Introduction: Intensive Care Unit (ICU) medical personnel, in an ongoing process termed ventilation management, utilize patient physiology and pathology data to define ventilator apparatus settings. Aims: The aim of the research is to develop and evaluate in comparison hybrid ventilation advisor systems, that could support ventilation management process, specific to lung pathology for patients ventilated in control mode. Methodology: A questionnaire was designed and circulated to Intensivists. Patient data, as defined by the questionnaire analysis, were collected and categorized into three lung pathologies. Three ICU doctors evaluated correlation analysis of the recorded data. Evaluation results were used for identifying models basic architecture. Two custom software toolboxes were developed for developing hybrid systems; namely the EVolution Of Fuzzy INference Engines (EVOFINE) and the FUzzy Neural (FUN) toolbox. Eight hybrid systems developed with EVOFINE, FUN, ANFIS and ANN techniques were evaluated against applied clinical decisions and patient scenarios. Results: Seventeen (17) models were designed for each of the eight (8) modeling techniques. The modelled process consisted of twelve physiology variables and six ventilator settings. The number of models’ inputs ranged from single to six based on correlation and evaluation findings. Evaluation against clinical recommendations has shown that ANNs performed better; mean average error as percentage for four of the applied techniques was 0.16%, 1.29% & 0.62 for ANN empirical, 0.05%, 2.23% & 2.30% for ANFIS, 0.93%, 2.33% & 1.89% for EVOFINE and 0.73%, 2.63% & 6.56 for FUN NM, in Normal, COPD and ALI-ARDS categories respectively. Additionally evaluation against clinical disagreement SD has shown that 70.6% of the NN empirical models were performing in 90% of their suggestions within clinical SD, while the percentages were 53%, 53% and 59% for the EVOFINE, ANFIS and NN Normalized models respectively. The EVOFINE and ANFIS produced Fuzzy Systems whose architecture is transparent for the user. Visual observation of ANFIS architectures revealed possibly hazardous advices. Evaluation against clinical disagreement has shown that the NN empirical was not producing hazardous advices, while EVOFINE, ANFIS and NN Normalized were shown to produce potentially hazardous advice in 17.6%, 23% and 5.8% of the developed models.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:508008 |
| Date | January 2009 |
| Creators | Tzavaras, Aris |
| Publisher | City University London |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://openaccess.city.ac.uk/12084/ |
Page generated in 0.0014 seconds