The study population is derived from the PLCO Trial, a multicenter clinical trial aimed at assessing whether specific screening tests can reduce cancer mortality. The detailed trial methodology has been extensively described elsewhere (22). In summary, eligible participants were recruited from 10 study centers across the U.S. in the period 1993–2001, all of whom provided informed consent. Upon randomization, participants were evenly allocated to the control group received standard medicine care, and the intervention group underwent regular cancer screening as per the PLCO protocol. At the onset of the PLCO trial, each participant completed a baseline questionnaire (BQ), voluntarily reporting baseline information such as demographic characteristics and medical history. In 1998, the PLCO screening centers introduced a dietary history questionnaire (DHQ), which was to be completed concurrently with the BQ for participants enrolled after 1998, and for those enrolled before 1998, they were required to supplement any previously incomplete DHQ. The DHQ comprised inquiries about the intake of 124 foods and supplements over the past year, including intake amount, daily frequency, and other details. Participants were followed up upon entry into the PLCO trial, with data collection including cancer diagnoses up to 2009 and mortality status up to 2018. This study has received approval from the National Cancer Institute (NCI) (Project ID: PLCO-1560).

In our study, we further excluded: (1) individuals who did not return or complete the BQ (n = 4,918); (2) individuals with invalid DHQ completion, defined as lacking completion date, confirmed death before completing, having ≥8 missing responses, or extreme calorie intake values (top, or bottom 1%) (n = 38,462); (3) individuals with any type of personal history of cancer before entering the DHQ (n = 9,684); (4) individuals who experienced outcome events between entering the DHQ and DHQ completion, including the occurrence of lung cancer, death, or loss to follow-up (n = 68); (5) individuals with extreme energy intake (less than 800 kcal or more than 4,200 kcal for males, and less than 600 kcal or more than 3,500 kcal for females) (n = 3,296). Ultimately, 98,459 individuals were incorporated in this research. The exclusion and inclusion process are depicted in the Supplementary Figure S1.

Baseline information of participants was collected from the BQ, including sex, race, marital status, education level, body mass index (BMI), pack-years, history of respiratory diseases, history of diabetes, history of hypertension, history of aspirin use, family history of lung cancer, cigarette smoking status, and randomization arm. Body mass index was calculated as weight (kg) divided by height (m) squared. Respiratory diseases included bronchitis and emphysema. Dietary data, age and alcohol drinking status were collected using the DHQ, whose validity has been confirmed through the United States National Health and Nutrition Examination Survey (23). The DHQ includes the 2015 Healthy Eating Index (HEI-2015), which is used to measure individual dietary quality (13). In this study, the HEI-2015 was used as a benchmark to compare the effectiveness and value of the LLDS in measuring dietary quality.

The LLDS is developed based on the 2015 Dutch Dietary Guidelines and international evidence on the relationship between diet and chronic disease (15). Specifically, the 110 items in the Food Frequency Questionnaire (FFQ) are divided into 22 food groups, which are categorized into four classes based on their impact on human health: positive, negative, neutral, and unknown. Ultimately, positive groups (including fruit, vegetables, fish, legumes and nuts, whole grains products, oils and soft margarines, coffee, unsweetened dairy, and tea) and negative groups (butter and hard margarines, red and processed meat, sugar-sweetened beverages) were selected for inclusion in the scoring scheme. Food intake was normalized to a standard energy intake of 1,000 kcal. The energy-adjusted food intake was then distributed into quintiles ranging from 0 to 4. The highest quintile was assigned a score of 4 for positive food groups, while the lowest quintile was assigned a score of 0. In contrast, the scoring methodology for negative food groups was inverted. After all, the scores for all food groups were aggregated to derive the LLDS. A higher score indicates greater adherence to LLDS, reflecting higher dietary quality (15).

Each participant received an annual report requesting information on any cancer diagnoses. In the event of a diagnosis report, they were asked to provide the diagnosis date, site, cancer type, the medical institution making the diagnosis, and contact information for the diagnosing physician. For non-responses, the research team reached out to participants via email or phone to verify their cancer diagnosis and survival status. For each case of lung cancer, researchers made efforts to confirm the diagnosis and obtain detailed information by contacting the diagnosing facility or physician. To ensure comprehensive mortality information, the trial also regularly checked national death registries and used the International Classification of Diseases, Ninth Revision (ICD-9), based on death certificates, to ascertain causes of death.

This study encountered missing data for several covariates. Among them, smoking pack-years, a continuous variable, had the largest proportion of missing data at 1.12%. Other covariates had missing values for <1% of participants (Supplementary Table S1). Therefore, we imputed missing values of continuous variables using median imputation, including BMI and pack-years. For categorical variables, including race, marital status, education level, smoking status, family history of lung cancer, history of emphysema, history of bronchitis, history of hypertension, history of diabetes, and aspirin use, the missing values were imputed using mode imputation. For mode imputation, we first identified the most frequently occurring category (i.e., the mode) for each variable, and then used this mode to fill in the missing values for the corresponding variable (24). After imputation, the distribution of categorical variables and the means of continuous variables remained virtually unchanged compared to the pre-imputation values (Supplementary Table S1).

Participants were grouped into quartiles depending on their LLDS. Cox proportional hazards regression model was employed, with follow-up time as the time variable. The lowest quartile is considered as the reference group. Hazard ratios (HRs) and 95% confidence intervals (CIs) were computed for each quartile relative to the reference group. As shown in Supplementary Figure S2. The follow-up data regarding the incidence of lung cancer spanned from the completion date of the DHQ to the occurrences of lung cancer diagnosis, loss to follow-up, or termination of follow-up (12/31/2009). For lung cancer mortality, the follow-up period concluded in 2018. The median LLDS was allocated to participants in each quartile and treated as a continuous variable in Cox regression analysis to derive a trend p-value, evaluating the statistical significance. Furthermore, LLDS was analyzed as a continuous variable to assess the risk estimate associated with 1-point increment in LLDS. Based on existing medical knowledge, clinical experience, and a review of relevant previous studies, we constructed three models: unadjusted model, Model 1 adjusted for basic demographic characteristics, and Model 2 as a fully adjusted model. Specially, Model 1 adjusted for age, sex, race, marital status, and education level. As the fully adjusted model, Model 2 not only adjusted for the variables in Model 1 but also additionally adjusted for BMI, alcohol drinking status, cigarette smoking status, pack-years, history of respiratory diseases, history of diabetes, history of hypertension, history of aspirin use, family history of lung cancer, and randomization arm. In addition, we also used restricted cubic spline model with three knots at the 10th, 50th, and 90th centiles to flexibly model the relationship between the overall LLDS and the incidence and mortality of lung cancer. The p-value for non-linearity was calculated by testing whether the regression coefficient of the second spline was equal to zero. Meanwhile, using the same methodology, we also examined the association between LLDS and specific lung cancer subtypes, including NSCLC and SCLC. Predetermined variables representing potential confounding factors was incorporated for conducting stratified analyses, including age (>65 years and ≤65 years), sex (male and female), smoking status (No and Current/former), family history of lung cancer (No and Yes/Possibly), BMI at baseline (>30 kg/m² and ≤30 kg/m²), aspirin use regularly (No and Yes), history of emphysema (No and Yes), history of bronchitis (No 11and Yes), and trail arm (Intervention and Control). Given that smoking is a significant risk factor for lung cancer, and smoking cessation can substantially reduce the risk of lung cancer mortality (25), we further categorized smoking status into three groups: never, current, and former. This classification enabled us to conduct a more in-depth investigation of the interaction effect of smoking on the association between LLDS and lung cancer. The significance of the multiplicative interaction between the above stratification factors and LLDS was examined using likelihood ratio tests.

Finally, a sequence of sensitivity analyses were conducted to reinforce the stability of the results: (1) exclusion of individuals with extreme BMI (top 1% and bottom 1%); (2) exclusion of individuals who experienced the outcome (including incidence and mortality of lung cancer) within the initial 2 or 4 years of follow-up, which may have partially mitigated the potential for reverse causality effects; (3) exclusion of individuals with respiratory comorbidities such as emphysema and chronic bronchitis, as these conditions may increase the risk of lung cancer (26); (4) adjustment for daily cigarette consumption (0, 1–20, or >20) instead of pack-years, to enhance the statistical power of the analysis; (5) excluded participants with missing data to determine whether the obtained results were influenced by the imputation of missing data; (6) HEI-2015 scores was used in place of LLDS for calculations to test the effectiveness of LLDS in assessing dietary quality.

All statistical analyses were performed using R software version 4.3.2, with a significance level set at p < 0.05.

This study encompassed 98,459 participants, with an average age of 65.5 years (standard deviation of 5.7 years). Participants were stratified based on their LLDS into four quartiles: Quartile 1 (0–19), Quartile 2 (20–24), Quartile 3 (25–29), and Quartile 4 (30–45), as detailed in Table 1. Those in the highest LLDS quartile typically exhibited characteristics such as a higher likelihood of being female, possessing greater educational levels, and a lower prevalence of smoking. Additionally, they had lower BMI and reduced energy intake. Regarding medical histories, the highest quartile had fewer instances of diabetes, hypertension, and emphysema compared to the lowest quartile. Notably, the increase in LLDS quartiles corresponded with higher HEI-2015 scores, validating LLDS as an effective measure of dietary quality against the widely recognized HEI-2015 standard.

During the mean follow-up period of 8.8 years, a total of 1,642 new lung cancer cases were diagnosed, with an incidence of 0.19 cases per 100 person-years. Among the confirmed cases, there were 1,408 instances of NSCLC and 234 cases of SCLC. In the fully adjusted multivariable model, individuals in the highest quartile of LLDS demonstrated a significantly lower incidence of lung cancer compared to those in the lowest quartile. HRs indicated inverse relationships across different types of lung cancer: for overall lung cancer, the HR_Q4:Q1 was 0.80 (95% CI = 0.68–0.94, P for trend = 0.003), and for SCLC, the HR_Q4:Q1 was 0.55 (95% CI = 0.35–0.87, P for trend = 0.002). Although the HR for NSCLC indicated a reduced risk at 0.85 (95% CI = 0.72–1.01), this association did not achieve statistical significance (P for trend = 0.063) (Table 2). When LLDS is analyzed as a continuous variable, a 1-point increment in LLDS was associated with a 1% decrease in the incidence of overall lung cancer, with a specific reduction of 3% for SCLC. Further analysis using restricted cubic spline models revealed an inverse linear dose–response relationship between the full LLDS and the incidence of lung cancer (P for non-linearity = 0.335) and SCLC (P for non-linearity = 0.548), as depicted in Figure 1.

Over a 15.1-year follow-up, the study documented 1,172 lung cancer-related deaths, comprising 967 NSCLC and 205 SCLC cases. Multivariable analyses revealed that individuals in the highest LLDS quartile had significantly lower mortality for both overall lung cancer (HR_Q4:Q1 = 0.81, 95% CI = 0.67–0.98, P for trend = 0.009) and SCLC (HR_Q4:Q1 = 0.42, 95% CI = 0.25–0.70, P for trend <0.001), as detailed in Table 3. No significant effect on NSCLC mortality was observed (HR_Q4:Q1 = 0.91, 95% CI = 0.74–1.11, P for trend = 0.268). For 1-point increment in LLDS, there was a corresponding 1% reduction in overall lung cancer mortality and a 4% decrease in SCLC mortality. Restricted cubic spline analysis confirmed an inverse linear dose–response relationship between LLDS and mortality for both overall lung cancer and SCLC, with respective P for non-linearity values of 0.812 and 0.315, as depicted in Figure 1.

The association between the LLDS and both the incidence and mortality of lung cancer is not affected by pre-defined potential effect modifiers (all P_interaction > 0.05) (Figures 2, 3). When we conducted subgroup analyses using a more refined classification of smoking status (never, current, and former smokers), this inverse relationship is not affected by the refined smoking status (P_interaction > 0.05) (Supplementary Table S5). The sensitivity analyses conducted reinforced the robustness of the association between LLDS and lung cancer outcomes (Supplementary Tables S2, S3). Key adjustments included the exclusion of individuals with extreme BMI values and those who experienced the outcome within the first 2–4 years of follow-up. Additionally, excluding participants with respiratory comorbidities such as emphysema and chronic bronchitis did not alter the results. Further adjustments for daily cigarette consumption instead of pack-years, and excluded participants with missing data, also supported the original findings. To further validate the effectiveness of LLDS as a measure of dietary quality, we substituted LLDS with the widely validated dietary assessment tool, HEI-2015. The results still demonstrated a relationship with higher HEI-2015 and decreased lung cancer incidence and mortality. This comparison underscores the reliability of LLDS in assessing dietary quality and its relationship with lung cancer incidence and mortality.