Metagenomics is one of the emerging concepts in personalized medicine approaches to take care of and improve human health. Numerous studies have revealed that metagenomic data can associate with a vast of human diseases. Recent advancements in machine learning techniques and computation resources enable us to speed up the data processing and also improve the performance in diagnosis accuracy. However, we face difficulties to process metagenomic data due to its complexities and high dimension. This work proposes an approach based on an explanation model to perform feature selection tasks. The proposed approach selects a small set of features from the original features with Interpretable Model-agnostic Explanations that can obtain better performances than feature selection based on importance scores generated from a robust learning machine learning such as Random Forests. We are expected that the approach can be an efficient feature selection method compared to classic feature selection techniques.