Temporal Features and Kernel Methods for Predicting Sepsis in Postoperative Patients

Objective: Sepsis represents a major factor in morbidity and mortality in postoperative patients. The systemic inflammatory response syndrome (SIRS) criteria are binary statistics used to identify patients with sepsis, and are based on four physiological variables: body temperature, heart rate, breathing rate, and white blood cell count. However, the SIRS criteria have been criticized for having reduced specificity (high false positive rate), which diminishes their utility in clinical settings. This paper presents new features derived from the same four variables, and a methodology for predicting sepsis in postoperative patients under moderate care. Methods and material: Data for 1213 sepsis and 26 non-sepsis patients are obtained from post-operative patients on telemetry. We propose new temporal features that capture trends and variability in the SIRS variables, and a framework for prediction based on kernel methods. Since the physiological variables of patients in moderate care are sampled irregularly, the temporal features often have missing values. We therefore propose modified kernels that account for these missing values, allowing us to apply existing kernel methods such as the two-class and one-class support vector machines. Results: We compare the predictive power of the temporal features to those of the SIRS criteria. Performance is evaluated not just when the patients are discharged or sent to intensive care unit (ICU), but also some number of hours in advance. The experimental results show that using temporal features leads to improvements over the SIRS criteria by a statistically significant amount. We also present 6 temporal features that appear to be