Low-Carbohydrate Diets and Mortality in Older Chinese: A 15-Year Follow-Up of Guangzhou Biobank Cohort Study

Ce Sun Sun Yat-sen University Wei Sen Zhang Guangzhou Twelfth People’s Hospital Chao Qiang Jiang Guangzhou Twelfth People’s Hospital Ya Li Jin Guangzhou Twelfth People’s Hospital Xue Qing Deng Sun Yat-sen University Jean Woo The Chinese University of Hong Kong Kar Keung Cheng University of Birmingham Tai Hing Lam Guangzhou Twelfth People’s Hospital G Neil Thomas University of Birmingham Lin Xu (  xulin27@mail.sysu.edu.cn ) Sun Yat-sen University

We therefore conducted a prospective cohort study, using data from Guangzhou Biobank Cohort Study (GBCS), to investigate the associations of types of LCD with the risk of all-cause, cancer and CVD mortality in an older Chinese sample. Furthermore, we also examined whether the associations of types of LCD with the risk of all-cause, cancer and CVD mortality varied by diabetes status.

Study design and sample
The GBCS is a population-based cohort study in South China [13]. Brie y, GBCS is a 3-way collaboration among the Guangzhou Twelfth People's Hospital and the Universities of Hong Kong and Birmingham. All participants were recruited from a community social and welfare association, the Guangzhou Health and Happiness Association for the Respectable Elders (GHHARE) from 2003 to 2008. GHHARE is a large uno cial organization with ten branches throughout all districts of Guangzhou. Membership of this association is open to Guangzhou residents aged 50 years or older for a nominal, monthly fee of four CNY (≈50 US cents). Baseline information was collected using computerassisted face-to-face interview by trained nurses. Information of anthropometrics, blood pressure, fasting plasma glucose, lipids and in ammatory markers was collected following standard protocols. Reliability of the questionnaire was tested 6 months into recruitment by recalling 200 randomly selected participants for re-interview, and the results were satisfactory [13]. Ethics approval was granted by the Guangzhou Medical Ethics Committee of the Chinese Medical Association, Guangzhou, China. As the Food Frequency Questionnaire (FFQ) was shortened in phase 3 of baseline (2006)(2007)(2008), participants from phase 3 were not included in the current analysis.
Assessment of LCD score Information of diet was collected using a FFQ validated by Woo et al [14]. The LCD diet score was calculated as per the method described in a recent study by Shan et al [8]. Brie y, percentages of energy from carbohydrate, fat and protein for each participant were each calculated and used to rank the participants into 11 strata. For carbohydrate, participants in the lowest group received 10 points and those in the highest group received 0 points. The order of the strata for fat and protein was reversed. The scores of the three macronutrients were summed to create an overall LCD score, which ranged from 0 to 30. Two additional LCD scores were also created: (1) vegetable-based LCD scores were calculated according to the percentage of energy from high-quality carbohydrate, plant protein and unsaturated fat; (2) meat-based LCD scores were calculated according to the percentage of energy from low-quality carbohydrate, animal protein and saturated fat (Supplementary Table 1). Based on the Healthy Eating Index (HEI) 2015, high-quality carbohydrate was de ned as carbohydrate from whole grains, whole fruits, legumes and non-starchy vegetables, and low-quality carbohydrate as carbohydrate from re ned grains, added sugar, fruit juice, potato and other starchy vegetables [3]. As we found a signi cant interaction between LCDs and diabetes on all-cause mortality (P for interaction<0.001), we also conducted pre-speci ed analyses by diabetes status (Supplementary Table 2). Diabetes was de ned by having a history of diabetes or fasting glucose ≥7.0 mmol/L at baseline.

Ascertainment of mortality
Details of the methods were described elsewhere, and information on the causes of death up to April 19, 2021 was obtained through record linkage with the Death Registry of the Guangzhou Center for Disease Control and Prevention (GCDC) [15]. Brie y, causes of death were coded by trained nosologists in each hospital according to the International Classi cation of Diseases, Tenth Revision (ICD-10). If death certi cates were not issued by medical institutions, the causes of death were veri ed by GCDC as part of its quality assurance program by cross-checking past medical history and conducting a verbal autopsy. Moreover, we also conducted verbal autopsy meetings in the Guangzhou Twelfth People's Hospital to further clarify the deaths of unclear causes. In the present study, the primary outcome was mortality from all causes, and the secondary outcome was mortality from cancer and CVD.

Potential confounders and mediators
As sex, age, socioeconomic factors (education, family income) [16], lifestyle factors (smoking, drinking and physical activity), body mass index (BMI) [17], and history of cancer and CVD were associated with both dietary pattern and mortality, these factors were considered as potential confounders and adjusted in the regression model. The potential mediators between LCD score and all-cause mortality risk included systolic blood pressure (SBP), fasting plasmaglucose (FPG) total cholesterol (TC) and self-rated health at baseline. Procedures for measuring these were reported previously [13].

Statistical analysis
Chi-square test and one-way analysis of variance (ANOVA) were used to compare baseline categorical and continuous variables by quartiles of LCD scores, respectively. Person-years of follow-up were assessed from the date of baseline enrollment to death or end of the present study on April 19, 2021, whichever came rst. The LCD scores were categorized into quartiles. Multivariable Cox proportional hazards regression was used to calculate hazard ratio (HRs) and 95% con dence intervals (CI) of mortality associated with LCD score. Schoenfeld's residuals were used to test the proportional hazard assumption and no violations of the proportional hazard assumption were found. Model 1 was the crude model without any adjustment. In multivariable analyses, model 2 adjusted for sex and age, and model 3 additionally adjusted for education, family income, smoking, drinking, physical activity, BMI and history of cancer and CVD. Model 4 adjusted for determinants considered potential mediators, namely, SBP, FPG, TC and self-rated health. In addition, the non-linearity of the effect of LCD score on mortality risk was estimated by adding a quadratic term to the model with the quartiles of LCD scores as a continuous variable and the tness of the models with and without the quadratic term was compared using the likelihood ratio (LR) test [18]. A non-signi cant P-value was interpreted as an indication of a linear effect of LCD score on mortality risk. Furthermore, strati cation analysis was done for the associations between diet score and all-cause mortality according to several potential effect modi ers at baseline. As many statistical tests were performed in subgroup analysis, we used the Bonferroni correction to account for multiple testing and the signi cance level was set at P < 0.002 (0.05/8 [subgroups] × 3 [dietary scores]). To assess the extent to which baseline risk factors explained the associations of LCD score with mortality, the percentage of excess risk mediated (PERM) was calculated for the following four groups of explanatory variables: (1) SBP; (2) FPG; (3) TC; (4) self-rated health. For each risk-factor group, we calculated PERM as: PERM= [HR (E+C)-HR (E+C+M)]/ [HR (E+C)-1] *100, where E=exposure, C=covariates (sex, age, education, family income, smoking, drinking, BMI, physical activity, and history of cancer and CVD), M=explanatory variable being tested [19]. A crude model was rst developed, then adjusted for age and sex and subsequently to address other potential confounders (education, family income, oil intake, smoking, drinking, physical activity, BMI, and history of cancer and CVD), and nally to examine for proposed mediators (systolic blood pressure, fasting plasma-glucose, total cholesterol and self-rated health). To account for potential bias due to reverse causality, we conducted a sensitivity analysis excluding participants who died within the rst three years of follow-up. Statistical analysis was done using Stata (STATA Corp LP, version 15). Two-sided P-values < 0.05 were considered as statistically signi cant.

Participant characteristics
Of 20,490 participants, 128 with potentially unreliable dietary intake (<800 or >4200 kcal/d in men, and <600 or >3500 kcal/d in women), 57 who were followed up for less than 1 year, and 99 lost to follow-up with unknown vital status were excluded (Supplementary Figure 1) Table 1 shows that compared with a low LCD score (Q1), high LCD score (Q4) was associated with being women, having a younger age, higher education level and family income, lower physical activity, never smoking and a current alcohol consumer (all P<0.05). However, the potential mediators, lower FPG and SBP level were found in those with higher overall LCD score (P<0.001). Similar results were found in meat-based LCD score though BMI was lower but with increased history of cancer and CVD. In contrast, those with a higher vegetable-based LCD score had lower education level, with a greater proportion of smokers and higher SBP and lower TC levels. Vegetable-based LCD score showed no association with sex, age, family income, drinking, BMI, and history of CVD. Participants without diabetes showed similar patterns as all participants (Supplementary Table 3). In contrast, participants with diabetes showed no association of overall LCD score with sex, age, smoking, drinking and FPG level, no association of vegetable-based LCD score with smoking, history of cancer and SBP level, no association of meat-based LCD score with age, BMI, history of CVD and FPG level but a negative association between meat-based LCD score and drinking (Supplementary Table 3).
Mortality and LCD score Table 2 shows that in all participants, after adjusting for sex, age, education, family income, smoking, drinking, physical activity, BMI and history of cancer and CVD, overall LCD score showed no association with all-cause mortality. For vegetable-based LCD score, the adjusted HR (95% CI) of all-cause mortality for the 2 nd , 3 rd and 4 th quartile, versus the 1 st quartile (Q1), was 0.99 (0.91-1.07), 1.11 (1.02-1.21) and 1.16 (1.05-1.27) (P for trend<0.001 and P for non-linear=0.18), respectively. For meat-based LCD score, the adjusted HR (95% CI) of all-cause mortality for the Q2, Q3 and Q4, versus the Q1, was 0.89 (0.83-0.97), 0.90 (0.83-0.97) and 0.89 (0.81-0.97) (P for trend=0.007 and P for non-linear=0.06), respectively. Table 3 shows no association between overall LCD score and all-cause mortality in participants with or without diabetes. In those without diabetes, the results on vegetable-based LCD and meat-based LCD score were generally similar with those from the total population. Comparing with the Q1 group, participants in Q4 of vegetable-based LCD score showed a higher risk of all-cause mortality (HR 1.10, 95% CI 1.01-1.23), whereas those with the highest quartile of meat-based LCD score showed a lower risk of all-cause mortality (HR 0.87, 95% CI 0.79-0.97). In participants with diabetes, no associations of vegetable-based LCD score and meat-based LCD score quartiles with all-cause mortality were found, although there was a linear trend between vegetable-based LCD score and all-cause mortality (P for trend= 0.04).
Supplementary Table 4 shows no association of the overall LCD score with mortality of cancer, CVD and other causes. The vegetable-based LCD score was associated with higher risk of CVD mortality (Q1: reference, Q2: 1.18 (1.03-1.34), Q3: 1.36 (1.18-1.56) and Q4 1.39 (1.19-1.62), P for trend<0.001 and P for non-linear=0.15), whereas the higher meatbased LCD score quartiles were associated with lower risk of CVD mortality (Q1: reference, Q2: 0.84 (0.75-0.95), Q3: 0.82 (0.72-0.93) and Q4: 0.81 (0.70-0.93), P for trend=0.02 and P for non-linear=0.10). Supplementary Table 5 shows similar results in participants without diabetes. In participants with diabetes, no association of overall LCD score and meatbased LCD score with CVD mortality was found. However, we found a positive association between vegetable-based LCD score quartiles and CVD mortality (Q1: reference, Q2: 1.21 (0.89-1.63), Q3: 1.59 (1.18-2.15), Q4: 1.54 (1.11-2.14), P for trend=0.003 and P for non-linear=0.25). Figure 1 shows that the HR of all-cause mortality comparing Q4 to Q1 of vegetable-based LCD score was 1.16 (1.05-1.27) after adjustment for potential confounders. The HR decreased by 14% after adjusting for SBP, 27% after adjusting for FPG and 2% after adjusting for self-rated health, and increased by 3% after adjusting for TC. The overall attenuation after adjustment for mediators was 41%. Similar patterns were found for the association between vegetable-based LCD score and cause-speci c mortality. FPG appeared to be the strongest mediator in vegetable-based LCD diet. For meatbased LCD, the HR of all-cause mortality was 0.89 (0.81-0.97) after adjustment for potential confounders, which increased by 10% after adjusting for SBP, 24% after adjusting for FPG and 1% after adjusting for self-rated health. FPG appeared to be the strongest mediator in meat-based LCD diet. Figure 2 shows that in participants without diabetes, vegetable-based LCD score was associated with higher risk of mortality from all-cause mortality (HR comparing Q4 to Q1=1.10, 95% CI 1.01-1.23) and CVD (HR 1.26, 95% CI 1.06-1.51). After adjustment for mediators, the HRs of all-cause mortality became non-signi cant (HR 1.06, 95% CI 0.96-1.18). SBP appeared to be the strongest mediator (PERM=26% for all-cause mortality and 19% for CVD mortality). In participants with diabetes, no association of vegetable-based LCD score with all-cause mortality was found. However, we found that vegetable-based LCD score was associated with a higher risk of CVD mortality (adjusted HR 1.54, 95% CI, 1.11-2.13), for which TC appeared to be the strongest mediator (PERM=16%). Figure 3 shows that in participants without diabetes, meat-based LCD score was associated with lower risk of all-cause and CVD mortality, which was partly mediated by SBP (PERM=13% for all-cause and CVD mortality). No association of meat-based LCD score with cancer and other-cause mortality was found.

Subgroup and sensitivity analyses
Supplementary Figures 2 to 7 show similar associations in most subgroups. After Bonferroni corrections for multiple testing, the association between vegetable-based LCD score and all-cause mortality appeared to be stronger in obese than non-obese participants (Supplementary Figure 3, P for interaction<0.001). Higher vegetable-based LCD score was associated with a higher risk of all-cause mortality (HR for Q4 versus quartile 1=1.55, 95% CI 1.18-2.04). Similar results were observed in participants without diabetes (Supplementary Figure 6, all P for interaction<0.001). Similar results were found after excluding deaths within the rst three years of follow-up (Supplementary Table 6 and 7).

Discussion
Although after a long-term follow-up for nearly 15 years, no association of overall LCD score with risk of all-cause and cause-speci c mortality was found in our study. In prespeci ed subgroup analyses, we found that vegetable-based LCD score was positively, whereas meat-based LCD score was negatively associated with all-cause and CVD mortality.
Similar associations were observed for participants without diabetes. In participants with diabetes, a positive association of vegetable-based LCD score with risk of CVD mortality was found.

Comparison with previous studies
Most studies considered LCD based on animal-derived protein and fat sources as a risk factor of mortality, whereas LCD based on plant-derived protein and fat reduced mortality [5,20,21]. Furthermore, studies show that higher level of whole grain intake was associated with lower risk of all-cause mortality, whereas re ned grain intake was associated with higher risk of all-cause mortality [22,23]. This highlights that, a healthy LCD diet is not only dependent on the sources of macronutrients but also on the quality of them. A previous study using a new classi cation approach of LCD score found that participants with low low-quality re ned carbohydrate, high unsaturated fat and plant protein intake had lower all-cause and cancer mortality risk, whilst those with low high-quality unre ned carbohydrate and high saturated fat and animal protein intake had higher all-cause mortality risk [8]. Our results generally supported the intake of highquality carbohydrate, and further showed that participants with low low-quality re ned carbohydrate, high saturated fat and animal protein intake had lower all-cause and CVD mortality risk, and with low high-quality unre ned carbohydrate, high unsaturated fat and plant protein intake had higher mortality risk. The discrepancy might be due to the differential amount and sources of carbohydrate, fat and protein in the West and non-West settings.
The percentage of energy from carbohydrate, fat and protein in our study were similar with the China Health and Nutrition Survey (CHNS) [24]. Notably, the percentage of energy from carbohydrate (especially high-quality carbohydrate) in our study was higher than that reported in the US (total carbohydrate, 57.1% versus 50.5%; high-quality carbohydrate, 10.6% versus 8.6%, respectively), whereas the percentage of energy from animal protein and saturated fat intake was much lower than the US (animal protein, 7.4% versus 10.4%; saturated fat, 4.9% versus 11.9%, respectively) [3]. Moreover, compared with the US, total per capita consumption of meat in Asians was much lower (49.4kg/year versus 122.8kg/year), whereas the percentage of energy from sh/sea food consumption was higher (43.5% versus 26.0%) [25]. Some recent studies showed that sh/sea food consumption was associated with lower risk of all-cause and CVD mortality in Asians, but not in the US populations [26,27]. Meta-analyses show that total mortality is in participants who have high intakes of both red and processed meat than in those with low meat intakes in western high-income countries (28). However, meat is good source of energy and a range of essential nutrients, including protein and micronutrients such as iron, zinc, and vitamin B12 for low-income countries. A previous study showed that Indian vegetarians had a more favorable cardiovascular risk pro le than that of non-vegetarians [29]. Along with these ndings, our results support the bene cial effects of moderate consumption of animal protein. In addition, the nonsigni cant association between meat-based LCD and CVD mortality in diabetic patients could also be explained by the lower levels of sh consumption than those without diabetes [26]. Moreover, a recent meta-analysis also showed that substituting sh with red and processed meat was associated with increased risks of all-cause mortality in patients with type 2 diabetes [30]. Apart from differential amount of high-quality carbohydrate, saturated fat and animal protein intake, the discrepancies could also be at least partly explained by the low-quality carbohydrate, unsaturated fat and plant protein. Notably, comparing with the US, the percentage of energy from low-quality carbohydrate (46.4% versus 41.8%), due in part to the high white rice intake, and plant protein (8.5% versus 5.8%) intake was higher in our sample, whereas percentage of energy from unsaturated fat was lower (monounsaturated fatty acids, 8.6% versus 13.1%; polyunsaturated fatty acids, 6.3% versus 8.2%).
Regarding the results on vegetable-based LCD, our results were generally consistent with previous studies in Asia showing positive associations between plant-based diets consisting of a high intake of re ned carbohydrates and the risk of metabolic syndrome and CVD [31, 32]. In our study, participants with higher vegetable LCD score had higher levels of unsaturated fat consumption and higher risks of all-cause and CVD mortality, which could be partly explained by the cooking method. In traditional Chinese cuisine, plant oil is often used for stir-frying, pan-frying and deep-frying and is heated to a high temperature [33]. High heat has been shown to cause partial hydrogenation of unsaturated plant oils to produce trans unsaturated fats. Studies have consistently shown trans unsaturated fats consumption to be associated with higher risk of all-cause and CHD mortality [34]. Moreover, as CVD is a leading cause of mortality in people with type 2 diabetes mellitus [35]. the stronger positive association between vegetable-based LCD and CVD mortality in participants with diabetes in our study also warrants further attention.
Regarding the null association between the overall LCD and mortality, our results were consistent with some [8, 36] but not all [20,37] previous studies. For example, a recent study in Japan showed a U-shape association between overall LCD score and all-cause mortality [37]. The authors suggested that sources of food might have modi ed the association [37]. Another study in the US showed a positive association between overall LCD and all-cause mortality [20]. The differences in the results could be partly due to the substantial variation in carbohydrate consumption across different populations (i.e., about 60% of the overall energy was from carbohydrate in Asians vs. 50.5% in the US) [2,3].
Higher carbohydrate consumption could lead to greater glycemic burden, and a subsequently higher risk of insulin resistance and vascular complications [38-40], which warrants further research in populations with high carbohydrate intake.

Strengths and limitations
Strengths of our study included the large sample size, long duration of follow-up and comprehensive adjustment for potential confounders. There were some limitations in the present study. Firstly, changes in dietary patterns were not assessed during follow-up. However, our previous study found that the dietary patterns of our sample were relatively stable [41, 42]. Secondly, residual confounding could not be fully ruled out, although we adjusted for a wide range of potential confounding factors reported in previous literature. Thirdly, our results may not be directly applicable to younger or western populations. Fourthly, the null association in the subgroup of participants with diabetes could be due to the relatively small sample size. Although a recent meta-analysis showed that patients adhering to an LCD for six months may experience remission of diabetes without adverse consequences [43], further studies on the health effects related to long-term and types of LCD patterns in diabetic patients are warranted.
In conclusion, we found that vegetable-based LCD score was positively, whereas meat-based LCD score was negatively associated with all-cause mortality. Inconsistencies in the literature on the health effect of LCD may re ect the importance of the geographical context and age-related nutrient composition of the diet.    Overall LCD score a Median score (IQR) 6 (4, 8)