Cardiovascular Disease Events and Life Expectancy Lost Attributable to Machine Learning-Derived Dietary Networks:...

Artificial intelligence and machine learning (ML) are transforming nutritional epidemiology by revealing dietary network structures invisible to conventional correlation-based methods. Although traditional approaches fail to capture conditional food-food dependencies, advanced computational techniques can identify co-consumption patterns and their differential associations with cardiovascular disease (CVD) and mortality.

To apply ML-based network analysis to characterize complex food-food relationships and identify dietary communities representing distinct food combinations consumed by Canadian adults, and to examine associations of these community scores with CVD risk, all-cause mortality, and life expectancy (LE).

Dietary intake was obtained from 2 cycles of the nationally representative samples from the Canadian Community Health Survey-Nutrition 2004 (n = 15,835; linked with health administrative databases) and 2015 (n = 13,557). We used semiparametric Gaussian Copula graphical models to infer conditional dependencies between food groups and the Louvain algorithm for unsupervised community detection. Individual-level community scores were derived, and communities' associations with mortality and CVD were measured using weighted multivariable-adjusted Cox models. LE was estimated across community scores using abridged life tables.

Three dietary communities were identified: vegetable-rich (VR), high-sugary beverage and low fruit (HSBLF), and high-fat breakfast. Higher VR scores were associated with lower all-cause mortality in the overall sample [hazard ratios (HR): 0.49; 95% confidence interval (CI): 0.36, 0.67], females (HR: 0.41; 95% CI: 0.26, 0.63), and males (HR: 0.52; 95% CI: 0.33, 0.81), and with lower CVD risk overall (HR: 0.55; 95% CI: 0.32, 0.94). Higher HSBLF scores were associated with increased mortality (HR: 1.31; 95% CI:1.06, 1.62) and among males (HR: 1.39; 95% CI:1.03, 1.88). LE at 45 y was longer with higher VR scores or lower HSBLF/high-fat breakfast scores (1.0-8.3 y for females, 0.8-6.1 y for males).

ML-based network analysis identifies dietary communities characterized by unique combinations of foods that are differentially associated with mortality and CVD, demonstrating the value of advanced computational approaches for characterizing dietary patterns relevant to chronic disease prevention.