Study population of the current study consisted of the participants from the PLCO Cancer Screening Trial, a large multicenter randomized clinical trial that aimed to examine the possible benefits of screening tests or exams in reducing cancer-specific death from these prostate, lung, colorectal, and ovarian cancers. The corresponding study protocol has been published previously (12). In brief, between November 1993 and September 2001, 10 study centers invited American persons between the ages of 55 and 74 to participate in the PLCO Cancer Screening Trial. A total of 154,887 persons were finally included, and they were then randomly divided into the control arm or the screening arm. Participants in the screening arm received the selected screening tests or exams, whereas those in the control arm received the usual care. The PLCO Cancer Screening Trial had been ethically approved by the National Cancer Institute and the Institutional Review Committee of each screening center. Written informed consent was available for each participant.

We took several steps to identify our study population. First, we identified all patients receiving a diagnosis of cancer during the trial (i.e., from trial entry to December 31, 2009) (n = 29,225). Then, we excluded the following patients: (1) those who did not to finish a diet history questionnaire (DHQ) (n = 7,092); (2) those who had an invalid DHQ (n = 1,021) (11); (3) those who did not return a baseline questionnaire (n = 363); and (4) those who had been diagnosed with cancer before the DHQ completion (n = 5,645). After exclusions, we obtained a cohort of cancer patients completing the DHQ before the diagnosis of the first cancer (n = 15,104), which was called hereafter the source population. Finally, we identified patients with colorectal (n = 1,100), lung (n = 1705), prostate (n = 4,336), or breast cancer (n = 2,443) from the source population (Supplementary Figure S1). Of note, all included cancer patients were followed up through December 31, 2018; the long follow-up duration allows us to obtain a large number of outcome events of interest and provides an ideal opportunity to determine the potential impacts of UPFs on cancer-related mortality. Importantly, standardized differences of most baseline characteristics between the source and excluded populations were found to be <0.1, reminding us that there was a small possibility of nonparticipation bias, although up to 14,121 participants were excluded (Supplementary Table S1).

Dietary assessment was conducted using the above-mentioned DHQ. The DHQ, a self-administered questionnaire, contains a total of 124 food item and is developed for each participant to collect the frequency and serving size of his or her food consumption during the past year. The performance of this questionnaire had been validated in terms of dietary assessment (13). We approximated daily food consumption by directly multiplying food frequency by serving size, and we estimated daily intakes of nutrients and energy through the DietCalc software (14). Healthy Eating Index-2015, a frequently used diet quality indicator, was calculated for each participant following the method described previously (15). Western diet score was calculated based on the reported method (16) and data availability in the PLCO Cancer Screening Trial, and was defined by eight food groups, namely sugar, animal fat, butter, margarine, eggs, chips, red and processed meat, and salad dressings.

The method used to assess UPF consumption has been mentioned in the literatures (11, 17). Briefly, based on the NOVA classification method (3), two experienced dietitians classified all food and drink items in the DHQ into the four food groups. The specific definition and example for each food group are provided in the literature (3). Our study concentrated on UPFs (group 4), which consist of 64 individual food items (Supplementary Table S2). As per an established categorization method (18), these individual food items were further divided into the following nine food subgroups: ultra-processed fruits and vegetables, cereals, sauces and dressings, meat and meat products, soft drinks, margarine, ultra-processed dairy products, salty snacks, and sugary products.

The overall UPF consumption for a given patient was calculated by summing the amounts consumed of the above-mentioned 64 individual food items. UPF consumption was expressed as servings per day in main analyses primarily based on the USDA Pyramid Servings Database (19), considering different water contents across different food items. Because the consumption of almost all UPFs was initially estimated and recorded in daily grams in the PLCO Cancer Screening Trial, thus UPF consumption was also expressed as grams per day in supplementary analyses for comparison with the results from main analyses. For determining the possible influence of body size, we examined the associations of interest by expressing UPF consumption as daily serving/kilogram body weight. Before the formal data analyses, UPF consumption was adjusted for dietary energy intake with the residual method (20). Notably, given the potential impacts of cancer diagnosis on dietary behaviors, we used UPF consumption before the diagnosis of the first cancer to represent UPF consumption in a patient’s daily life and to perform all data analyses, which exclude the potential that our observed associations are actually caused by reverse causation.

Mortality status of each cancer patient was ascertained predominantly via an annual study update form. For those who did not return this form, repeat contacts were performed using e-mail or telephone. The ascertainment of mortality status was further supplemented by linkage to the National Death Index. Copies of death certificates were collected for died patients and were used as the primary source of dates and underlying causes of death. For the PLCO cancers, the relevant medical records were additionally reviewed to determine underlying causes of death. The ninth revision of International Classification of Diseases was used for coding the causes of death: colorectal cancer (codes 153.0–154.1 and 209.10–209.17), lung cancer (codes 162.2–162.5, 162.8, and 162.9), prostate cancer (codes 185), and breast cancer (codes 174).

Colorectal, lung, prostate, and breast cancers were primarily ascertained through an annual study update form, which was sent to each alive participant for asking if they had received a diagnosis of cancer, and if so, the date and place of diagnosis, and the site and type of cancer. Reports of cancer from the annual study update form were further validated by checking any available medical records. Of note, in the PLCO Cancer Screening Trial, cancer ascertainment also used the data from death certificates and family reports.

Most baseline characteristics shown in Table 1 were assessed using a self-administered baseline questionnaire. Body mass index (BMI) was calculated by dividing body weight (kg) by height squared (m²). Aspirin use referred to regularly taking aspirin or aspirin-containing drugs over the past year. Age at diagnosis and data on treatment, staging, and diagnostic were extracted from patient’s medical records. Notably, data were extracted only for treatments patients received within 1 year of cancer diagnosis. Cancer staging was performed using the AJCC 7th edition staging manual (21). Alcohol consumption was assessed using the DHQ. Physical activity level was expressed as total time of moderate-to-vigorous activity each week, which was estimated based on the information from a self-report supplemental questionnaire.

Missing data were imputed using the methods described as below. Categorical and continuous covariates with <10% missing data were imputed with the modal value and the median, respectively; the covariate “physical activity level,” which had 32.72% missing data, was imputed with multiple imputation by chained equations (the number of imputations = 25) under the assumption that these data were missing at random (22). Supplementary Tables S3, S4 present the distribution of covariates with missing data before and after imputation in the source population and the included cancer patients, respectively.

We used Cox proportional hazards regression model to calculate hazard ratios (HRs) and their 95% confidence intervals (CIs) of all-cause and cancer-specific mortality in relation to prediagnosis UPF consumption. In this model, follow-up duration was treated as time metric, and was calculated from the date of cancer diagnosis to loss to follow-up, death date, or the end of follow-up, whichever came earlier (Supplementary Figure S2). We examined the proportional hazard assumption using the Schoenfeld residuals (23); as the exposure variable “prediagnosis UPF consumption” was indicated to violate this assumption in analyzing its association with all-cause mortality among breast cancer patients (P for global test = 0.020), thus time-dependent Cox regression was used to calculate the corresponding HRs and 95% CIs. To evaluate the potential effects of competing risk bias on the associations of interest, we used competing risk regression to calculate subdistribution HRs and 95% CIs, with other causes of death than death from that cancer studied as competing events. In regression models, prediagnosis UPF consumption was split into quartiles, with the first quartile as the reference group. For examining the linear trends in effect sizes across quantiles, we first assigned the median of each quartile to each patient in that quartile to yield an ordinal variable and then regarded it as a continuous variable in regression models, with its P indicating the significance of linear trends. Notably, we chose covariates controlled in multivariable regression models using our causal knowledge of the current literature instead of statistical criteria (24). Specifically, model 1 controlled for age at diagnosis, sex (only for colorectal and lung cancers), race/ethnicity; and model 2 additionally controlled for trial arm, BMI, physical activity, alcohol consumption, smoking status, aspirin use, energy intake from diet, family history of each cancer we studied, history of diabetes, history of hypertension (only for all-cause mortality), and clinical covariates (mainly including cancer stage and treatments, see footnotes of the relevant tables for the exact list of covariates). We used Kaplan–Meier curves to show the cumulative incidence of cancer-related deaths by quarters of prediagnosis UPF consumption. To provide more stable estimates and lower random variability, we classified patients in the first and second quartiles into one group and those in the third and fourth quartiles into another group (25). The difference in cumulative incidence between groups was compared using the log-rank test.

We used restricted cubic spline regression to explore the potential dose–response associations between prediagnosis UPF consumption and all-cause and cause-specific mortality, with the reference level set at 0 servings/day. We used the Akaike’s information criterion and the Bayesian information criterion to determine the number of knots, with the lowest penalized likelihood suggesting the best fitted model. Thus, four knots located at 5th, 35th, 65th, and 95th percentiles were used in exploring dose–response associations with all-cause mortality in prostate cancer patients and with breast-cancer specific mortality in breast cancer patients, whereas three knots located at 10th, 50th, and 90th percentiles were used in exploring other dose–response associations (Supplementary Table S5). A P_nonlinearity was calculated by testing the null hypothesis that the regression coefficient(s) of the second spline (for three knots) or the second and third splines (for four knots) equal(s) to zero.

We performed a series of sensitivity analyses to determine the stability of our results: (1) excluding patients whose colorectal, lung, prostate, or breast cancer was not the first diagnosed cancer; (2) excluding patients whose colorectal, lung, prostate, or breast cancer was diagnosed ≤2 years after dietary assessment to test the potential influence of the potential reverse causation; (3) excluding patients who died within 30 days or 90 days after cancer diagnosis; (4) excluding patients with extreme UPF consumption (top 2.5% or bottom 2.5%); (5) excluding patients with extreme energy intake (26); (6) repeating the analysis with sex-specific quartiles, since the distribution of UPF consumption was found to be significantly different by sex; (7) additionally adjusting for intakes of fruit, vegetable, coffee, dairy, fish, whole grain, and red and processed meat or intakes of dietary fiber, added sugar, saturated fatty acids, and polyunsaturated fatty acids on model 2; (8) further adjusting for Healthy Eating Index-2015 or Western diet score on model 2 to test whether the observed associations were mediated by diet quality; and (9) additionally adjusted for glycemic index or glycemic load on model 2 to test whether the observed associations were influenced by dietary sugar intake.

We performed several prespecified subgroup analyses to explore whether the observed associations were modified by age at diagnosis (>69 vs. ≤65 years), sex (males vs. females, only in colorectal and lung cancer patients), BMI (≥25 vs. <25), current or formers smokers stopping smoking ≤15 years (yes vs. no), trial arm (screening vs. control arms), and cancer stage (stages I and II vs. stages III and IV). A P_interaction was obtained via comparing regression models with and without interaction terms prior to the formal subgroup analyses for avoiding the potentially spurious subgroup differences.

To identify the main driver(s) to the observed associations, we examined the associations of each food subgroup consumption with all-cause and cause-specific mortality. Statistical analyses were completed using STATA software version 12.0 (StataCorp). Two-sided p < 0.05 was considered statistically significant.

Regardless of cancer site, in most patients, UPFs contributed to 20–40% of total energy intake before cancer diagnosis (Supplementary Figure S3). Overall, compared with patients in the lowest quartile of prediagnosis UPF consumption, those in the highest quartile had younger age at cancer diagnosis, higher BMI and dietary energy intake while lower alcohol intake and Healthy Eating Index-2015, had a higher possibility of being Non-Hispanic White and present smokers and having histories of hypertension and diabetes; also, patients in the highest vs. the lowest quartiles of prediagnosis UPF consumption had higher intakes of vegetables, dairy, fish, whole grain, red and processed meat, dietary fiber, added sugars, as well as saturated and polyunsaturated fatty acids (Table 1). Supplementary Table S6 presents cancer characteristics and treatment information of included patients.

We observed 643 all-cause deaths and 324 colorectal cancer deaths in colorectal cancer patients during an average follow-up of 8.04 years, 1,525 all-cause deaths and 1,272 lung cancer deaths in lung cancer patients during an average follow-up of 2.95 years, 1,634 all-cause deaths and 254 prostate cancer deaths in prostate cancer patients during an average follow-up of 10.76 years, and 755 all-cause deaths and 189 breast cancer deaths in breast cancer patients during an average follow-up of 10.89 years. After fully adjusting for the potential confounders, high UPF consumption before cancer diagnosis was significantly associated with an elevated risk of all-cause mortality in lung (HR_{quartile 4 vs. 1}: 1.18; 95% CI: 0.98, 1.40; P_trend = 0.021) and prostate (HR_{quartile 4 vs. 1}: 1.18; 95% CI: 1.00, 1.39; P_trend = 0.017) cancer patients (Table 2). When prediagnosis UPF consumption was expressed as daily gram or daily serving/kilogram body weight, the initial results did not change substantially (Supplementary Tables S7, S8). After considering the potential competing risk bias, we obtained similar results for cancer-specific mortality (Supplementary Table S9). In addition, high UPF consumption before cancer diagnosis was found to confer increased risks of lung cancer-specific (HR_{quartile 4 vs. 1}: 1.13; 95% CI: 0.93, 1.36) and prostate cancer-specific (HR_{quartile 4 vs. 1}: 1.30; 95% CI: 0.84, 2.01) mortality, although the linear trend tests did not reach statistical significance (P_trend = 0.137 for lung cancer-specific mortality and P_trend = 0.112 for prostate cancer-specific mortality) (Table 2).

Kaplan–Meier curves showed that the cumulative incidence of death from all causes was higher among lung (p = 0.032) and prostate (p = 0.049) cancer patients in the third and fourth quartiles of prediagnosis UPF consumption compared with those in the first and second quartiles, while no significant differences were found for other outcomes of interest (Figure 1).

A nonlinear dose–response trend was found for the association of prediagnosis UPF consumption with all-cause mortality in lung (P_nonlinearity = 0.022) and prostate (P_nonlinearity = 0.042) cancer patients, while such a trend was not found for other associations of interest (Figure 2). The initially observed associations did not change substantially in a wide range of sensitivity analyses (Supplementary Table S10). Interestingly, prediagnosis UPF consumption was positively associated with colorectal cancer-specific mortality among colorectal cancer patients in stages I and II (HR_{quartile 4 vs. 1}: 2.52; 95% CI: 1.10, 5.80; P_trend = 0.033) but not among those in stages III and IV (HR_{quartile 4 vs. 1}: 0.96; 95% CI: 0.64, 1.45; P_trend = 0.674) (P_interaction = 0.006) (Supplementary Figure S4); moreover, prediagnosis UPF consumption was positively associated with prostate cancer-specific mortality among prostate cancer patients with BMI <25 (HR_{quartile 4 vs. 1}: 3.08; 95% CI: 1.25, 7.60; P_trend = 0.004) but not among those with BMI ≥25 (HR_{quartile 4 vs. 1}: 1.16; 95% CI: 0.69, 1.95; P_trend = 0.577) (P_interaction = 0.001) (Supplementary Figure S5). No significant interaction was detected in other subgroup analyses (Supplementary Figures S4–S7).

Overall, in our study population, main food subgroups contributing to total serving size of UPFs were cereals, soft drinks, meat and meat products, and margarine (Supplementary Figure S8). Higher consumption of sauces and dressings (HR_{quartile 4 vs. 1}: 1.17; 95% CI: 1.01, 1.36; P_trend = 0.046) and ultra-processed fruits and vegetables (HR_{quartile 4 vs. 1}: 1.15; 95% CI: 0.99, 1.33; P_trend = 0.016) before cancer diagnosis conferred a higher risk of all-cause mortality in lung cancer patients, whereas higher consumption of soft drinks (HR_{quartile 4 vs. 1}: 1.11; 95% CI: 0.96, 1.28; P_trend = 0.042), meat and meat products (HR_{quartile 4 vs. 1}: 1.16; 95% CI: 1.00, 1.33; P_trend = 0.019), and ultra-processed fruits and vegetables (HR_{quartile 4 vs. 1}: 1.14; 95% CI: 1.00, 1.31; P_trend = 0.041) conferred a higher risk of all-cause mortality in prostate cancer patients (Supplementary Table S11). Moreover, prediagnosis consumption of sauces and dressings (HR_{quartile 4 vs. 1}: 1.23; 95% CI: 1.05, 1.44; P_trend = 0.033) was positively associated with the risk of lung cancer-specific mortality, and prediagnosis consumption of meat and meat products (HR_{quartile 4 vs. 1}: 1.60; 95% CI: 1.08, 2.36; P_trend = 0.010) and sugary products (HR_{quartile 4 vs. 1}: 1.41; 95% CI: 1.00, 2.00; P_trend = 0.036) was positively associated with the risk of prostate cancer-specific mortality.

In addition, prediagnosis consumption of ultra-processed fruits and vegetables was positively associated with all-cause (HR_{quartile 4 vs. 1}: 1.16; 95% CI: 0.91, 1.49) and colorectal cancer-specific mortality (HR_{quartile 4 vs. 1}: 1.30; 95% CI: 0.92, 1.85) in colorectal cancer patients, with marginal significance in linear trend tests (both P_trend = 0.098). Similarly, prediagnosis consumption of meat and meat products (HR_{quartile 4 vs. 1}: 1.25; 95% CI: 0.88, 1.79; P_trend = 0.053) and salty snacks (HR_{quartile 4 vs. 1}: 1.31; 95% CI: 0.92, 1.87; P_trend = 0.064) was positively associated with colorectal cancer-specific mortality. Also, despite the lack of statistical significance in linear trend tests (all P_trend > 0.05), higher consumption of ultra-processed fruits and vegetables conferred an increased risk of lung cancer-specific mortality (HR_{quartile 4 vs. 1}: 1.10; 95% CI: 0.93, 1.29), and higher consumption of meat and meat products conferred increased risks of all-cause (HR_{quartile 4 vs. 1}: 1.18; 95% CI: 0.94, 1.49) and breast cancer-specific (HR_{quartile 4 vs. 1}: 1.35; 95% CI: 0.87, 2.11) mortality in breast cancer patients.