GLIM Achieves Best Diagnostic Performance in Non-Cancer Patients with Low BMI: A Hierarchical Bayesian Latent-Class Meta-Analysis

Abstract

Context

Global Leadership Initiative on Malnutrition (GLIM) and Patient-Generated Subjective Global Assessment (PG-SGA) are commonly used nutrition assessment tools, whose performance does not reach a consensus due to different and imperfect reference standards.

Objective

This study aimed to evaluate and compare the diagnostic accuracy of GLIM and PG-SGA, using a hierarchical Bayesian latent class model, in the absence of a gold standard.

Data Sources

A systematic search was undertaken in PubMed, Embase, and Web of Science from inception to October 2022. Diagnostic test studies comparing (1) the GLIM and/or (2) PG-SGA with “semi-gold” standard assessment tools for malnutrition were included.

Data Extraction

Two authors independently extracted data on sensitivity, specificity, and other key characteristics. The methodological quality of each included study was appraised according to the criteria in the Quality Assessment of Diagnostic Accuracy Studies-2.

Data Analysis

A total of 45 studies, comprising 20 876 individuals evaluated for GLIM and 11 575 for PG-SGA, were included. The pooled sensitivity was 0.833 (95% CI 0.744 to 0.896) for GLIM and 0.874 (0.797 to 0.925) for PG-SGA, while the pooled specificity was 0.837 (0.780 to 0.882) for GLIM and 0.778 (0.707 to 0.836) for PG-SGA. GLIM showed slightly better performance than PG-SGA, with a higher diagnostic odds ratio (25.791 vs 24.396). The diagnostic performance of GLIM was most effective in non-cancer patients with an average body mass index (BMI) of <24 kg/m², followed by non-cancer patients with an average age of ≥60 years. PG-SGA was most powerful in cancer patients with an average age of <60 years, followed by cancer patients with an average BMI of <24 kg/m².

Conclusion

Both GLIM and PG-SGA had moderately high diagnostic capabilities. GLIM was most effective in non-cancer patients with a low BMI, while PG-SGA was more applicable in cancer patients.

Systematic Review Registration

PROSPERO registration No. CRD42022380409.

INTRODUCTION

Malnutrition is a state resulting from a lack of intake or uptake of nutrition that leads to decreased fat-free mass and body cell mass, resulting in diminished physical and mental function and impaired clinical outcome from disease.¹ As an imperative and prevalent public health challenge, malnutrition has detrimental effects on human physical health and quality of life and is one of the world’s leading causes of illness and death. The situation became even more serious under the shadow of the COVID-19 pandemic; the prevalence of undernourishment climbed to around 9.9% in 2020, with 768 million people suffering worldwide.² According to a previous systematic review that covered 66 studies from Latin American countries, the incidence of malnutrition was estimated at 40% in hospitalized patients.³ Malnutrition, which leads to an increase in adverse reactions and treatment complications, seriously reduces patients’ tolerance of treatments and the efficacy of those treatments, and is a powerful predictor of poor prognosis, long-term hospitalization, readmission, and mortality.⁴ To address these issues, it is proposed that nutrition therapy be added to conventional management. In light of this, effective methods for diagnosing malnutrition are the premise of and critical to the development of personalized therapy.

Recently, the Global Leadership Initiative on Malnutrition (GLIM) proposed a 2-step process for assessing nutritional status, based on phenotypic and etiologic criteria,⁵ and the validation work for that method of assessment is in progress.⁶ Another assessment tool, Patient-Generated Subjective Global Assessment (PG-SGA), consisting of patients’ self-reports and medical staff assessments, has served as an acceptable method for malnutrition diagnosis in the last 20 years.⁷ At present, both the GLIM and PG-SGA are commonly used for nutritional assessment. To meet the needs for precision nutrition treatment, a systematic evaluation and comparison of the diagnostic accuracy of the GLIM and PG-SGA among various groups of patients is warranted.

The lack of a gold standard reference is an acknowledged problem in studies of malnutrition diagnostics. Heterogeneity in the reference standard between studies is one of the sources of variability and confounding in the validation studies. For example, when Subjective Global Assessment (SGA) was used as the reference, the sensitivity of PG-SGA was 91.1%,⁸ but when the malnutrition universal screening tool (MUST) was used, the PG-SGA sensitivity was only 63.1%.⁹ There would be substantial bias in the pooled estimates in diagnostic test accuracy meta-analyses if the imperfect nature of the reference standard was ignored. To solve this problem, Rutter and Gadtsonis proposed the hierarchical summary receiver operating characteristic (HSROC) model to take into account the difference in the kind, number, and cut-off values of measurement indices across studies.¹⁰ Dendukuri et al developed a hierarchical Bayesian latent-class approach based on the HSROC model to control the imperfect gold standard bias, with the hope of producing unbiased pooled outcomes in the absence of a gold standard reference.¹¹ Without a gold standard, the current validated nutritional assessment tools were regarded as “semi-gold” standards and the preferred criteria for validation, such as GLIM, SGA, PG-SGA, the European Society for Clinical Nutrition and Metabolism (ESPEN 2015), and the American Society for Parenteral and Enteral Nutrition 2012 (ASPEN), all these widely used tools could be served as reference standards. Hence, a meta-analysis, using the hierarchical Bayesian latent-class model, of all the available eligible studies was conducted to evaluate and compare the performance of the GLIM and the PG-SGA in assessing malnutrition among various populations.

METHODS

A diagnostic test accuracy meta-analysis was performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analysis extension for Diagnostic Test Accuracy Studies (PRISMA-DTA) guidelines (Supplementary PRISMA DTA Checklist),¹² and the protocol was registered with PROSPERO (Registration No. CRD42022380409). The PICOS (Participants, Intervention, Comparison, Outcomes, Study design) criteria used to structure the research question are shown in Table 1.

Data sources and search strategy

A comprehensive search strategy combining the keywords and Medical Subject Heading terms was undertaken to identify relevant studies in PubMed, Embase, and Web of Science from inception to October 2022. “Malnutrition,” “Nutritional Deficiency,” “Global Leadership Initiative on Malnutrition,” “Patient generated subjective global assessment” and their synonyms were used as key terms to define the scope of the meta-analysis and construct the search formula. To make our analysis more comprehensive, reference lists of previous reviews and meta-analyses were retrieved and reviewed to reveal additional relevant primary diagnostic studies. Note, the searches for GLIM and PG-SGA were separate. The full search strategy can be found in Table S1.

Inclusion and exclusion criteria

Studies were included in the meta-analysis based on the following criteria: (1) all of the sample populations were adults, regardless of disease status, disease type, or treatment status; (2) the validation study used all of the GLIM or PG-SGA criteria for malnutrition diagnostic tests; (3) concurrent criterion validity of GLIM or PG-SGA compared with other nutritional assessment tools (GLIM, SGA, PG-SGA, the ESPEN criteria); (4) the results were reported in English or Chinese, with sufficient information to reconstruct a 2 × 2 contingency table.

Publications were eliminated if one of the following existed: (1) the research was not related to malnutrition; (2) the publication was in the form of a review, conference abstract, comment, or letter; (3) the study was not reported in English or Chinese; (4) the publication provided insufficient data to construct a contingency table; (5) a repeated study with the same population; (6) use of an isolated nutritional indicator or nutritional screening tool as the reference standard. The detailed reasons for exclusion of publications are listed in the flow diagrams (Figure 1). Regarding studies comparing the GLIM and PG-SGA directly, they were regarded as GLIM diagnostic tests if PG-SGA was used as a reference standard, and as PG-SGA diagnostic tests if the GLIM was used as a reference standard. Two authors separately filtered eligible studies for inclusion. Disagreements were resolved according to the opinion of a third author.

Data extraction and synthesis

The numbers required for the construction of contingency tables, including true positive (TP), false negative (FN), false positive (FP), true negative (TN), sensitivity, and specificity, were extracted from the eligible articles for analysis of the concurrent validity of the two tools. In addition, general information on the study characteristics was collected (title, first author, year of publication, country of study, sample size), as well as population characteristics (age, gender distribution, body mass index [BMI], disease type), reference standard, and cut-off value.

Data from each of the included studies were extracted and documented independently by 2 authors. Discrepancies in opinion were resolved by discussion with 2 additional reviewers. If a study was published multiple times, the same data was included only once to avoid overlapping, and the one with the larger sample size or the most recently published was chosen. If more than 1 reference standard or cut-off were provided in the same article, the optimal reference standard and cut-off points with the best balance of sensitivity and specificity were selected.

Quality assessment

The methodological quality of each of the included studies was appraised according to the criteria in the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool,¹³ which evaluates the risk of bias and the applicability of diagnostic accuracy studies focusing on 4 domains: patient selection, index test, reference standard, and flow and timing. The risk of bias and concern of applicability for each domain were categorized as “low,” “high,” or “unclear.” The appraisal of study quality was administered independently by 2 reviewers. Any differences were resolved by reaching a consensus through negotiation and discussion.

Statistical analysis

With the available information extracted from the eligible studies, the results of the 2 diagnostic tests were converted into dichotomous variables (at risk of malnutrition or malnutrition vs no nutritional risk). The TP, FP, FN, and TN values in contingency tables were directly obtained or calculated indirectly from the sensitivity and specificity. Then, the hierarchical Bayesian latent-class model was applied to jointly estimate the sensitivity, specificity, diagnostic odds ratio (DOR), positive and negative likelihood ratio (LR+, LR−), and corresponding 95% confidence intervals (95% CI) in the absence of a gold standard, with the Stata package “metandi.”¹⁴ HSROC curves were drawn to show the diagnostic effects under different positive definitions. The DOR represents the combined values of sensitivity and specificity, ranging from 0 to infinity, with a higher value signifying better diagnostic accuracy. LRs, including LR+ and LR−, were ideal indices for the comprehensive evaluation of the diagnostic test, as they combine the clinical significance of sensitivity and specificity. LR is a ratio of the probability that a test result is correct to the probability that the test result is incorrect. LRs have unique properties that make them particularly useful for clinician decision-makers. Higher LR+ corresponds to the clinical concept of “rule in disease,” and lower LR− means “rule out disease.” A good diagnostic test typically has a LR+ of greater than 5.0 and a LR− of less than 0.2.

The Bayes theorem was utilized to estimate the post-test probability of malnutrition based on pre-test probability and LR. The pre-test probability was 40%, according to previous studies.^15–17 The calculation formula is as follows: Pre-test odds = pre-test probability/(1 − pre-test probability), post-test odds = pre-test odds × LR, post-test probability = post-test odds/(1 + post-test odds). These results are shown by the Fagan plot.

Heterogeneity was assessed using Cochran’s Q test and inconsistency index (I²), with P < .05 or I² > 50% indicating the existence of heterogeneity. Subgroup analyses were established to find out which population was more suitable for GLIM/PG-SGA and to explore potential sources of heterogeneity. Studies were stratified according to 5 dimensions, including the average age of the population (over or equal to 60 years old vs less than 60 years old), average BMI of the population (over or equal to 24 kg/m² vs less than 24 kg/m²), population type (cancer patients vs non-cancer patients), continent, and reference standard. It should be noted that some groups with fewer than 4 primary studies could not be converged using the HSROC model. Deeks’ funnel plot analysis was carried out to evaluate publication bias, with P < .1 indicating that publication bias might exist. STATA 17.0 with the “metandi” command was adopted to run the hierarchical Bayesian latent-class model. The R package “meta4diag” was used to help with visualization of the summarized sensitivity and specificity in R 4.2.1. Cochrane Collaboration Review Manager 5.4.1 statistical software was employed to make assessments of quality.

RESULTS

Study identification and characteristics

A total of 45 articles conducted across 15 countries and covering 2 types of nutritional assessment tools were available in this meta-analysis, with 32 studies assessing GLIM^18–50 (Figure 1A) and 26 studies assessing PG-SGA^{7,8,23,25,27,29,35,37,43–48,50–61} (Figure 1B). Among these, 13 articles involved both GLIM and PG-SGA, 15 were conducted in China, 8 in Australia, 4 in Japan, and 3 in Brazil. The remaining 15 articles were conducted in 11 other countries. The sample size ranged from 42 to 3777 patients, with a total of 20 876 individuals for GLIM and 11575 for PG-SGA included in our final analysis. The detailed characteristics of the eligible studies are summarized in Table 2.

Quality of the included studies

The methodological quality of each included study was evaluated using QUADAS-2. The results of the risk-of-bias assessments and applicability concerns of the individual studies assessing the GLIM and PG-SGA are exhibited in Figure S1. The overall methodological quality of the included studies was considered moderately high, as shown in Figure S2.

Result of meta-analysis

Summary estimates of sensitivity, specificity, LR+, and LR− generated from the HSROC model are presented in Table 3. The pooled sensitivity was 0.833 (95% CI 0.744 to 0.896) for the GLIM and 0.874 (95% CI 0.797 to 0.925) for PG-SGA, while the pooled specificity was 0.837 (95% CI 0.780 to 0.882) for GLIM and 0.778 (95% CI 0.707 to 0.836) for PG-SGA. Notably, the GLIM showed lower sensitivity than PG-SGA, but higher specificity.

Regarding the LR, the LR+ was 5.124 (95% CI 3.797 to 6.914) for the GLIM and 3.939 (95% CI 3.014 to 5.148) for PG-SGA. The LR− was 0.199 (95% CI 0.128 to 0.309) for the GLIM and 0.161 (95% CI 0.101 to 0.258) for PG-SGA. The overall DOR was 25.791 (95% CI 14.574 to 45.640) for the GLIM and 24.396 (95% CI 14.408 to 41.308) for PG-SGA.

The HSROC curve presents summary points, 95% confidence regions, and 95% prediction regions (Figure 2), which intuitively illustrates the performance of the 2 nutritional assessment tools in diagnosing malnutrition. The forest plots reveal the sensitivity and specificity of the GLIM (Figure 3A) and PG-SGA (Figure 3B) in each of the included studies, and considerable heterogeneity in terms of the sensitivity and specificity of the GLIM and PG-SGA was observed (all P < .01, I² > 50%).

Subgroup analysis

Subgroup analyses were performed to investigate the diagnostic efficacy of the GLIM and PG-SGA in various groups, and the results are shown in Table 4. According to various DOR values, GLIM was found to have better diagnostic performance in patients with an average age of ≥60 years (DOR: 38.454, 95% CI: 10.675 to 138.519) than those with an average age of <60 years (DOR: 30.761, 95% CI: 16.294 to 58.070); showed superior diagnostic capability in patients with an average BMI of <24 kg/m² (DOR: 56.944, 95% CI: 19.244 to 168.499) than those with an average BMI of ≥24 kg/m² (DOR: 14.358, 95% CI: 7.144 to 28.858); and had better diagnostic ability in non-cancer patients (DOR: 27.223, 95% CI: 10.872 to 68.165) than cancer patients (DOR: 22.618, 95% CI: 14.423 to 35.470). GLIM exhibited better discrimination ability in Asia (DOR: 42.473, 95% CI: 19.330 to 93.327) than in Europe (DOR: 6.793, 95% CI: 3.046 to 15.147). Comparing 3 reference standards, GLIM showed the best diagnostic accuracy while considering ESPEN 2015 as the reference standard (DOR: 119.495, 95% CI: 13.801 to 1034.658), compared with PG-SGA (DOR: 17.755, 95% CI: 11.252 to 28.017) and SGA (DOR: 21.482, 95% CI: 7.816 to 59.039).

PG-SGA had better diagnostic performance in patients with an average age of <60 years (DOR: 32.381, 95% CI: 13.575 to 77.237) than those with an average age of ≥60 years (DOR: 19.577, 95% CI: 9.708 to 39.480); had better diagnostic accuracy in patients with an average BMI of <24 kg/m² (DOR: 42.236, 95% CI: 17.317 to 103.015) than those with an average BMI of ≥24 kg/m² (DOR: 38.277, 95% CI: 13.727 to 106.734); and had clearly better discrimination ability in cancer patients (DOR: 29.628, 95% CI: 16.172 to 54.282) than in non-cancer patients (DOR: 16.771, 95% CI: 6.398 to 43.958). PG-SGA was observed to show the best performance in Oceania (DOR: 61.271, 95% CI: 14.101 to 266.229) compared with Asia and Europe. After comparing performance according to 5 reference standards, PG-SGA exhibited better diagnostic value while considering SGA as the reference standard (DOR: 103.455, 95% CI: 26.023 to 411.282) than GLIM (DOR: 16.683, 95% CI: 11.940 to 23.311). The results of the subgroup analyses indicated that age, BMI, population, continent, and reference standard were sources of heterogeneity, and could interactively affect the accuracy of the 2 tools.

In addition, more detailed subgroup analyses according to population type were carried out. The pooled sensitivity and specificity are exhibited in Tables S2 and S3. Among the various subgroup populations, the diagnostic performance of GLIM was most effective in non-cancer patients with an average BMI of <24 kg/m² (DOR: 561.088, 95% CI: 13.987 to 22507.470), followed by non-cancer patients with an average age of ≥60 years (DOR: 56.624, 95% CI: 8.468 to 378.623). PG-SGA was most effective in cancer patients with an average age of <60 years (DOR: 44.918, 95% CI: 18.338 to 110.021), followed by cancer patients with an average BMI of <24 kg/m² (DOR: 42.236, 95% CI: 17.317 to 103.015).

Result of Bayes analysis

Clinical utilization of the 2 tools for diagnosing malnutrition was evaluated using the LRs. Post-test probability was calculated based on Bayes’ theorem. The pre-test probability was assessed as 40%, which had previously been reported, and the corresponding post-test probability, determined using Fagan’s nomograms, is shown in Figure 4. The Fagan plots suggested that the GLIM is more effective in ruling in a diagnosis of malnutrition, with a 77% post-test probability of malnutrition following a “positive” GLIM result; a “positive” PG-SGA result is associated with a lower risk of malnutrition (post-test probability 72%). However, PG-SGA was more useful for ruling out a diagnosis of malnutrition, with a 10% post-test probability of malnutrition following a “negative” PG-SGA test, as compared with the GLIM (12%). Summing up, the GLIM and PG-SGA have different advantages in clinical application, but the gap is small.

Publication bias

The results of Deeks’ funnel plot and Deeks’ asymmetry test indicated that the likelihood of publication biases in the included studies were low for both the GLIM (P = .141, Figure S3A) and PG-SGA (P = .324, Figure S3B).

DISCUSSION

In the current study, a hierarchical Bayesian latent-class meta-analysis was carried out to evaluate the capacity of the GLIM and PG-SGA to diagnose malnutrition among patients in the absence of a gold standard. Our meta-analysis revealed that the GLIM had a pooled sensitivity of 0.833 (0.744, 0.896), and a pooled specificity of 0.837 (0.780, 0.882). In comparison, the PG-SGA was more sensitive, with a higher pooled sensitivity of 0.874 (0.797, 0.925), but less specific, with a lower pooled specificity of 0.778 (0.707, 0.836). Taken together, both the GLIM and PG-SGA, which are the preferred diagnostic tools for identifying malnutrition, showed fairly good sensitivity and specificity. The GLIM had a higher LR+ (5.124, 95% CI 3.797 to 6.914) compared with the PG-SGA (3.939, 95% CI 3.014 to 5.148), while the PG-SGA had a lower LR− (0.161, 95% CI 0.101 to 0.258) than the GLIM (0.199, 95% CI 0.128 to 0.309). These results indicated that the GLIM was more likely to correctly identify malnutrition patients, and the PG-SGA was less likely to misdiagnose non-malnutrition patients. In addition, the GLIM exhibited a slightly higher DOR value (25.791, 95% CI 14.574 to 45.640) than the PG-SGA (24.396, 95% CI 14.408 to 41.308), which suggested a slightly superior diagnostic value in the GLIM. More specifically, the diagnostic performance of GLIM was more accurate in non-cancer patients, especially those with a low BMI or older. PG-SGA was more powerful in cancer patients, especially those of younger age.

The diagnostic accuracy of GLIM observed in our research was better than that in a previous meta-analysis, which reported that the sensitivity and specificity of GLIM in diagnosing malnutrition were 0.72 (95% CI, 0.64 to 0.78) and 0.82 (95% CI, 0.72 to 0.88), respectively.⁶² The larger number of primary studies included in our meta-analysis may partly explain the discrepancy. In addition, the previous meta-analysis was conducted with a bivariate random effect model, which could not take into account the variation due to the absence of a true gold standard. There is evidence that ignoring the imperfect nature of the reference standard can lead to substantial bias in the pooled estimates in meta-analyses of screening accuracy.⁶³

Because some subgroups with fewer than 4 primary studies could not be converged using the HSROC model, such as cancer patients with an average BMI of ≥24 kg/m², our research could not thoroughly investigate the accuracy of GLIM and PG-SGA in each subset. According to the available data, the GLIM was most powerful in non-cancer patients with an average BMI of <24 kg/m², having the highest DOR value among several different subgroup populations (DOR: 561.088, 95% CI: 13.987 to 22507.470). In 2020, Henrique et al²⁶ conducted subset analyses and observed poorer diagnostic performance of GLIM in overweight and obese non-cancer patients, and their findings are supported by our results. In addition, GLIM showed excellent specificity among cancer patients, although the sensitivity was relatively inferior.

Ruan et al⁶⁴ carried out a meta-analysis comparing screening accuracy between the Mini Nutritional Assessment (MNA), the nutritional risk screening 2002 (NRS-2002), and the PG-SGA in cancer patients, using the same hierarchical Bayesian latent class model used in our study. They reported that, of the 3 nutritional screening tools, the PG-SGA had the best diagnostic performance among cancer patients. The sensitivity and specificity of PG-SGA in their findings were 0.964 (0.913, 0.986) and 0.905 (0.807, 0.956), much higher than in our results. We speculated that the difference might be due to the inferior applicability of the PG-SGA in non-cancer patients. It is acknowledged that PG-SGA is a nutritional assessment tool developed specifically for cancer patients. The speculation was confirmed in our subgroup analysis, which showed that the sensitivity and specificity of PG-SGA were indeed evidently higher in cancer patients (DOR: 29.628, 95% CI: 16.172 to 54.282) than in non-cancer patients (DOR: 16.771, 95% CI: 6.398 to 43.958). More specifically, PG-SA was most accurate in cancer patients with an average age of <60 years (DOR: 44.918, 95% CI: 18.338 to 110.021), among several types of patients, closely followed by cancer patients with an average BMI of <24 kg/m² (DOR: 42.236, 95% CI: 17.317 to 103.015).

Comparing the GLIM with PG-SGA, the GLIM had higher DOR values than PG-SGA in patients with an average age of ≥60 years, patients with an average BMI of <24 kg/m², and non-cancer patients. PG-SGA had higher DOR values than the GLIM in patients with an average age of <60 years, patients with an average BMI of ≥24 kg/m², and cancer patients. Our results could offer guidance to help clinicians find which test should be used for particular patient populations. Comparing the diagnostic performance of the GLIM and PG-SGA in cancer patients with different ages and BMI, PG-SGA showed better discrimination ability than the GLIM in cancer patients with an average of <60 years and patients with an average BMI of <24 kg/m². The GLIM displayed better diagnostic value than PG-SGA in non-cancer patients with an average age of ≥60 years and patients with an average BMI of <24 kg/m². Overall, our findings indicated that PG-SGA is more suitable for cancer patients, while the GLIM is more suitable for non-cancer patients.

To make our analysis more comprehensive, both the original diagnostic studies identified and the reference lists of previous literature reviews and meta-analyses were searched. By doing this, publication bias was reduced. In addition, this meta-analysis was conducted using a hierarchical Bayesian latent class model, which accounts for both within- and between-study variability. Hence, the tremendous variation resulting from using different reference standards was avoided. Nonetheless, the limitation of the lack of a gold standard reference also needed to be addressed in this study. Our inclusion criteria were not aimed at specific disease populations, which might be a source of heterogeneity. However, subgroup analysis was performed to evaluate the accuracy of the GLIM and PG-SGA among cancer patients and non-cancer patients.

CONCLUSION

In conclusion, the hierarchical Bayesian latent-class meta-analysis evidence demonstrates that both the GLIM and PG-SGA had moderately high evaluation performance in discriminating malnutrition. Moreover, the diagnostic performance of GLIM was more accurate in non-cancer patients, especially in those who had a low BMI or who were older. PG-SGA was more accurate for cancer patients, especially younger patients. These findings might assist clinicians in assessing malnutrition, to help guide personalized management strategies that are optimized for the patient’s nutritional status.

Author Contributions

T.W. and C.S. contributed to the concept and design; M.Z., K.X., and Y.Z. contributed to the acquisition and analysis of the data; S.Z., H.C., and P.G. contributed to the interpretation of the data; and T.W. drafted the manuscript. All authors critically revised the manuscript, agreed to be fully accountable for ensuring the integrity and accuracy of the work, and read and approved the final manuscript.

Supplementary Material

Supplementary Material is available at Nutrition Reviews online.

Funding

This work was supported by the National Key Research and Development Program (2022YFC2009600, 2022YFC2009601). The funders had a role in the study design.

Conflicts of Interest

None declared.

Data availability

The data set is available from the corresponding author upon request.

REFERENCES