Identifying cardiovascular risk factor-related dietary patterns with reduced rank regression and random forest in the EPIC-NL cohort.

Several methods are used to determine dietary patterns. Hybrid methods incorporate information on nutrient intake or biological factors to extract patterns relevant to disease etiology. We explore differences between patterns derived with 2 hybrid methods with those obtained by a posteriori methods and compare associations of these patterns with coronary artery disease (CAD) and stroke risk. Food-frequency questionnaires were used to estimate dietary intake in 34,644 participants of European Prospective Investigation into Cancer-Netherlands at baseline (1993-1997). Follow-up was complete until 31 December 2007. Hybrid methods to determine dietary patterns were reduced rank regression (RRR) and random forest with classification tree analysis (RF-CTA). Included risk factors were body mass index, total:high-density lipoprotein cholesterol ratio, and systolic blood pressure. Results were compared with those from principal component analysis (PCA) and k-means cluster analysis (KCA), respectively. Both RRR and PCA derived a "Western," "prudent," and "traditional pattern." All RRR patterns were significantly associated with CAD risk [highest vs. lowest quartile factor score; HR: 1.45 (95% CI: 1.25, 1.69), 0.86 (0.74, 0.99), and 1.25 (1.07, 1.47), respectively]. Only the prudent RRR factor was statistically significant associated with stroke (HR: 0.76; 95% CI: 0.59, 0.97). From the PCA patterns, only the traditional pattern was associated with CAD (HR: 1.29; 95% CI: 1.11, 1.50). RF-CTA derived 7 dietary patterns that could be categorized as "Western-like," "prudent-like," and "traditional-like." KCA established a prudent and Western cluster. Compared with the RF-CTA "prudent-like 1" pattern, only the "traditional-like 1" pattern was associated with CAD (HR: 1.36; 955 CI: 1.12, 1.65). None of the RF-CTA groups were associated with stroke. Compared with the Western KCA cluster, the prudent cluster was not associated with CAD or stroke. Including risk factors in RRR and RF-CTA resulted in small differences in food groups, contributing to similar patterns that showed in general stronger associations with CAD than PCA and KCA, respectively.