Sepsis is currently defined as a “life-threatening organ dysfunction due to a dysregulated host response to infection”, and has a very high mortality rate. Although early diagnosis and prompt interventions are associated with better outcomes, early clinical recognition is frequently impeded by lack of confirmed infection diagnosis and the poorly understood nature of the sepsis pathophysiological process. Consequently, clinicians commonly rely on inferences based on “indirect” clinical evidence of infection and signs of severe organ failure measured by rule-based scoring tools to assist in mortality risk recognition. The Sequential Organ Failure Assessment (SOFA) score has been proposed for the identification of hospitalized patients at risk of sepsis mortality. A “quick SOFA” (qSOFA) score was similarly derived and validated as a simple surrogate for organ dysfunction for use in prehospital settings.
In search of improved sepsis predictive performance, dynamic Bayesian networks (DBN) machine learning algorithms using temporal/longitudinal data have been proposed. DBNs have also been used to model the complex dynamics in organ failure in infected ICU patients and identify meaningful clinical patterns in multi-parameter EMR data streams that precede sepsis-induced clinical deterioration. Additionally, customized evidence-based ontological models that express clinical knowledge/reasoning towards semantic annotation of clinical data have been studied but to our knowledge not yet applied in the sepsis domain. These models could be used to semantically characterize EMR extracted data for enhanced natural language processing, rule-based detection, and machine learning predictive analytics.
In this context, we explore the predictive performance of semantically enhanced DBN compared to SOFA and qSOFA used in sepsis clinical decision support. We also assess comparative performance of our DBN against the Simplified Acute Physiological Score (SAPS-II) and the Modified Early Warning Score (MEWS) tools used to identify general hospitalized patients at risk of catastrophic deterioration.
The network model and features used to train the DBN were guided by sepsis domain hierarchical concept maps. The concept maps used for our study are “deep” sepsis-specific ontological “structured reasoning” models developed with the support of expert clinicians and published medical evidence. We used a supervised training algorithm and evaluated the model’s predictive performance based on ICU encounter data and outcomes from patients with infections recorded in the openly available Beth Israel Deaconess MIMIC-II database. We compared semantically enhanced DBN predictive performance with SOFA, qSOFA, SAPS-II, and MEWS scores and evaluated differences in discrimination/calibration and risk classification between these algorithms.
18,607 ICU stays from 15,092 patients were analyzed with 2,452 patients deceased at discharge. The area under receiver operating curve (AUROC) for predicting mortality from the DBN was found to be 0.91 compared to qSOFA (0.645), first SOFA score (0.756), maximum SOFA score (0.843), MEWS (0.716) and SAPS-II (0.754). Continuous Net Reclassification Index and Integrated Discrimination Improvement analysis supported the superiority of semantically enhanced DBN with respect to SOFA, qSOFA, MEWS, and SAPS-II.
Compared with automatically computed conventional severity scores, the semantically enhanced DBN algorithm offers improved performance for predicting mortality of infected patients in intensive care units.
DECISION SUPPORT & HOSPITAL MONITORING
Author: Tom Velez
Coauthor(s): Tony Wang, PhD, Tom Velez, PhD, Emilia Apostolova, PhD, Tim Tschampel, Thuy Ngo, MD, Joy Hardison, MD
Status: Completed Work
Funding Acknowledgment: NIH SBIR GRANT: 1R43LM012291-01 PROJECT TITLE: HYBRID ONTOLOGY AND MACHINE LEARNING BASED METHODS USING EMR DATA FOR EFFECTIVE CLINICAL DECISION SUPPORT (CDS); A SEPSIS CASE STUDY AWARDEE: COMPUTER TECHNOLOGY ASSOCIATES, INC PRINCIPAL INVESTIGATOR: TOM VELEZ, PHD