Using machine learning to predict the consumption of a Mediterranean diet with untargeted metabolomics data from controlled feeding studies

BACKGROUND AND AIMS: Developing metabolomic signatures of diets is a promising strategy to better understand the diet-health paradigm. Our objective was to develop and validate machine learning (ML) models that predict the consumption of a Mediterranean diet (MedDiet) versus a control diet using untargeted metabolomics data from controlled feeding studies. METHODS AND RESULTS: In the Development set, 26 participants (100 % men) aged 24-62 years consumed a North American diet for 5 weeks followed by a MedDiet for 5 weeks in full-feeding conditions. In the Validation set, 70 participants (54 % men) aged 25-50 years were instructed to follow Canada's Food Guide recommendations for 4 weeks and then consumed a MedDiet for 4 weeks in full-feeding conditions. Plasma metabolites were analyzed using a MPLEx method and an untargeted metabolomics approach. Random forest (RF) and decision tree (DT) models were developed to predict diet assignment using data from the Development set and validated using data from the Validation set. The RF model from the Development set predicted diet assignment with an accuracy of 0.97 (95 %CI: 0.81-1.00). When applied to the Validation set, the RF model had an accuracy of 0.79 (95 %CI:0.71-0.86). Similar results were obtained using the DT model. CONCLUSION: RF and DT models can predict the consumption of a MedDiet diet with high accuracy in full-feeding conditions in males based on plasma untargeted metabolomics data. However, accuracy is reduced when models are applied to a more heterogenous sample (sample of males and females and less controlled feeding conditions).