Phenotyping women based on dietary macronutrients, physical activity and body weight using machine-learning tools

Abstract

AbstractBackgroundNutritional phenotyping is a promising approach to achieve personalized nutrition. While conventional statistical approaches haven’t enabled personalizing well yet, machine-learning tools may offer solutions that haven’t been evaluated yet.ObjectiveThe primary aim of this study was to use energy balance components – input (dietary energy intake and macronutrient composition), output (physical activity) to predict energy stores (body weight) as a way to evaluate their ability to identify potential phenotypes based on these parameters.MethodsWe obtained data from the Women’s Health Initiative –Observational Study (WHI-OS) from BioLINCC. We chose dietary macronutrients – carbohydrate, protein, fats, fiber, sugars & physical activity variables – energy expended from mild, moderate and vigorous intensity activity h/wk to predict current body weight either numerically (as kg of body weight) or categorically (as BMI categories). Several machine-learning tools were used for this prediction – k-nearest neighbors (kNN), decision trees, neural networks (NN), Support Vector Machine (SVM) regressions and Random Forest. Further, predictive ability was refined using cluster analysis, in an effort to identify putative phenotypes.ResultsFor the numerical predictions, kNN performed best (Mean Approximate Error (MAE) of 2.71kg, R2 of 0.92, Root mean square error (RMSE) of 4.96kg). For categorical prediction, ensemble trees (with nearest neighbor learner) performed best (93.8% accuracy). K-means cluster analysis identified 11 clusters suggestive of phenotypes, based on significantly improved predictive accuracy. Within clusters, individual macronutrient gain and loss modeling identified that some clusters were strongly predicted by dietary carbohydrate while others by dietary fat.ConclusionsMachine-learning tools in nutritional epidemiology could be used to identify putative phenotypes.