ABSTRACT
Over the last decade, metabolomics research has produced thousands of research works. Many of them were describing the ability of machine learning methods to detect diverse health conditions based on mass-spectrometry, nuclear magnetic resonance or artificial olfaction analysis of body fluids.
While few success stories exist, most described applications never found the road to clinical exploitation. Most described methodologies were not reliable and were plagued by numerous problems that prevented practical application beyond the lab. This work gives some insight on the reasons behind this lack of generalizability and emphasizes the need of external validation in metabolomics research. We describe some statistical and methodological pitfalls of the current data analysis practice and we give some best practice recommendations for researchers in the field.
