The National Health and Nutrition Examination Survey (NHANES), conducted by the American National Center for Health Statistics, is a periodical program that adopts a complex, multistage, probability sampling design to recruit individuals that are representative of the general US population [19]. Participants of this work come from the NHANES 2007–2018 cycle. We included individuals who met the following criteria: (1) age ≥20 years old; (2) participants without pregnancy; (3) Participants with available 24-hour dietary interview data. The survival analysis extracted participants from the NHANES 2007–2014 cycle and excluded individuals without follow-up information in combination with other prespecified inclusion criteria.

This retrospective cohort study was carried out in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology reporting guideline for cohort studies [20].

Original dietary data were collected from NHANES dietary interview, which contains a dietary recall about the food, beverages, and supplements that consumed during the 24-hour prior to the interview and has been validated elsewhere [21].

HEI-2015 is the latest version of HEI, which consists of 13 food parameters including total vegetables, greens & beans, total fruits, whole fruits, whole grains, dairy, total protein foods, seafood & plant proteins, fatty acids, sodium, refined grains, saturated fats, added sugars. The total and component HEI scores were calculated following the four steps of the HEI scoring algorithm on the website and a simple HEI scoring algorithm was adopted to calculate the total and component HEI scores [22]. According to the DII calculating method reported by N. Shivappa et al. [16], we extracted 28 food parameters from the 24-hour dietary recall, including protein, carbohydrates, total fat, cholesterol, saturated fat, polyunsaturated fatty acids (PUFAs), monounsaturated fatty acids (MUFAs), alcohol, niacin, vitamin A, riboflavin, thiamin, vitamin B6, vitamin B12, vitamin C, vitamin D, vitamin E, fiber, Mg, Fe, Zinc, Selenium, β-carotene, folic acid, n-3 fatty acids, n-6 fatty acids, caffeine, and energy. The process of DII calculation was presented in previous work [23]. Overall, a higher and positive DII score represents a pro-inflammatory dietary potential, while a lower and negative DII score means an anti-inflammatory dietary potential. The intact DII score system consists of 45 food parameters, but it has been demonstrated that DII calculated from less than 30 food parameters remained its predictive value [24, 25].

The follow-up was ended on December 31, 2019. The primary endpoint was all-cause death, and the secondary endpoints included death of malignant neoplasms, death of heart disease, death of chronic lower respiratory diseases, and death of cerebrovascular diseases, all of which were selected according to the leading causes of death in the United States and were extracted from the National Death Index (NDI) records [26, 27]. The definition of cause-specific death was based on the International Classification of Diseases -10 codes (Malignant neoplasms: C00-C97; heart disease: I00-I09, I11, I13, I20-I51; chronic lower respiratory diseases: J40-J47; cerebrovascular diseases: I60-I69).

Participants were diagnosed as diabetes if they met one of the following diagnostic criteria: (1) a self-reported physician diagnosis of diabetes; (2) fasting glucose ≥7.0 mmol/L; (3) 2-hour plasma glucose levels in oral glucose tolerance test ≥11.1 mmol/L; (4) glycated hemoglobin A1c (HbA1c) ≥6.5%; (5) use of diabetes medication or insulin. Participants who met one of 2–5 criteria but without a self-reported physician diabetes diagnosis were divided into the undiagnosed diabetes group, while participants who met the first criterion were grouped into diagnosed diabetes group. The definition of hyperlipidemia was: (1) total cholesterol ≥200 mg/dL; (2) low-density lipoprotein-cholesterol (LDL-C) ≥ 130 mg/dL; (3) high density lipoprotein-cholesterol (HDL-C) < 40 mg/dL in men and 50 mg/dL in women; (4) triglyceride ≥ 150 mg/dL; (5) use of cholesterol-lowering drugs. The diagnostic criteria for hypertension were (1) a self-reported diagnosis of hypertension from a physician or other health professional; (2) systolic blood pressure ≥140 mmHg and/or diastolic blood pressure ≥90 mmHg (at least 3 times); (3) use of antihypertensive drugs. The moderate or heavy drink was defined as: (1) ≥3 drinks per day for men or ≥2 drinks per day for women; (2) binge drinking ≥2 days per month. Binge drinking was defined as ≥5 drinks at a time for men or ≥4 drinks for women [28].

Due to the multistage complex design of NHANES, all analyses in this work have incorporated oversampling, clustering, and stratification to estimate the representative statistics of the general US population [19]. According to the type of variables, baseline characteristics were listed as weighted mean and 95% confidence interval (CI) (continuous variables) or weighted proportions (categorical variables). Comparison of parameters among three groups (without diabetes, with undiagnosed diabetes, and with diagnosed diabetes) was performed by weighted generalized linear regression models (continuous variable), weighted chi-square test (categorical variable), or weighted logistic regression models (categorical variable). The radar plot was plotted to show the distribution of HEI component scores, which was generated by divide the component score by the maximum score of that component. The max scores of each HEI food parameter were listed in Supplementary Table S1.

Participants were equally divided into 3 groups according to their HEI or DII scores: low HEI (0 ≤ HEI ≤ 44.4), medium HEI (44.4 < HEI ≤ 56.8), high HEI (56.8 < HEI ≤ 96.0), and low DII (−5.28 ≤ DII ≤ 0.81), medium DII (0.81 < DII ≤ 2.65), high DII (2.65 < DII ≤ 5.47). To investigate whether dietary patterns correlate with the long-term mortality of the general US population, we performed age-adjusted weighted Cox proportional hazard regression models and included HEI or DII as continuous or categorical variables respectively. Furthermore, the weighted Cox regression models have been adjusted for age, sex, educational level, body mass index (BMI), smoking, hypertension, hyperlipidemia, diabetes diagnosis, and alcohol consumption. P for interaction was calculated by baseline diabetes diagnosis.

We considered a two-sided p value < 0.05 as statistically significant. All statistics were conducted using the R software version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria).