Predicting human health from biofluid-based metabolomics using machine learning

Biofluid-based metabolomics enables the profiling of thousands of molecules and has the potential to provide highly accurate, minimally invasive diagnostics for a range of health conditions. However, typical metabolomics studies focus on only a few statistically significant features. We study the applicability of machine learning for health state-prediction across 35 human mass spectrometry-based metabolomics studies. Models trained on all features outperform those using only significant features and frequently provide high predictive performance across nine health states, despite disparate experimental conditions and disease contexts. Combining data from different experimental settings (e.g. sample type, instrument, chromatography) within a study minimally alters predictive performance, suggesting information overlap between different methods. Using only non-significant features, we still often obtain high predictive performance. To facilitate further advances, we provide all data online. This work highlights the applicability of biofluid-based metabolomics with data-driven analysis for health state diagnostics.