Abstract
There is a serious threat of high-resistance pneumonia bacteria in the Intensive Care Unit. If a doctor finds out patients infected with pneumonia, he needs to determine which types of bacteria causes the trouble in order to prescribe the right antibiotic medicine. Up to now bacteria culture is the common way of diagnosis, but it takes five to six days to get the results. This is usually slow for emergency and some patients do not survive during the wait. Therefore, the idea of installing an electronic nose with pneumonia bacteria recognition function on artificial respiration is born. The electronic nose aims to detect whether the constituents of the gas exhaled by the patient imply bacterial infection in the lung. The monitoring can be done continually, helping physicians to perform early diagnosis and prescribe the right medicine, grasping the prime-time on saving the patient’s life. Nano composite-array sensors are used here to get the signal from exhaled gas by the patients. Pattern recognition approaches were adopted to analyze the data; in this thesis, we choose the K nearest neighbor method (KNN) as our classifier and use sequential feature selection to obtain features that are most effective in discriminating between different types of pneumonia bacteria. The results show that the recognition rate of pneumonia detection increased slightly from 73% to 75% and the recognition rate of pneumonia bacteria recognition improved from 66% to 73%, thanks to the sequential feature selection. This thesis also proposes a novel mixture gas recognition method, which we call the Individual Constituent-Decision Method (ICDM). The method utilizes the physical meaning of mixture, and it can be optimized separately to detect each constituent of interest. Results from all constituent-decision makers can be combined so as to produce a final result of mixture gas recognition. To validate ICDM in this thesis, because the constituting gases are unknown in the gas exhaled from pneumonia patients, we use fruit juice mixtures instead, to emulate the scenario of mixture-gas sensing. We compare ICDM with the traditional method that aims to classify all mixture combinations at one shot. Results show that ICDM has a better performance because it can find different recognition models (the best feature subset and parameters of the classifiers) for each individual constituent. This validates the idea that ICDM should be able to optimize on each individual constituent-decision.