Substantial response heterogeneity is commonly seen in dietary intervention trials. In larger datasets, this variability can be exploited to identify predictors, for example genetic and/or phenotypic baseline characteristics, associated with response in an outcome of interest.
Using data from a placebo-controlled crossover study (the FINGEN study), supplementing with two doses of long chain n-3 polyunsaturated fatty acids (LC n-3 PUFAs), the primary goal of this analysis was to develop models to predict change in concentrations of plasma triglycerides (TG), and in the plasma phosphatidylcholine (PC) LC n-3 PUFAs eicosapentaenoic acid (EPA) + docosahexaenoic acid (DHA), after fish oil (FO) supplementation. A secondary goal was to establish if clustering of data prior to FO supplementation would lead to identification of groups of participants who responded differentially.
To generate models for the outcomes of interest, variable selection methods (forward and backward stepwise selection, LASSO and the Boruta algorithm) were applied to identify suitable predictors. The final model was chosen based on the lowest validation set root mean squared error (RMSE) after applying each method across multiple imputed datasets. Unsupervised clustering of data prior to FO supplementation was implemented using k-medoids and hierarchical clustering, with cluster membership compared with changes in plasma TG and plasma PC EPA + DHA.
Models for predicting response showed a greater TG-lowering after 1.8 g/day EPA + DHA with lower pre-intervention levels of plasma insulin, LDL cholesterol, C20:3
Our models established new predictors of response in TG (plasma insulin, LDL cholesterol, C20:3n-6, saturated fat consumption, TG, IL-10 and VCAM-1) and in PC EPA + DHA (age and sex) upon intervention with fish oil. We demonstrate how application of statistical methods can provide new insights for precision nutrition, by predicting participants who are most likely to respond beneficially to nutritional interventions.