Bacteria and Wound Stage Classification based on Machine Learning and E-nose

• Main Authors: Yi-JhenWu, 吳宜臻
• Other Authors: Che-Wei Lin
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/2b76b7

碩士 === 國立成功大學 === 生物醫學工程學系 === 107 === The development of electronic nose combining the technology of software and hardware on “Volatile organic compounds sensor array” and “Multiple gas classification algorithms in machine learning” has gradually supported industrial technology analysis in food safety, public health, air pollution, and other applications and let our life getting a better trend. To the application of home care, it is expected that the user can immediately receive the evaluation results from various types of wounds by a simple method, and also that provides an appropriate treatment or prevent deterioration measures. For the experiment of detecting wound odor, we specially designed a soft cover with a tube to isolate the external ambient and to collect the specific volatile organic compounds by E-nose. In the development of the algorithm, we compared the raw signals with six channels and obtained the feature extraction from raw signals per second (including mean, standard deviation, variance, quartile deviation, Shannon entropy, and logical energy entropy) and also adding the feature selections method, like NCA or PCA for dimensionality reduction. Then validating the results of performance with four classifiers (decision tree, linear discriminant analysis, quadratic discriminant analysis, and k-nearest neighbor) by k-fold and leave-one-out cross validation. There were two items in the collection of samples, one was classification of bacteria and another one was classification of wound. In the case of bacteria, we targeted two kinds of bacteria with bio-safety level 1 (BSL-1), including Bacillus subtilis and Escherichia coli; two with bio-safety level 1 (BSL-1), including Staphylococcus aureus and Candida albicans; also adding one of growth media-LB. To discriminate the results of classification in five classes, and the total sample datapoints were 4500. And in the case of healing wounds, we measured the complete comprehensive healing phase of wounds, the classes from healthy skin to injured fresh skin, and through coagulation, inflammation, proliferation, remodeling, and finally healed wound completely. Multiple categories of phase could be effectively predicted and the total sample datapoints were 105,000. In this research, we also compare the classification results between for classifier with feature extraction and selection to improve the sensitivity and specificity of the prediction. As the results, we obtained that the best results in 10-fold validation were in the performance of bacteria could reach 99.29% accuracy, 99.29% sensitivity and 99.82% specificity; the performance of wound could reach 99.79% accuracy, 99.79% sensitivity and 99.97% specificity.