Low-carbohydrate diets and the risk of pancreatic cancer: a large prospective cohort study

Abstract

Low-carbohydrate diets have become a popular approach for weight loss in recent years. However, whether low-carbohydrate diets are associated with the risk of pancreatic cancer remains to be elucidated. Hence, we examined the association of low-carbohydrate diets with the risk of pancreatic cancer in a US population. A population-based cohort of 95 962 individuals was identified. A low-carbohydrate-diet score was calculated to quantify adherence to this dietary pattern, with higher scores indicating greater adherence. Cox regression was used to calculate risk estimate for the association of the low-carbohydrate-diet score with the risk of pancreatic cancer. Subgroup analysis was used to identify the potential effect modifiers. After an average follow-up of 8.87 years (875856.9 person-years), we documented a total of 351 pancreatic cancer cases. In the fully adjusted model, the highest versus the lowest quartiles of the overall low-carbohydrate-diet score were found to be associated with a reduced risk of pancreatic cancer (hazard ratio_{quartile 4} versus ₁: 0.61; 95% confidence interval: 0.45, 0.82; P_trend < 0.001). Subgroup analysis found that the inverse association of low-carbohydrate diets with the risk of pancreatic cancer was more pronounced in individuals aged ≥65 years than in those aged <65 years (P_interaction = 0.015). Similar results were obtained for animal and vegetable low-carbohydrate-diet scores. In conclusion, low-carbohydrate diets, regardless of the type of protein and fat, are associated with a lower risk of pancreatic cancer in the US population, suggesting that adherence to low-carbohydrate diets may be beneficial for pancreatic cancer prevention. Future studies should validate our findings in other populations.

Introduction

Pancreatic cancer is an extremely lethal malignancy, with a total of 45 8918 incident cases and 43 2242 million cancer deaths worldwide in 2018 (1). In the USA, it is the fourth most common cause of cancer-associated mortality, with an estimated 47 050 deaths in 2020 (2). The etiology of pancreatic cancer remains unclear, but dietary habits have been suggested to play an important role (3). Indeed, epidemiological studies have revealed some dietary factors in relation to the risk of pancreatic cancer, such as intakes of coffee and folate (4,5). However, the World Cancer Research Fund/American Institute for Cancer Research suggested in their 2018 report that there was limited evidence to support a causal association of diets with the risk of pancreatic cancer (6), indicating that more studies are warranted to investigate this association.

Low-carbohydrate diets, which are low in carbohydrate and high in fat and protein, have become a particularly popular approach for weight reduction in recent years (7). Observational studies have found that low-carbohydrate diets are associated with increased risks of type 2 diabetes (8), atrial fibrillation (9) and mortality (10) but a reduced risk of cardiovascular disease (11). However, the effects of low-carbohydrate diets on carcinogenesis remain largely unknown. As glucose is the main source of energy of cancer cells, carbohydrate-restricted diets are thus expected to have a protective role in cancer development. Indeed, several observational studies have revealed inverse associations between overall or plant-based low-carbohydrate diets and risks of breast cancer (12), glioma (13) and hepatocellular carcinoma (14). To our knowledge, only one study in Sweden has investigated the association between low-carbohydrate diets and the risk of pancreatic cancer, with a null association observed (15). However, this study only considered intakes of carbohydrate and protein in the calculation of low-carbohydrate-diet score and documented limited pancreatic cancer cases (n = 70), raising a possibility that it had insufficient power to detect the potential significant association. Moreover, whether the results from European population can be extended to other populations needs to be further confirmed.

Given the high popularity of low-carbohydrate diets and the dismal prognosis of pancreatic cancer, clarifying the potential association of low-carbohydrate diets with the risk of pancreatic cancer will have important public health implications. Hence, we conducted a prospective study to determine whether low-carbohydrate diets are associated with a reduced risk of pancreatic cancer in a US population.

Materials and methods

Study population

Approximately 15 5000 individuals between the ages of 55 and 74 were enrolled to the Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial between 1993 and 2001. The PLCO Cancer Screening Trial is a multicenter randomized controlled study, which was designed to investigate the potential beneficial effects of selected screening examinations on the risks of death from prostate, lung, colorectal and ovarian cancers. Study design of this trial has been described previously (16). All individuals provided their written informed consents, and the trial was approved by the Institutional Review Boards at 10 screening centers and the US National Cancer Institute.

The following individuals were not eligible for the present study: (i) individuals failing to return a baseline questionnaire (n = 4918); (ii) individuals with an invalid Diet History Questionnaire (DHQ), which is defined as death date prior to DHQ completion date, missing DHQ completion date and the presence of eight or more missing DHQ items or extreme energy intakes (>99th percentile or <1st percentile) (n = 38 462); (iii) individuals diagnosed with any cancer before randomization or DHQ completion (n = 9684); (iv) individuals with a diagnosis of primary adenocarcinoma of the endocrine pancreas (n = 19); (v) individuals receiving a diagnosis of pancreatic cancer between randomization and DHQ completion (n = 75); (vi) individuals whose pancreatic cancer was not the first diagnosed cancer (n = 42) and (vii) individuals with extreme values of energy intake [i.e. <800 or >4000 kcal/day for male and <500 or >3500 kcal/day for female (17)] (n = 2887). Finally, a total of 98 800 individuals were included in this study (Supplementary Figure 1, available at Carcinogenesis Online). Of note, all participants were followed up from DHQ completion to the occurrence of pancreatic cancer, death, loss to follow-up or the end of follow-up (i.e., 31 December 2009), whichever came first (Supplementary Figure 2, available at Carcinogenesis Online).

Calculation of the low-carbohydrate-diet score

We calculated a score to quantify adherence to low-carbohydrate diets following a method proposed by the Halton et al. (18). Briefly, we classified participants into 11 strata according to the percentage of energy from carbohydrate, protein and fat. For carbohydrate, participants in the lowest stratum earned 10 points and those in the highest stratum earned 0 points; in contrast, for fat and protein, participants in the lowest stratum earned 0 points and those in the highest earned 10 points (Supplementary Table 1, available at Carcinogenesis Online). The points for each of macronutrients were then summed to calculate an individual’s overall low-carbohydrate-diet score, which ranges from 0 to 30 points. Higher scores indicate greater adherence to low-carbohydrate diets. Similarly, we calculated an animal low-carbohydrate-diet score based on the percentage of energy from carbohydrate, animal protein and animal fat, as well as a vegetable low-carbohydrate-diet score based on the percentage of energy from carbohydrate, vegetable protein and vegetable fat for each participant. The information used for the calculation of low-carbohydrate-diet score was obtained from the DHQ. The DHQ is a 137-item food frequency questionnaire developed for assessing the consumption of foods and supplements during the past year. Importantly, dietary information collected with the DHQ was not cumulatively updated during follow-up; instead, all participants were asked only once at baseline for their dietary information. Of note, the reproducibility and validity of the DHQ have been confirmed elsewhere, indicating that the DHQ enjoyed good performance in assessing dietary consumption (19).

Assessment of incident pancreatic cancer

Incident pancreatic cancer was confirmed through a mailed annual study update form that asked participants to answer whether they were diagnosed with cancer, the date and site of cancer diagnosis and the type of cancer. Participants who did not return the form were contacted again via e-mail or telephone. Death certificates and family reports were used as supplementary files for case ascertainment. Relevant medical records were checked carefully for the further confirmation of pancreatic cancer cases. Importantly, in this study, pancreatic cancer was defined as primary adenocarcinoma of the exocrine pancreas (ICD-O-2 codes: C25.0-C25.3, C25.7-C25.9).

Assessment of covariates

Age at DHQ completion, single or multivitamin supplement use, and alcohol and food consumption were collected with the DHQ. Daily food consumption was calculated by multiplying the food frequency by portion size; daily nutrient intake was estimated using the USDA’s 1994–96 Continuing Survey of Food Intakes by Individuals (20) and the Nutrition Data Systems for Research (21). Physical activity level was estimated based on the frequency and duration of moderate and strenuous activities that were collected using a self-administrated supplemental questionnaire. Other covariates, including sex, race, history of diabetes, family history of pancreatic cancer, aspirin use, body weight, and height, were collected using a self-administrated baseline questionnaire. Body mass index (BMI) was calculated as body weight (kg) divided by height squared (m²).

Statistical analysis

For covariates with <5% missing values, we used the modal value and median to impute missing data of categorical and continuous variables, respectively; for the covariate ‘physical activity level’ with 25% missing values, to decrease selection bias and increase statistical power, we assumed that these data were missing at random, and then used multiple imputation with chained equations to impute missing data (the number of imputations = 25) (22). The distribution of variables with missing data before and after imputation is shown in Supplementary Table 2 (available at Carcinogenesis Online). All variables involved in statistical analyses were used to produce the imputed data sets.

Cox proportional hazards regression was performed to calculate hazard ratios (HRs) and 95% confidence intervals (CIs) for the association of low-carbohydrate diets and the risk of pancreatic cancer, with follow-up length as time metric. No evidence for the violation of proportional hazards assumption was detected with Schoenfeld residuals approach (P values for global test: 0.335 for overall low-carbohydrate-diet score, 0.535 for animal low-carbohydrate-diet score and 0.613 for vegetable low-carbohydrate-diet score); hence, in Cox regression models, low-carbohydrate-diet score and all covariates were not treated as time-varying variables. In regression models, low-carbohydrate-diet score was split into quartiles, with the first quartile as the reference group. To determine whether a linear trend existed across quartiles of low-carbohydrate-diet score, we first assigned the median of each quartile to each individual in the quartile, and then treated it as a continuous variable in regression models. Covariate selection in multivariable analyses was based on the existing literature and the change-in-estimate approach (23). Specifically, model 1 was adjusted for age, sex and race; model 2 was additionally adjusted for educational level, smoking status, alcohol consumption, physical activity level, BMI, aspirin use, single or multivitamin supplement use, history of diabetes, family history of pancreatic cancer and energy intake from diet.

We employed restricted cubic spline models (24) with three knots located at the 10th, 50th and 90th percentiles to provide a thorough description for pancreatic cancer risk across the full range of the low-carbohydrate-diet score. For comparison with the results from Cox proportional hazards regression, we used the median of the first quartile of the low-carbohydrate-diet score as the reference level. A P_{non-linearity} was estimated by testing the null hypothesis that the regression coefficient of the second spline equals 0 (24). A series of subgroup analyses were conducted after stratifying for age (≥65 versus <65 years), sex (male versus female), BMI (≥25 versus <25 kg/m²), aspirin use (yes versus no), alcohol consumption (≥ median versus <median), smoking status (current or former versus never) and dietary fiber consumption (≥ median versus <median). We employed a likelihood ratio test to examine the significance of interaction between low-carbohydrate-diet score and stratified factors. We conducted the following sensitivity analyses to examine the stability of our results: (i) excluded participants with a history of diabetes, as they might have a more favorable diet in terms of low-carbohydrate diets; (ii) excluded cases observed within the first 2 and 4 years of follow-up to evaluate the potential influence of reverse causation; (iii) repeated the analysis in participants with complete data and (iv) further adjusted for consumption of fruits, coffee, tea and eggs or intakes of dietary fiber, folate, calcium, magnesium, iron and potassium on model 2.

Statistical analyses were conducted with STATA version 12.0 (StataCorp, College Station, TX). The results were considered statistically significant when a two-tailed P value was <0.05.

Results

Participant characteristics

In the whole study population, the mean (SD) was 15.0 (7.1) for the overall low-carbohydrate-diet score, 15.0 (7.8) for the animal low-carbohydrate-diet score and 15.0 (5.9) for the vegetable low-carbohydrate-diet score. Compared with participants in the lowest quartile of the overall low-carbohydrate-diet score, those in the highest quartile were more likely to be male, non-Hispanic white, current or former smokers, and have a history of diabetes, had higher energy intake from diet, but were less likely to be single or multivitamin supplement users, and had a lower physical activity level (Table 1). Moreover, participants in the highest versus the lowest quartiles of the overall low-carbohydrate-diet score had higher intakes of red meat, coffee, tea, fish, eggs, calcium, magnesium, potassium but lower intakes of fruits, dietary fiber and folate.

Table 1.

Baseline characteristics of study population according to overall low-carbohydrate-diet score

		Quartiles of overall low-carbohydrate-diet score
Characteristics	Overall	≤9	10–14	15–19	≥20
Number of participants	98 800	23 266	23 950	23 644	27 940
Age (yeas)	65.5 ± 5.7	66.4 ± 5.9	65.9 ± 5.8	65.5 ± 5.7	64.4 ± 5.4
Male	46 859 (47.4)	9566 (41.1)	10 771 (45.0)	11 462 (48.5)	15 060 (53.9)
Race
Non-Hispanic white	90 042 (91.1)	19 964 (85.8)	22 039 (92.0)	22 065 (93.3)	25 974 (93.0)
Non-Hispanic black	3162 (3.2)	1209 (5.2)	691 (2.9)	561 (2.4)	701 (2.5)
Hispanic	1415 (1.4)	336 (1.4)	288 (1.2)	309 (1.3)	482 (1.7)
Othersa	4181 (4.2)	1757 (7.6)	932 (3.9)	709 (3.0)	783 (2.8)
Educational level
College below	62 918 (63.7)	14 805 (63.6)	14 862 (62.1)	15 002 (63.4)	18 249 (65.3)
College graduate	17 371 (17.6)	3833 (16.5)	4377 (18.3)	4297 (18.2)	4864 (17.4)
Postgraduate	18 511 (18.7)	4628 (19.9)	4711 (19.7)	4345 (18.4)	4827 (17.3)
Body mass index (kg/m2)	27.2 ± 4.8	26.2 ± 4.6	26.8 ± 4.6	27.3 ± 4.7	28.3 ± 5.0
Physical activity (min/week)b	121.7 ± 122.3	133.4 ± 128.9	124.8 ± 121.7	118.4 ± 119.3	112.1 ± 118.7
Energy intake from diet (kcal/day)	1712.5 ± 650.5	1535.2 ± 583.1	1659.4 ± 617.6	1761.1 ± 645.0	1864.6 ± 693.2
Smoking status
Current
>60 pack-years	2749 (2.8)	335 (1.4)	546 (2.3)	665 (2.8)	1203 (4.3)
30–60 pack-years	4117 (4.2)	679 (2.9)	809 (3.4)	1022 (4.3)	1607 (5.8)
<30 pack-years	2139 (2.2)	443 (1.9)	471 (2.0)	551 (2.3)	674 (2.4)
Former
>60 pack-years	5371 (5.4)	883 (3.8)	1182 (4.9)	1337 (5.7)	1969 (7.0)
30–60 pack-years	11 677 (11.8)	2230 (9.6)	2757 (11.5)	2837 (12.0)	3853 (13.8)
<30 pack-years	25 253 (25.6)	5883 (25.3)	6175 (25.8)	6086 (25.7)	7109 (25.4)
Never	47 494 (48.1)	12 813 (55.1)	12 010 (50.1)	11 146 (47.1)	11 525 (41.2)
Alcohol consumption (g/day)	8.6 ± 18.6	6.2 ± 13.8	11.9 ± 28.1	9.3 ± 16.6	7.3 ± 12.0
History of diabetes	6509 (6.6)	1061 (4.6)	1318 (5.5)	1537 (6.5)	2593 (9.3)
Family history of pancreatic cancer	2545 (2.6)	647 (2.8)	623 (2.6)	609 (2.6)	666 (2.4)
Aspirin use	46 309 (46.9)	10 665 (45.8)	11 199 (46.8)	11 091 (46.9)	13 354 (47.8)
Single or multivitamin supplement use	76 983 (77.9)	18 759 (80.6)	19 055 (79.6)	18 222 (77.1)	20 947 (75.0)
Trial arm
Intervention	50 346 (51.0)	11 338 (48.7)	12 016 (50.2)	12 307 (52.1)	14 685 (52.6)
Control	48 454 (49.0)	11 928 (51.3)	11 934 (49.8)	11 337 (47.9)	13 255 (47.4)
Food consumption
Fruits (g/day)	273.8 ± 212.1	392.1 ± 284.1	291.8 ± 187.7	243.4 ± 161.4	185.5 ± 138.4
Vegetables (g/day)	282.8 ± 180.9	283.7 ± 206.3	282.8 ± 181.2	282.7 ± 170.1	282.2 ± 166.5
Red meat (g/day)	60.2 ± 48.4	31.3 ± 24.0	47.1 ± 30.9	62.7 ± 38.9	93.3 ± 61.3
Dairy (cups/day)	1.4 ± 1.1	1.1 ± 0.9	1.4 ± 1.1	1.4 ± 1.1	1.5 ± 1.2
Coffee (g/day)	632.1 ± 777.2	482.0 ± 674.9	584.7 ± 738.5	674.4 ± 783.9	761.9 ± 854.5
Tea (g/day)	178.7 ± 398.8	177.4 ± 408.9	176.0 ± 387.2	180.0 ± 393.2	181.0 ± 404.9
Whole grains (servings/day)	1.2 ± 0.8	1.3 ± 0.9	1.2 ± 0.8	1.2 ± 0.8	1.0 ± 0.7
Fish (g/day)	15.4 ± 18.2	11.0 ± 12.7	14.0 ± 15.4	15.9 ± 18.2	19.8 ± 22.7
Eggs (g/day)	12.0 ± 15.7	6.2 ± 9.0	9.2 ± 11.3	12.2 ± 13.6	19.0 ± 21.2
Nutrient intake
Dietary fiber (g/day)	17.9 ± 8.0	19.1 ± 9.0	18.2 ± 8.0	17.6 ± 7.6	16.8 ± 7.4
Folate (mcg/day)	375.1 ± 157.6	389.4 ± 171.4	379.8 ± 157.5	372.3 ± 151.3	361.6 ± 149.4
Calcium (mg/day)	743.5 ± 385.3	668.3 ± 336.3	756.5 ± 388.9	766.0 ± 390.6	776.1 ± 407.2
Magnesium (mg/day)	319.7 ± 118.7	305.5 ± 116.2	318.6 ± 115.9	322.5 ± 115.9	330.0 ± 124.3
Iron (mg/day)	14.3 ± 5.9	14.2 ± 6.0	14.4 ± 5.9	14.4 ± 5.8	14.4 ± 5.8
Potassium (mg/day)	3227.5 ± 1180.6	3152.3 ± 1229.4	3217.1 ± 1156.2	3239.7 ± 1144.1	3288.8 ± 1186.6
Carbohydrate (% energy)	52.0 ± 9.4	63.1 ± 6.2	54.4 ± 6.3	49.7 ± 4.6	42.9 ± 5.6
Total protein (% energy)	15.4 ± 2.9	13.4 ± 2.2	15.0 ± 2.7	15.6 ± 2.6	17.4 ± 2.6
Animal protein (% energy)	3.9 ± 2.5	2.5 ± 1.5	3.3 ± 1.8	4.0 ± 2.2	5.4 ± 3.1
Vegetable protein (% energy)	1.7 ± 1.0	1.8 ± 1.2	1.7 ± 1.0	1.6 ± 0.9	1.6 ± 0.9
Total fat (% energy)	31.8 ± 7.5	24.2 ± 4.9	28.8 ± 5.1	33.6 ± 4.7	39.0 ± 5.3
Animal fat (% energy)	10.1 ± 4.3	6.6 ± 2.5	8.7 ± 2.7	10.5 ± 3.0	14.0 ± 4.5
Vegetable fat (% energy)	13.5 ± 5.2	11.4 ± 3.9	12.6 ± 4.4	14.2 ± 5.0	15.5 ± 5.9
Monounsaturated fat (% energy)	11.9 ± 3.1	8.9 ± 2.1	10.8 ± 2.1	12.7 ± 2.1	14.9 ± 2.4
Polyunsaturated fat (% energy)	7.3 ± 2.2	5.8 ± 1.6	6.7 ± 1.7	7.6 ± 1.9	8.6 ± 2.3
Saturated fat (% energy)	10.1 ± 3.0	7.4 ± 2.1	9.1 ± 2.2	10.7 ± 2.3	12.5 ± 2.5

Data are mean (SD) or number (percentage) as indicated.

^a“Others” refers to Asian, Pacific Islander or American Indian.

^bTotal time of moderate-to-vigorous physical activities per week.

Low-carbohydrate-diet score and pancreatic cancer risk

During 875856.9 person-years of follow-up, we documented a total of 351 pancreatic cancer cases, with the overall incidence rate of 4.01 cases per 10 000 person-years. The mean (standard deviation) follow-up length was 8.87 (1.90) years. In univariable analysis, participants in the highest versus the lowest quartiles of the overall low-carbohydrate-diet score were found to be at a reduced risk of pancreatic cancer (HR_{quartile 4 versus 1}: 0.65; 95% CI: 0.49, 0.87; P_trend = 0.001) (Table 2). After the full adjustment for potential confounders, the observed inverse association of the overall low-carbohydrate-diet score with the risk of pancreatic cancer remained (HR_{quartile 4 versus 1}: 0.61; 95% CI: 0.45, 0.82; P_trend < 0.001). Similar results were observed for animal and vegetable low-carbohydrate-diet scores.

Table 2.

Hazard ratios of the association of low-carbohydrate-diet scores with the risk of pancreatic cancer

			Hazard ratio (95% confidence interval)
Quartiles of low-carbohydrate-diet score	Number of cases	Person-years	Unadjusted	Model 1a	Model 2b
Overall low-carbohydrate-diet score
≤9	111	209305.6	1.00 (reference)	1.00 (reference)	1.00 (reference)
10–14	87	213595.7	0.77 (0.58, 1.02)	0.79 (0.60, 1.05)	0.74 (0.56, 0.99)
15–19	69	209677.4	0.62 (0.46, 0.84)	0.65 (0.48, 0.88)	0.59 (0.44, 0.81)
≥20	84	243278.2	0.65 (0.49, 0.87)	0.71 (0.53, 0.95)	0.61 (0.45, 0.82)
Ptrend			0.001	0.009	<0.001
Animal low-carbohydrate-diet score
≤8	98	209184.0	1.00 (reference)	1.00 (reference)	1.00 (reference)
9–14	89	212381.3	0.89 (0.67, 1.19)	0.91 (0.68, 1.21)	0.85 (0.63, 1.13)
15–20	73	215433.3	0.72 (0.53, 0.98)	0.74 (0.55, 1.01)	0.67 (0.49, 0.91)
≥21	91	238858.3	0.81 (0.61, 1.08)	0.85 (0.64, 1.14)	0.72 (0.53, 0.98)
Ptrend			0.092	0.180	0.018
Vegetable low-carbohydrate-diet score
≤10	100	202551.7	1.00 (reference)	1.00 (reference)	1.00 (reference)
11–14	89	206906.5	0.87 (0.65, 1.16)	0.90 (0.68, 1.20)	0.89 (0.67, 1.18)
15–18	81	213975.1	0.77 (0.57, 1.03)	0.81 (0.60, 1.09)	0.79 (0.59, 1.07)
≥19	81	252423.6	0.65 (0.48, 0.87)	0.72 (0.54, 0.96)	0.68 (0.51, 0.92)
Ptrend			0.003	0.021	0.010

^aAdjusted for age (years), sex (male, female) and race (non-Hispanic white, non-Hispanic black, Hispanic others).

^bAdjusted for model 1 plus educational level (college below, college graduate, postgraduate), smoking status [current (>60 pack-years, 30–60 pack-years, <30 pack-years), former (>60 pack-years, 30–60 pack-years, <30 pack-years), never], alcohol consumption (g/day), physical activity level (min/week), body mass index (kg/m²), aspirin use (yes, no), single or multivitamin supplement use (yes, no), history of diabetes (yes, no), family history of pancreatic cancer (yes, no), and energy intake from diet (kcal/day).

Additional analyses

Restricted cubic spline models found that the overall low-carbohydrate-diet score was inversely associated with the risk of pancreatic cancer in a non-linear dose–response manner (P_{non-linearity} = 0.012) (Figure 1). A similar dose–response pattern was found for animal (P_{non-linearity} = 0.029) and vegetable (P_{non-linearity} = 0.007) low-carbohydrate-diet scores. Subgroup analyses found that the observed inverse associations could not be modified by sex, BMI, aspirin use, alcohol consumption, smoking status and dietary fiber consumption (all P_interaction > 0.05) (Table 3). However, we observed that the inverse association of low-carbohydrate diets with the risk of pancreatic cancer was more pronounced in participants aged ≥65 years than in those aged <65 years (all P_interaction < 0.05). The primary associations of low-carbohydrate-diet score with pancreatic cancer risk did not change substantially in sensitivity analyses (Table 4).

Figure 1.

Dose–response analyses on the association of LCD scores with the risk of pancreatic cancer. Hazard ratio was adjusted for age, sex, race, educational level, smoking status, alcohol consumption, physical activity level, body mass index, aspirin use, single or multivitamin supplement use, history of diabetes, family history of pancreatic cancer, and energy intake from diet. LCD, low-carbohydrate diet.

Table 3.

Subgroup analyses on the association of low-carbohydrate-diet scores with the risk of pancreatic cancer

	Overall low-carbohydrate-diet score		Animal low-carbohydrate- diet score		Vegetable low-carbohydrate- diet score
Subgroup variable	HRquartile 4 versus 1 (95% CI)a	Pinteraction	HRquartile 4 versus 1 (95% CI)a	Pinteraction	HRquartile 4 versus 1 (95% CI)a	Pinteraction
Age (years)
≥65	0.49 (0.34, 0.72)	0.015	0.57 (0.39, 0.82)	0.012	0.52 (0.35, 0.80)	0.038
<65	1.03 (0.59, 1.81)		1.31 (0.72, 2.36)		1.00 (0.61, 1.66)
Sex
Male	0.54 (0.36, 0.80)	0.380	0.69 (0.46, 1.05)	0.558	0.57 (0.38, 0.87)	0.346
Female	0.69 (0.43, 1.10)		0.78 (0.49, 1.23)		0.83 (0.54, 1.29)
Body mass index (kg/m2)
≥25	0.69 (0.48, 1.00)	0.261	0.76 (0.52, 1.10)	0.652	0.69 (0.48, 1.00)	0.951
<25	0.49 (0.28, 0.85)		0.65 (0.37, 1.12)		0.69 (0.41, 1.16)
Aspirin use
Yes	0.49 (0.31, 0.76)	0.403	0.71 (0.46, 1.10	0.645	0.79 (0.51, 1.23)	0.470
No	0.73 (0.48, 1.10)		0.73 (0.48, 1.12)		0.61 (0.40, 0.91)
Alcohol consumption (g/day)
≥ median	0.68 (0.42, 1.09)	0.450	0.83 (0.51, 1.34)	0.721	0.74 (0.47, 1.18)	0.645
< median	0.59 (0.39, 0.87)		0.70 (0.47, 1.06)		0.68 (0.46, 1.02)
Smoking status
Current or former	0.58 (0.39, 0.86)	0.934	0.74 (0.50, 1.10)	0.993	0.66 (0.44, 0.99)	0.817
Never	0.70 (0.44, 1.13)		0.75 (0.47, 1.22)		0.73 (0.47, 1.13)
Dietary fiber consumption (g/day)
≥ median	0.49 (0.31, 0.78)	0.133	0.64 (0.40, 1.01)	0.348	0.62 (0.40, 0.97)	0.256
< median	0.70 (0.46, 1.07)		0.78 (0.51, 1.20)		0.75 (0.50, 1.13)

Abbreviations: HR, hazard ratio; CI, confidence interval.

^aAdjusted for age (years), sex (male, female), race (non-Hispanic white, non-Hispanic black, Hispanic, others), educational level (college below, college graduate, postgraduate), smoking status [current (>60 pack-years, 30–60 pack-years, <30 pack-years), former (>60 pack-years, 30–60 pack-years, <30 pack-years), never], alcohol consumption (g/day), physical activity level (min/week), body mass index (kg/m²), aspirin use (yes, no), single or multivitamin supplement use (yes, no), history of diabetes (yes, no), family history of pancreatic cancer (yes, no), and energy intake from diet (kcal/day). In subgroup analyses stratified by sex, aspirin use, and smoking status, hazard ratios were not adjusted for the stratification factor.

Table 4.

Sensitivity analyses on the association of low-carbohydrate-diet scores with the risk of pancreatic cancer^a

	Overall low- carbohydrate-diet score		Animal low-carbohydrate-diet score		Vegetable low-carbohydrate-diet score
Categories	HRquartile 4 versus 1 (95% CI)	Ptrend	HRquartile 4 versus 1 (95% CI)	Ptrend	HRquartile 4 versus 1 (95% CI)	Ptrend
Excluded subjects with a history of diabetesb	0.61 (0.44, 0.85)	0.002	0.77 (0.55, 1.07)	0.056	0.73 (0.52, 1.00)	0.055
Excluded cases observed within the first 2 years of follow-up	0.70 (0.50, 0.99)	0.019	0.73 (0.54, 0.99)	0.021	0.71 (0.53, 0.98)	0.018
Excluded cases observed within the first 4 years of follow-up	0.76 (0.54, 1.07)	0.063	0.75 (0.54, 1.06)	0.091	0.74 (0.52, 1.04)	0.068
Repeated analysis in participants with complete data	0.68 (0.48, 0.97)	0.014	0.74 (0.52, 1.06)	0.093	0.72 (0.51, 1.03)	0.058
Further adjusted for consumption of fruits, coffee, tea and eggsc	0.56 (0.40, 0.78)	<0.001	0.71 (0.51, 0.99)	0.025	0.69 (0.51, 0.93)	0.013
Further adjusted for intakes of dietary fiber, folate, calcium, magnesium, iron and potassiumc	0.56 (0.40, 0.77)	<0.001	0.68 (0.49, 0.94)	0.011	0.65 (0.48, 0.89)	0.006

Abbreviations: HR, hazard ratio; CI, confidence interval.

^aHRs were adjusted for the following variables unless otherwise specified: age (years), sex (male, female), race (non-Hispanic white, non-Hispanic black, Hispanic, others), educational level (college below, college graduate, postgraduate), smoking status [current (>60 pack-years, 30–60 pack-years, <30 pack-years), former (>60 pack-years, 30–60 pack-years, <30 pack-years), never], alcohol consumption (g/day), physical activity level (min/week), body mass index (kg/m²), aspirin use (yes, no), single or multivitamin supplement use (yes, no), history of diabetes (yes, no), family history of pancreatic cancer (yes, no) and energy intake from diet (kcal/day).

^bHRs were not adjusted for history of diabetes.

^cAll covariates were treated as the continuous variable in multivariable Cox regression.

Discussion

Based on prospective data from a large randomized controlled trial, our study found that greater adherence to low-carbohydrate diets, regardless of the type of protein and fat, was associated with a lower risk of pancreatic cancer. Subgroup analysis further found that this inverse association was more evident in participants aged ≥65 years. In addition, dose–response analyses provided a thorough description for the risk of pancreatic cancer across the entire range of low-carbohydrate-diet scores and revealed non-linear dose–dependent effects of adhering to low-carbohydrate diets on the risk of pancreatic cancer, which seems to suggest that adherence to a vegetable-based low-carbohydrate diet may be more beneficial in the prevention of pancreatic cancer than adherence to an animal-based low-carbohydrate diet.

A previous study in Sweden found that there was no significant association of low-carbohydrate diets with the risk of pancreatic cancer (HR for the highest versus the lowest scores: 0.77; 95% CI: 0.39, 1.50; P_trend = 0.584) (15), which is inconsistent with our study. The median age of participants at baseline in the previous study was around 40 years (15), whereas the median age of our participants was 65 years. Our subgroup analyses observed that the inverse association of low-carbohydrate diets with the risk of pancreatic cancer was non-significant in participants aged <65 years. Hence, age difference between two populations may contribute to the raised inconsistency. An alternative explanation is the difference in the calculation of low-carbohydrate-diet score. Of note, the previous study only considered carbohydrate and protein intakes (15). In fact, low-fat dietary pattern has been associated with a reduced risk of pancreatic cancer in overweight or obese women (25). Therefore, using intakes of carbohydrate, protein and fat at the same time to construct the low-carbohydrate-diet score may better represent the true conditions in the real world. In addition, the non-significant association observed in the previous study could result from its limited power, as it documented only 70 pancreatic cancer cases (15).

Interestingly, we observed that the favorable association of low-carbohydrate diets with the risk of pancreatic cancer was statistically significant in individuals aged ≥65 years but not in those aged <65 years. Similar to our observation, several studies also found that age was a potential effect modifier for the association of low-carbohydrate diets and health outcomes (26–28). The exact mechanisms behind these observations are unclear. A straightforward explanation is that the old people could have higher adherence to a given dietary pattern. Indeed, a large cross-sectional study observed that the degree of adhering to Mediterranean diet increased with increases in age (29). A biological explanation is that these observations are due to a possible interaction between aging and low-carbohydrate diets in the biological pathway. Indeed, in an animal study, the adverse influence of low-carbohydrate diets on the metabolism and cardiovascular system was found in adult mice but not in young mice (30). However, it should be reminded that we cannot exclude the possibility that age difference in the association of low-carbohydrate diets with the risk of pancreatic cancer is a chance finding, although this phenomenon could be explained by the above points. Therefore, the results on effect modification by age should be treated with caution, and need to be verified by future studies.

Although animal-based and vegetable-based low-carbohydrate diets may include similar macronutrient contents, the type of macronutrient (e.g. saturated fat versus polyunsaturated fat) can lead to major differences in dietary composition that may influence the association between low-carbohydrate diets and health outcomes. This hypothesis has been confirmed in numerous studies in this area (12,14,26,31). For example, a prospective study of 136 967 individuals found that plant-based low-carbohydrate diets were associated with a lower risk of hepatocellular carcinoma, whereas there was no significant association with animal-based low-carbohydrate diets (14). Hence, in this study, we examined the association by the source of macronutrient. Unexpectedly, we found that both animal-based and vegetable-based low-carbohydrate diets were associated with a reduced risk of pancreatic cancer, indicating that the source of macronutrient plays little role in this association. In addition, the type of carbohydrate should also be considered in the evaluation of the association of low-carbohydrate diets with health outcomes, as different types of carbohydrate may have different associations with the risk of disease. For example, a meta-analysis found that whole grain and refined grain consumption were associated with a reduced risk and an increased risk of gastric cancer, respectively (32). However, establishing a practical classification system for the type of carbohydrate seems to be much challenging. Hence, all studies on low-carbohydrate diets and health outcomes, including ours, do not consider the potential impacts of the type of carbohydrate. More studies are needed to solve this key question.

Biologically, the observed inverse association of low-carbohydrate diets with the risk of pancreatic cancer may be explained by the following mechanisms. Unlike normal cells, cancer cells are severely dependent on glucose as the energy source to meet their biomass generating demands during the processes of proliferation and invasion. Moreover, carbohydrate consumption can promote the secretion of insulin and the production of insulin-like growth factor-1 (33,34), which can stimulate the proliferation of cancer cells though phosphoinositide 3-kinase/Akt/ mammalian target of rapamycin signaling pathway (35). Hence, carbohydrate-restricted diets may inhibit the malignant behaviors of cancer cells through reducing their energy supply and growth signaling. Additionally, low-carbohydrate diets are rich in fat, which is metabolized in the human body to produce ketone bodies. Generally, cancer cells have an inability of converting ketone bodies to ATP because of the lack of the relevant mitochondrial enzymes (35), thus a high-fat, ketogenic diet is assumed to have anti-cancer effects, which has been observed in human and animal studies (36–39).

Our study has several limitations. First, in this study, daily food consumption used for the calculation of low-carbohydrate-diet score was assessed once at baseline. The assessment of food consumption at one time point may result in non-differential bias, given that dietary habits can change over time. Nonetheless, it has been suggested that the approach using baseline diet only generally yields a weaker association than that using the cumulative averages (40). Second, we have fully adjusted the potential confounders, but we cannot exclude a possibility that our results have been biased by unrecognized or unmeasured confounders, considering the observational nature of our study. Moreover, in our study, covariate data used in the multivariable Cox regression models were not validated previously and cumulatively updated during follow-up, which might further make our results be subject to residual confounding. Third, misclassification bias might occur when participants self-reported their consumption of macronutrients, However, this bias is non-differential, as it was not expected to be associated with future pancreatic cancer incidence, and thus potentially biases risk estimates toward the null. Finally, our results were originated from a US population aged 55–74 years. Given different genetic backgrounds and dietary habits in different populations, our results may not be extended to other populations or other age groups.

In conclusion, low-carbohydrate diets, regardless of the type of protein and fat, are associated with a lower risk of pancreatic cancer in the US population. This inverse association may be more pronounced in individuals aged ≥65 years. These findings suggest that adherence to low-carbohydrate diets is helpful for pancreatic cancer prevention. Future studies should validate our findings in other populations.

Data availability

Data described in the manuscript are from the PLCO Cancer Screening Trial. Data can be made available upon the application and approval.

Supplementary material

Supplementary data are available at Carcinogenesis online.

Supplementary Figure 1. The study flow chart of identifying eligible subjects.

Supplementary Figure 2. The timeline and follow-up scheme of our study.

Supplementary Table 1. Criteria for determining the low-carbohydrate-diet score.

Supplementary Table 2. Distribution of variables with missing data before and after imputation.

Funding

This study did not receive any specific funding.

Acknowledgements

The authors sincerely thank the NIH PLCO study group and the National Cancer Institute (NCI) for access to NCI’s data collected by the PLCO Cancer Screening Trial. The authors also sincerely thank Prof. Yong Zhao (Department of Nutrition and Food Hygiene, School of Public Health and Management, Chongqing Medical University) for his suggestions and help in statistical analyses. This study has been approved by the NCI (approval number: PLCO-654); the corresponding study protocol is available online (https://cdas.cancer.gov/approved-projects/2690/). Of note, the statements contained herein are solely those of the authors and do not represent or imply concurrence or endorsement by NCI. G.-C.Z. conceived the study idea and was responsible for the response to reviewers’ comments. Q.-J.L. drafted the study protocol and the initial manuscript. Y.-B.W., F.-B.H., K.W. and J.-J.W. made critical comments and revisions for the initial manuscript. P.-F.Y., K.W. and J.-J.H. were responsible for statistical analyses. G.-C.Z. and P.-F.Y. interpreted the results of statistical analyses together. All authors approved the final version of the article, including the authorship list. All authors made critical comment and revision for the initial manuscript. All authors approved the final version of the article, including the authorship list. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Conflict of Interest Statement: None declared.