Sepsis remains the second largest killer in the Intensive Care Unit (ICU), giving rise to a significant economic burden ($17b per annum in the US, 0.3% of the gross domestic product). The aim of the work described in this thesis is to improve the estimation of severity in this population, with a view to improving the allocation of resources. A cohort of 2,143 adult patients with sepsis and hypotension was identified from the MIMIC-II database (v2.26). The implementation of state-of-the-art models confirms the superiority of the APACHE-IV model (AUC=73.3%) for mortality prediction using ICU admission data. Using the same subset of features, state-of-the art machine learning techniques (Support Vector Machines and Random Forests) give equivalent results. More recent mortality prediction models are also implemented and offer an improvement in discriminatory power (AUC=76.16%). A shift from expert-driven selection of variables to objective feature selection techniques using all available covariates leads to a major gain in performance (AUC=80.4%). A framework allowing simultaneous feature selection and parameter pruning is developed, using a genetic algorithm, and this offers similar performance. The model derived from the first 24 hours in the ICU is then compared with a “dynamic” model derived over the same time period, and this leads to a significant improvement in performance (AUC=82.7%). The study is then repeated using data surrounding the hypotensive episode in an attempt to capture the physiological response to hypotension and the effects of treatment. A significant increase in performance (AUC=85.3%) is obtained with the static model incorporating data both before and after the hypotensive episode. The equivalent dynamic model does not demonstrate a statistically significant improvement (AUC=85.6%). Testing on other ICU populations with sepsis is needed to validate the findings of this thesis, but the results presented in it highlight the role that data mining will increasingly play in clinical knowledge generation.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:669721 |
Date | January 2014 |
Creators | Mayaud, Louis |
Contributors | Tarassenko, Lionel ; Clifford, Gari ; Annane, Djillali |
Publisher | University of Oxford |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://ora.ox.ac.uk/objects/uuid:55a57418-de16-4932-8a42-af56bd380056 |
Page generated in 0.0015 seconds