Predicting Plasma Vitamin C Using Machine Learning

Precision Nutrition makes use of personal information about individuals to produce nutritional recommendations that have more utility than general population level recommendations. In many cases, being able to predict current status is a necessary first step in offering tailored nutritional advice. The objective of this study is to predict plasma vitamin C using machine learning. The NHANES dataset was used to predict plasma vitamin C in a cohort of 2952 American adults using regression algorithms and clustering in a way that a hypothetical health application might. Variables were selected based on a known or hypothesized relationship with plasma vitamin C, and variables that are expensive or difficult to obtain were excluded in order to more closely replicate the situation of a real health application. The best performance was seen with the XGBoost regressor, with random forest performing almost identically. Clustering was also investigated as a means of improving regression accuracy by splitting the data up into smaller yet more homogeneous groups, however, this was not successful. The low R-squared scores obtained by the models are likely to be due to the low resolution of the NHANES data, particularly the dietary data. This emphasizes the need for high-quality data sets in Precision Nutrition research.