Investigation of Pre-Symptomatic Biomarkers of Sepsis (303242)

Keywords: Prediction, Neural Networks, Sepsis, large scale data, complex data, modelling

Infection driven conditions such as sepsis are difficult to treat and lead to millions of deaths per year. The ability to predict sepsis in the pre-symptomatic phase maximizes the effect of medical intervention. A study was designed to investigate the ability to classify between patients that develop sepsis and comparator patients, before symptoms show. A large scale observational study collected blood from patients before and following elective surgery to allow analysis of a time course of up to 7 days after surgery. The study produces large scale, noisy multivariate data which has proven difficult for standard statistical techniques. Exploratory analysis and statistical testing is carried out, using a range of ranking tests in combination, to test for differences in gene expression between groups. This allows the reduction of the parameters that are taken forward to build classification models. A range of classifications models have been investigated to derive a predictive model for probability of sepsis. Artificial Neural Networks ability to deal with large amounts of noisy data has been vital to the analysis of this data and have been used to achieve high predictive accuracies