Detection of Colorectal Carcinoma Based on Microbiota Analysis Using Generalized Regression Neural Networks and Nonlinear Feature Selection
Detection of Colorectal Carcinoma Based on Microbiota Analysis Using Generalized Regression Neural Networks and Nonlinear Feature Selection
复制标题
DOI：
10.1109/tcbb.2018.2870124
复制
发表时间：
2020-03-01
影响因子：
4.5
通讯作者：
Teymourpour, Pegah
中科院分区：
文献类型：
--
作者：
Arabameri, Abazar;Asemani, Davud;Teymourpour, Pegah
为了获得基于肠道微生物群的结直肠癌 (CRC) 筛查工具，我们在此寻求确定用于 CRC 检测的最佳分类器以及用于确定最具辨别力的微生物种类的新型非线性特征选择方法。在这项研究中，使用通用回归神经网络（GRNN）结合所提出的特征选择方法对 141 名患者粪便中的肠道微生物群进行建模。所提出的模型的准确度（AUC = 0.911）比之前的研究（AUC < 0.87）略高。结果表明，梭菌和角双歧杆菌是健康肠道菌群的指标，而 CRC 恰好减少了这些细菌种类。此外，Fusobacter gonidiaformans被发现与CRC密切相关。根据粪便微生物群，结直肠腺瘤的发生没有足够的歧视性，这表明结肠菌群的变化发生在结直肠癌发展的晚期阶段，而不是最初的腺瘤。将所提出的模型与粪便潜血测试（FOBT）相结合，CRC 检测准确性几乎保持不变（AUC = 0.915）。使用来自美国和奥地利的独立队列验证了所提出方法的性能。我们的结果表明，所提出的特征选择方法与 GRNN 相结合可能是一种准确的 CRC 检测方法。
To obtain a screening tool for colorectal cancer (CRC) based on gut microbiota, we seek here to identify an optimal classifier for CRC detection as well as a novel nonlinear feature selection method for determining the most discriminative microbial species. In this study, the intestinal microflora in feces of 141 patients were modeled using general regression neural networks (GRNNs) combined with the proposed feature selection method. The proposed model led to slightly higher accuracy (AUC = 0.911) than previous studies (AUC < 0.87). The results show that the Clostridium scindens and Bifidobacterium angulatum are indicators of healthy gut flora and CRC happens to reduce these bacterial species. In addition, Fusobacterium gonidiaformans was found to be closely correlated with the CRC. The occurrence of colorectal adenoma was not sufficiently discriminatory based on fecal microbiota implicating that the change of colonic flora happens in the advanced phase of CRC development rather than initial adenoma. Integrating the proposed model with fecal occult blood test (FOBT), the CRC detection accuracy remained nearly unchanged (AUC = 0.915). The performance of the proposed method is validated using independent cohorts from America and Austria. Our results suggest that the proposed feature selection method combined with GRNN is potentially an accurate method for CRC detection.