, 2016
Bayshore Ballroom ABC, Lobby Level (Westin Bayshore Vancouver)
Poster Board # PS 4-29

Muge Capan, PhD1, Nisha Nataraj, PhD Student2, Liana Lin, PhD Student2, Min Chi, PhD2, Julie Ivy, PhD2 and Ryan Arnold, MD1, (1)Christiana Care Health System, Newark, DE, (2)North Carolina State University, Raleigh, NC
Purpose: This research aims at developing data-driven models to classify inpatients into sepsis presentation categories based on clinical observations, and determine the impact of comorbidities and pre-existing conditions on sepsis presentation.

Method: We used retrospective data from the in- and outpatient units of a hospital containing 77,975 patient visits. Patients’ sepsis presentations were analyzed using Hidden Markov Models (HMM) and Input-Output Hidden Markov Models (IOHMM). Natural Language Processing was used to combine observed measurements and providers’ documentation, e.g., progress notes. We used area under the curve to measure the discriminative performance, Akaike Information Criterion and Bayesian Information Criterion to evaluate the models’ fit.

Result: Descriptive analysis showed that 6,109 patient visits had a clinical diagnosis of sepsis as defined by the International Classification of Diseases, Ninth Revision (ICD9) codes. Chi-square tests found significant association between sepsis presence and primary malignancy, malignancy with metastases, urinary tract infections (p < 0.0001) and history of depression, diabetes, paralysis, peripheral vascular disease, among others (p < 0.0001). The HMM used five input variables: 1) Sepsis at admission probability (i.e., probability that the patient presents a sepsis phenotype at the time of admission), 2) Deterioration rate (i.e., rate of in-hospital deterioration resulting in a sepsis phenotype), 3) Recovery probability (i.e., probability that the patient recovers from sepsis presentation), 4) False positive rate (i.e., probability that the patient does not have sepsis, but expresses sepsis-related clinical symptoms) and 5) False negative rate (i.e., probability that the patient has sepsis, but does not express any sepsis-related clinical symptoms). The IOHMM extended the HMM where the sepsis at admission probability, deterioration rate, false positive and false negative rates were dependent on the therapeutic interventions, such as, antibiotic administration.

Conclusion: Sepsis is a complex condition identified as a major healthcare challenge. The lack of a gold standard diagnostic test results in inconsistencies in categorizing sepsis phenotypes. Understanding the impact of comorbidities and pre-existing conditions on sepsis presentation can inform categorization of sepsis phenotypes. We found that several pre-existing conditions and comorbidities are associated with presence of sepsis, indicating the need to examine comorbidities and health conditions when diagnosing and treating sepsis. Additionally, the IOHMM highlighted the importance of adaptive therapeutic interventions informed by the dynamic changes of sepsis presentation during a hospital visit.