基于傅里叶近红外特征光谱的血流感染致病菌鉴别研究

利用傅里叶变换近红外光谱(FT-NIR)收集1 000~1 852 nm范围内3种常见病原菌大肠杆菌(ATCC25922)、金黄色葡萄球菌(ATCC 29213)、铜绿假单胞菌(ATCC 27853)的近红外透射光谱,采用竞争性自适应重加权算法(CARS)对波长变量进行筛选,并分别结合偏最小二乘判别分析(PLS-DA)、最小二乘-支持向量机(LS-SVM)建立鉴别模型.比较两种鉴别模型在进行波长变量优选前后的性能发现,采用全波段建模的PLS-DA与LS-SVM两种模型的预测性能较低;利用CARS对波长变量进行筛选后,对优选的24个特征波长分别建立两种鉴别模型,模型预测性能明显提高,其中以LS-SVM模型最优,3种病原菌准确率分别为85.0%,100%和100%.研究结果表明,利用CARS能够有效去除光谱无用信息,减少模型复杂度,增强模型预测性能,结合LS-SVM可为临床利用近红外快速检测血流感染病原菌提供一种新的方法.

Research on identification of bacteria in bloodstream infections based on FT-NIR characteristic spectrum selection

In order to establish a method that can quickly and accurately discriminate clinically bacterial species in bloodstream infections. Fourier transform near-infrared spectroscopy (FT-NIR) was used to collect the range from 1000~1852nm near-infrared spectra of the three common pathogen bacteria which included Escherichia coli, Pseudomonas aeruginosa, and Staphylococcus aureus. Competitive adaptive reweighted sampling (CARS) algorithm to be used to select the characteristic wavelength variables which were used to build the PLS-DA model and LS-SVM model. Moreover, performance of PLS-DA and LS-SVM model built by the characteristic wavelength variables were also compared with PLS-DA and LS-SVM model built by full spectra. Studies showed that full spectra model had poor prediction performance due to some wavelength variables contains irrelevant information. The prediction performance of PLS-DA model and LS-SVM had improved significantly by using twenty four characteristic wavelengths selected by CARS. Meanwhile, LS-SVM model achieved the optimal performance which he correct rate of three pathogen bacteria was 80%, 100%, 85% respectively. The results showed that CARS can remove useless information, reducing model complexity, enhanced model prediction performance and combined with LS-SVM can accurately identify clinical pathogens.