Abstract
Randomized controlled trials (RCTs) have demonstrated benefits of marine omega-3 polyunsaturated fatty acids (omega-3 FA) supplementation for the prevention of coronary heart disease (CHD). However, it has not been clear which individuals benefit the most from supplementation. We sought to develop an omega-3 effect score to stratify individuals according to their expected benefit from supplementation. Among the 25,871 randomized participants without a history of cardiovascular disease in the VITamin D and OmegA-3 TriaL (VITAL), we applied machine-learning (ML) approaches to predict individual treatment effect of omega-3 FA supplementation on 5-year CHD risk using 11 covariates pre-specified in the VITAL protocol. An omega-3 effect score was developed such that each covariate contributed linearly. ML algorithms effectively stratified participants by their expected benefit according to individual factors; for example, there was 1.21% absolute CHD risk reduction in the top tertile of the expected benefit, compared with the average effect of 0.47% risk reduction. Baseline diabetes, race, hypertension, sex, and fish intake contributed the most to the omega-3 effect score. Five-year CHD risk was 2.5% among those in the omega-3 arm and 3.2% among those in the placebo arm with omega-3 effect score ≥ 4 (upper 70th percentile), and 1.4% among the omega-3 arm and 1.3% among the placebo arm in those with the score < 4, respectively. The transportability of the score to the National Health and Nutrition Examination Survey (NHANES) data was confirmed. Although testing of the score in a new RCT is warranted, the proposed omega-3 effect score holds promise for guiding decision making for omega-3 FA supplementation in the US primary prevention population.