Predicting Sepsis Severity from Limited Temporal Observations

Abstract

Sepsis, an acute systemic inflammatory response syndrome caused by severe infection, is one of the leading causes of in-hospital mortality. Our recent work provides evidence that mortality rate in sepsis patients can be significantly reduced by Hemoadsorption (HA) therapy with duration determined by a data-driven approach. The therapy optimization process requires predicting high-mobility group protein B-1 concentration 24 hours in the future. However, measuring sepsis biomarkers is very costly, and also blood volume is limited such that the number of available temporal observations for training a regression model is small. The challenge addressed in this study is how to balance the trade-off of prediction accuracy versus the limited number of temporal observations by selecting a sampling protocol (biomarker selection and frequency of measurements) appropriately for the prediction model and measurement noise level. In particular, to predict HMGB1 concentration 24 hours ahead when limiting the number of blood drawings before therapy to three, we found that the accuracy of observing HMGB1 and three other cytokines (Lsel, TNF-alpha, and IL10) was comparable to observing eight cytokines that are commonly used sepsis biomarkers. We found that blood drawings 1-hour apart are preferred when measurements are noise free, but in presence of noise, blood drawings 3 hours apart are preferred. Comparing to the data-driven approaches, the sampling protocol obtained by using domain knowledge has a similar accuracy with the same cost, but half of the number of blood drawings.