Multiscale network representation of physiological time series for early prediction of sepsis

• Authors: Supreeth P. Shashikumar, Georgia Institute of Technology; Gari Clifford, Emory University; Qiao Li, Emory University; Shamim Nemati, Emory University
• Journal Title: Physiological Measurement
• Volume: Volume 38, Number 12
• Publisher: IOP Publishing: Hybrid Open Access | 2017-12-01, Pages 2235-2248
• Type of Work: Article | Post-print: After Peer Review
• Publisher DOI: http://doi.org/10.1088/1361-6579/aa9772
• Subjects: Engineering, Biomedical
• Keywords: Science & Technology; Life Sciences & Biomedicine; Technology; Biophysics; Engineering, Biomedical; Physiology; Engineering; sepsis; multiscale network; blood pressure; predictive analytics; network physiology; intensive care; critical care; INTENSIVE-CARE-UNIT; VARIABILITY; DYNAMICS; HEALTH

Abstract

Objective and Approach: Sepsis, a dysregulated immune-mediated host response to infection, is the leading cause of morbidity and mortality in critically ill patients. Indices of heart rate variability and complexity (such as entropy) have been proposed as surrogate markers of neuro-immune system dysregulation with diseases such as sepsis. However, these indices only provide an average, one dimensional description of complex neuro-physiological interactions. We propose a novel multiscale network construction and analysis method for multivariate physiological time series, and demonstrate its utility for early prediction of sepsis. Main results: We show that features derived from a multiscale heart rate and blood pressure time series network provide approximately 20% improvement in the area under the receiver operating characteristic (AUROC) for four-hour advance prediction of sepsis over traditional indices of heart rate entropy ( versus ). Our results indicate that this improvement is attributable to both the improved network construction method proposed here, as well as the information embedded in the higher order interaction of heart rate and blood pressure time series dynamics. Our final model, which included the most commonly available clinical measurements in patients' electronic medical records and multiscale entropy features, as well as the proposed network-based features, achieved an AUROC of . Significance: Prediction of the onset of sepsis prior to clinical recognition will allow for meaningful earlier interventions (e.g. antibiotic and fluid administration), which have the potential to decrease sepsis-related morbidity, mortality and healthcare costs.