Data Transformation method for Comprehensive Classification to Blood Test Data

Abstract

This paper proposes a method to transform data to better fit a normal distribution in medical record analysis. We proposed a method to comprehensively analyze medical record using various machine learning method. Because some machine learning method assume that the data follow normal distribution, we need to transform the data to fit normal distribution. Ordinary method used Box-Cox transform that is a power transform, but it cannot apply negative values. In this paper, we propose to use Yeo-Johnson transform instead of Box-Cox transform, because it can apply negative values. We also propose one-class power transform that estimates the transform parameter only from one class. We confirmed that our proposed method can transform the data even if some values are negative and the transformed data fit normal distribution better than Yeo-Johnson transform using both classes.