Abstract: Objective To investigate the variation trend of the classification accuracy while the number of normal variables increases. Methods Firstly, 11 metabolomics datasets were simulated whose variables in normal distribution increased gradually. Secondly, 5 statistical methods were adopted to compare the classification accuracy, including the traditional methods: Bayes discrimination, Fisher discrimination, Partial least squares discrimination analysis(PLS-DA)and the machine learning methods: Random Forest(RF), Support Vector Machine(SVM). Lastly, the rationality of the results in simulations were evaluated with 2 sets of real data. Results The normality of matabolomics data could influence the classification accuracy of Bayes discrimination, Fisher discrimination and PLS-DA. Besides, the classification accuracy increased with the larger number of normal variables. However, the normality of data did not have discernible effect on their classification accuracy for SVM and RF. Conclusion The traditional statistical methods have requirements for the normality of metabolomics data, such as Bayes discrimination, Fisher discrimination and PLS-DA, while the machine learning methods such as SVM and RF require little of it and can produce higher and more stable classification accuracy.

Key words: Data Normality, Classification Accuracy, Bayes Discrimination, Fisher Discrimination, Random Forest, Partial Least Squares Discrimination Analysis, Support Vector Machine