Data were based on the CLHLS study. The CLHLS is a nationwide community-based prospective longitudinal study, aiming to investigate health status and factors influencing longevity in the Chinese elderly. Its baseline survey began in 1998, and seven follow-up survey waves were conducted in 2000, 2002, 2005, 2008–2009, 2011–2012, 2014, and 2017–2018, respectively (Yi, 2008; Zeng, 2012). To obtain representative information of Chinese populations at an advanced age, the CLHLS study oversampled the oldest-old and was enriched for centenarians, such as 19.5 thousands centenarians, 26.8 thousands nonagenarians, and 29.7 thousands octogenarians (Zeng, 2013; Zeng et al., 2017). The validity of the CLHLS has been widely reported elsewhere (Gu et al., 2009, 2017).

In addition, a biomedical in-depth CLHLS study was performed in the CLHLS of waves 2008–2009, 2011–2012, and 2014 in sequence. During the in-depth study, physical examinations and almost 30 biomarkers in blood and urine samples were collected by the trained medical staff of the Chinese Center for Disease Control and Prevention local network. More details of the in-depth CLHLS study have been published elsewhere (Zeng et al., 2017; Lv et al., 2019; Liu et al., 2020). The CLHLS study was approved by the Institutional Review Board, Duke University (Pro00062871), and the Biomedical Ethics Committee, Peking University (IRB00001052–13074). All participants gave written informed consent.

In this study, we used three waves of the CLHLS data in 2011–2012, 2014, and 2017–2018. The inclusion criteria were as follows: participants aged 60 years or older; and participants with complete baseline information of the GNRI and Mini-Mental State Examination (MMSE) score. The exclusion criteria were as follows: participants with missing data in covariates; and participants with a history of intracranial tumor, craniocerebral injury, acute infection, mental illness, and cerebrovascular diseases, which were self-reported. Eventually, 1,632 eligible participants were included. The detailed information is presented in Figure 1. If the baseline survey was conducted in the wave 2011–2012, participants were followed up consecutively in the 2014 and 2017–2018 waves. Otherwise, if the baseline survey was only in the 2014 wave, follow-up was conducted in the wave 2017–2018. Therefore, there is no overlap in our final analyzed population.

The baseline data were collected using a standardized questionnaire in a home-based interview by trained investigators in 2011–2012 and 2014 waves. The information included (1) sociodemographic characteristics such as age, sex (male/female), ethnicity (Han/other), residence (urban/rural), marital status (married/other), education level (no schooling/some schooling), and living alone (yes/no), (2) lifestyle-related variables such as current smoking status (yes/no), drinking status (yes/no), and regular exercising (yes/no), and (3) health status involving a self-reported history of heart disease (yes/no), hypertension (yes/no), diabetes (yes/no), and arthritis (yes/no). Participants were divided into three age groups: 60–79 years, 80–89 years, and ≥90 years.

The participants were asked to report their food intake frequency of last month of vegetables, meat, eggs, and fish. The intake frequencies of meat, fish, and eggs were recorded as “almost every day,” “occasionally,” or “rarely or never.” While for vegetables, the intake frequency was “almost every day,” “almost every day except in winter,” “occasionally,” and “rarely or never” (Shi et al., 2015). In the analysis, we classified the frequency of “almost every day except in winter” as “almost every day.” The ability of basic activities of daily living (BADL) was measured by a question: Do you need assistance in bathing/dressing/toileting/transferring/eating/continence? According to a previous study using the data of the CLHLS, a score of 0 was given if no help was needed, and a score of 1 was given if some or complete help was needed. The BADL score ranged from 0 to 6 (Zeng et al., 2017). Based on the BADL score, we divided participants into having BADL disability (BADL score > 0) and not having BADL disability (BADL score = 0).

After the interview, the anthropometric measurements were carried out to obtain height (in meters) and weight (in kilograms). All measurements were carried out three times, and the average values were recorded. Body mass index (BMI) was calculated as weight divided by the square of height.

Serum albumin was measured by trained personnel using an automatic biochemistry analyzer (Hitachi 7180, Japan; Roche Diagnostic, Mannheim, Germany) at the central laboratory at the Capital Medical University in Beijing.

In this study, the GNRI was used to evaluate nutritional status. The GNRI is calculated by albumin and weight loss as follows: GNRI = [1.489 × serum albumin (g/L)] + [41.7 × (actual weight/ideal weight)] (Bouillanne et al., 2005). Weight loss is reflected by ideal body weight and actual weight. Different Lorentz formulas were used to calculate the ideal weight according to sex. For males, ideal weight was calculated by 0.75 × height (cm)–62.5, and ideal weight was obtained using 0.60 × height (cm)–40 for females (Matsuzawa et al., 1990). If the actual weight is greater than the ideal weight, the actual weight/ideal weight is set to 1 (Matsuzawa et al., 1990; Bouillanne et al., 2005). Participants were divided into four groups based on the GNRI as follows: normal nutrition group, a GNRI > 98; mild malnutrition group, a GNRI ≥ 92 but ≤ 98; moderate malnutrition group, a GNRI ≥ 82 but < 92; and severe malnutrition group, a GNRI < 82 (Bouillanne et al., 2005). In this study, since serum albumin failed to be collected in the wave 2008–2009, the GNRI was calculated only using the data of 2011–2012 and 2014 waves in the CLHLS. Moderate and severe malnutrition groups were combined into a moderate-to-severe group. The validity and reliability of the GNRI have been reported in previous studies (Abd Aziz et al., 2019; Nishi et al., 2019; Zhao Q. et al., 2020).

In each wave of the CLHLS, cognitive function was assessed by the Chinese version of MMSE, which was adapted from the international version that was proposed by Folstein et al. (1975) and validated in previous publications (An and Liu, 2016; Zeng et al., 2017; Zhang et al., 2019). In view of cultural and socioeconomic factors in the Chinese elderly, the Chinese version of the MMSE consists of 13 questions, covering 5 domains of cognitive function, namely, orientation, registration, attention and calculation, recall, and language. With total scores ranging from 0 to 30, a lower score of MMSE indicated a poorer cognitive function. More details about the Chinese version of MMSE are shown in Supplementary Table 1.

Continuous variables with normal distribution were presented as means ± SD. Non-normal variables were reported as median (P₂₅, P₇₅), whereas categorical variables were expressed as frequencies (percentages). The distributions of baseline characteristics were compared by one-way ANOVA, Kruskal–Wallis non-parametric test, and χ² test across the normal, mild malnutrition, and moderate-to-severe malnutrition groups. For the primary analysis, a linear mixed-effects model was employed to estimate the effects of malnutrition on cognitive function, adjusting for age, sex, ethnicity, residence, marital status, education, living alone, regular exercising, smoking, drinking, history of heart diseases, diabetes, hypertension, and arthritis, BADL disability, and the frequencies of meat, fish, eggs, and vegetables. In the linear mixed-effects model, the MMSE scores from waves 2011–2012, 2014, and 2017–2018 were response variables, and malnutrition assessed by the GNRI was an independent variable. Furthermore, the linear mixed-effects model was stratified by sex, age, and education level. Four sensitivity analyses were conducted. First, the GNRI as a continuous variable was considered as an independent variable in the linear mixed-effects model. Second, serum albumin and BMI were used to construct linear mixed-effects models, respectively. Third, the mediation analysis was used to examine whether the association between malnutrition and cognitive impairment was mediated by hypertension and diabetes. Finally, we imputed the missing covariates using the R software package “mice” to evaluate the influence of missing covariates on the results. Two-tailed P < 0.05 was considered to be statistically significant, and all analyses were conducted using SAS 9.4 (SAS Institute Inc., Cary, NC, United States).

Among 1,632 subjects, the average age was 85.57 ± 12.35 years old, and 54.6% of subjects were females. The proportion of participants in the normal nutrition, mild malnutrition, and moderate-to-severe malnutrition groups was 65.0, 19.4, and 15.6%, respectively. As shown in Table 1, the risk of malnutrition was higher in an older female with a lower MMSE score, BMI, and serum albumin concentration (P < 0.001). Also, subjects who have been married (P < 0.001), had regular exercises (P < 0.001), and ever attended school (P < 0.001) were less likely to be malnourished. Additionally, in contrast with normal nutritional group, people under malnutrition status tended to have BADL disability (P < 0.001), rare or never eat meat (P = 0.001), fish (P = 0.007), eggs (P = 0.015), and vegetables (P < 0.001). However, ethnicity, residence, living alone, and history of heart diseases, diabetes, and arthritis were not significantly different across the three groups.

Table 2 shows the association between malnutrition and cognitive function in the total population. Whether adjusting for covariates or not, people in mild and moderate-to-severe malnutrition groups have a poorer cognitive function (adjusted P = 0.007 and P < 0.001, respectively). Compared with the normal nutrition group, those with mild and moderate-to-severe malnutrition have a lower MMSE score of 0.95 and 1.39 points (adjusted β = –0.95, 95% CI = –1.60 to –0.25; adjusted β = –1.39, 95% CI = –2.21 to –0.57, respectively). Moreover, the MMSE score has shown a significant linear downward trend with increasing malnutrition degrees (adjusted P for trend < 0.001).

Table 3 shows the relationship between malnutrition and cognitive function stratified by sex. In males, after adjusting for covariates, a significant association was not found in both mild and moderate-to-severe malnutrition groups (adjusted P = 0.196 and P = 0.514, respectively). While for female subjects, there was a negative association of mild and moderate-to-severe malnutrition with cognitive function regardless of adjusting covariates or not (adjusted P = 0.024, β = –1.12, and 95% CI = –2.08 to –0.15; and P = 0.001, β = –1.94, and 95% CI = –3.09 to –0.79, respectively). A significant trend of a cognitive decline with increasing malnutrition degrees was observed in females (adjusted P for trend < 0.001) but not in males (adjusted P for trend = 0.289).

The association between malnutrition and cognitive function stratified by age is presented in Table 4. After adjusting for covariates, there was no association of mild and moderate-to-severe malnutrition with cognitive function among the elderly aged 60–79 years (adjusted P = 0.057 and P = 0.821, respectively) and 80–89 years (adjusted P = 0.695 and P = 0.490, respectively). While among the elderly above 90 years old, a significant association was observed between mild and moderate-to-severe malnutrition and cognitive function (adjusted P = 0.003, β = –2.04, and 95% CI = –3.37 to –0.71; and P < 0.001, β = –2.54, and 95% CI = –3.93 to -1.15, respectively). Moreover, a linear trend in the association of malnutrition risk with cognitive performance was also found in individuals aged 90 years and above (adjusted P for trend < 0.001).

After adjusting for covariates, a significant association and trend between malnutrition and cognitive impairment were found for the illiterate Chinese elderly (adjusted P = 0.004 and 0.007 for mild and moderate-to-severe malnutrition, respectively; adjusted P for trend = 0.002), while the association disappeared among the elderly who ever attended school (Table 5).

As shown in Supplementary Table 2, there is a significant positive relationship of the continuous GNRI and serum albumin with cognitive function, regardless of adjusting or not (crude and adjusted P < 0.001). Namely, subjects with a higher GNRI or serum albumin have a better cognitive function, while BMI was not associated with cognitive decline (adjusted P = 0.499). In the mediation analysis, as shown in Supplementary Table 3, hypertension, rather than diabetes, had a significant partially mediating effect on the relationship between malnutrition and cognitive impairment. The mediating effect was –0.095 (–0.164 to –0.025), and the proportion mediated of hypertension was 9.65%. Moreover, as shown in Supplementary Table 4, a significant association remained between mild (adjusted P = 0.024) and moderate-to-severe (adjusted P = 0.004) malnutrition and cognitive function in the complete dataset.

Our study has some strengths. First, because the CLHLS is a representative study in the Chinese elderly, the conclusion of our current study was more credible and reproducible. Second, compared with the previous study, we used the CLHLS, a longitudinal cohort study, to examine the association between malnutrition and cognitive function, which makes it easier to interpret the causal link between malnutrition and cognitive function. Finally, to our knowledge, the GNRI was rarely used to evaluate the relationship between malnutrition and cognitive function, particularly in the Chinese elderly. Compared to another single indictor, the GNRI has been considered as a reliable comprehensive tool for assessing malnutrition in the elderly. Thus, our study could provide solid evidence on the association of the GNRI with cognitive impairment, which is beneficial for delaying or preventing the occurrence and progression of cognitive impairment. However, this study had some limitations. First, since the CLHLS was designed for the oldest-old Chinese, generalizing the results to other populations should be cautioned. Second, although we collected confounding factors such as smoking, drinking, food intake frequency, hypertension, and history of diabetes, it may lead to a potential recall bias as some of them were self-reported. Third, since biomarkers were assayed for once in the random sample, the changes over time of biomarkers failed to be considered in this study. Finally, this study was enriched for octogenarians and nonagenarians, who might be healthier and less likely to suffer from malnutrition and cognitive impairment. Therefore, this study might underestimate the association of malnutrition with cognitive function.