Dietary information improves cardiovascular disease risk prediction models

Background/objectives: Data are limited on cardiovascular disease (CVD) risk prediction models that include dietary predictors. Using known risk factors and dietary information, we constructed and evaluated CVD risk prediction models. Subjects/methods: Data for modeling were from population-based prospective cohort studies comprised of 9026 men and women aged 40-69 years. At baseline, all were free of known CVD and cancer, and were followed up for CVD incidence during an 8-year period. We used Cox proportional hazard regression analysis to construct a traditional risk factor model, an office-based model, and two diet-containing models and evaluated these models by calculating Akaike information criterion (AIC), C-statistics, integrated discrimination improvement (IDI), net reclassification improvement (NRI) and calibration statistic. Results: We constructed diet-containing models with significant dietary predictors such as poultry, legumes, carbonated soft drinks or green tea consumption. Adding dietary predictors to the traditional model yielded a decrease in AIC (delta AIC=15), a 53% increase in relative IDI (P-value for IDI <0.001) and an increase in NRI (category-free NRI=0.14, P <0.001). The simplified diet-containing model also showed a decrease in AIC (delta AIC=14), a 38% increase in relative IDI (P-value for IDI <0.001) and an increase in NRI (category-free NRI=0.08, P<0.01) compared with the office-based model. The calibration plots for risk prediction demonstrated that the inclusion of dietary predictors contributes to better agreement in persons at high risk for CVD. C-statistics for the four models were acceptable and comparable. Conclusions: We suggest that dietary information may be useful in constructing CVD risk prediction models.