关键词: 疾病预测, 肠道菌群, 特征筛选, 融合模型, 宏基因组

Abstract:

The utilization of metagenomic analysis to investigate human diseases and predict health conditions is a current focal point of research. Through the application of bioinformatics tools such as KneadData and MetaPhlAn2， the raw metagenomic data undergoes quality control and host contamination removal were carried out to obtain the pure sequences. Subsequently， dimensionality reduction methods and a random forest model were employed to identify microbial taxa that were highly correlated with disease occurrence， serving as replacements for the original data features in the disease prediction model. A fusion disease prediction model was constructed by integrating multilayer perceptron （MLP）， support vector machine （SVM）， and extreme gradient boosting （XGBoost） as sub-models. Following feature selection and cross-validation on datasets pertaining to liver cirrhosis， type 2 diabetes， and obesity， the obtained AUC values were 0.928 6， 0.652 1， and 0.574 7， respectively. The area under the ROC curve demonstrated that the model augmented with the selected microbial taxa， which could efficiently and accurately screen and diagnose diseases， effectively distinguishing between healthy individuals and patients. This work provided valuable insights for the establishment of a novel non-invasive and quantifiable auxiliary diagnostic method.

Key words: disease prediction, intestinal microbiota, feature screening, fusion model, metagenomics