Association of Prognostic Nutritional Index with Severity and Mortality of Hospitalized Patients with COVID-19: A Systematic Review and Meta-Analysis

The associations of prognostic nutritional index (PNI) with disease severity and mortality in patients with coronavirus disease 2019 (COVID-19) remain unclear. Electronic databases, including MEDLINE, EMBASE, Google scholar, and Cochrane Library, were searched from inception to 10 May 2022. The associations of PNI with risk of mortality (primary outcome) and disease severity (secondary outcome) were investigated. Merged results from meta-analysis of 13 retrospective studies (4204 patients) published between 2020 and 2022 revealed a lower PNI among patients in the mortality group [mean difference (MD): −8.65, p < 0.001] or severity group (MD: −5.19, p < 0.001) compared to those in the non-mortality or non-severity groups. A per-point increase in PNI was associated with a reduced risk of mortality [odds ratio (OR) = 0.84, 95% CI: 0.79 to 0.9, p < 0.001, I2 = 67.3%, seven studies] and disease severity (OR = 0.84, 95% CI: 0.77 to 0.92, p < 0.001, I2 = 83%, five studies). The pooled diagnostic analysis of mortality yielded a sensitivity of 0.76, specificity of 0.71, and area under curve (AUC) of 0.79. Regarding the prediction of disease severity, the sensitivity, specificity, and AUC were 0.8, 0.61, and 0.65, respectively. In conclusion, this study demonstrated a negative association between PNI and prognosis of COVID-19. Further large-scale trials are warranted to support our findings.


Introduction
The outbreak of the coronavirus disease 2019 (COVID- 19) pandemic since the end of 2019 has already claimed millions of lives, as well as imposed enormous threats to healthcare systems and economies worldwide [1]. Despite mild symptoms in the majority of patients contracting the disease, a significant population still experienced severe symptoms and later developed life-threatening complications including acute respiratory distress

Materials and Methods
We report this systematic review in compliance with the Preferred Reporting Items for Systematic Reviews and Meta-analysis statement. Our study protocol has been registered in the International Prospective Register of Systematic Review [CRD42022331321]. For improving the quality of the current meta-analysis, study selection, data collection, and risk of bias assessment were independently executed by two authors. All disagreements were settled by discussion. The protocol and procedures of the current meta-analysis has been described in our previous study [22].

Data Sources and Searches
Using different combinations of keywords and MeSH terms, we searched on 10 May 2022 for studies reporting the association between PNI and prognostic outcomes in adult patients diagnosed with COVID-19 from the following databases: Medline, Embase, and Cochrane Library. Additionally, a manual search was performed in Google scholar to identify relevant articles. Supplemental Table S1 depicts the process of the literature search, taking Medline as an example. There was no restriction on language, year of publication, and sample size when conducting the database search. For a completeness of our search, we examined the reference lists of the acquired articles and published meta-analyses to further retrieve eligible studies for the current investigation.

Study Selection and Data Extraction
We included studies that fulfilled the following criteria: (1) observational studies including cross-sectional, cohort study, and case-control study designs, (2) adult patients who were diagnosed with COVID-19 and admitted to hospital, (3) available PNI values at hospital admission, (4) reporting of the association of PNI values with disease severity or mortality, and (5) studies with adequate details for the calculation or extraction of individual odds ratio (OR) as well as corresponding 95% confidence intervals (CIs). We excluded studies that were (1) duplicated, (2) conducted in the pediatric population or those not admitted to hospital, (3) presented as conference abstracts, animal studies, systematic reviews, case reports, and editorials or commentaries as well as different forms of publication other than the original investigation.

Data Extraction
The following items were retrieved from each study: first author/publication year, age, male gender, number of patients, prognostic outcomes (i.e., mortality and disease severity), PNI values, sensitivity/specificity, negative predictive value (NPV), positive predictive value (PPV), and country. We acquired the odds ratio (OR) and 95% confidence intervals (CI) from matched or adjusted data for each study. For studies that provided both unadjusted and adjusted ORs, we adopted the adjusted ORs. On encountering a categorical variable from the dichotomization of a continuous variable, we calculated the OR by the number of cases and controls with exposure to the prognostic factors (e.g., PNI) according to the cut-off point reported in that study. The authors of our included articles with missing data were contacted for additional information.

Outcomes and Definitions
Our primary outcome was the correlation between PNI and the all-cause mortality during hospitalization, while the association of PNI with disease severity and the diagnostic efficacy of PNI for mortality and disease severity served as the secondary outcomes. The definition of disease severity was based on that in each study. Investigations into the associations of other risk factors (e.g., age and biomarkers) with prognostic outcomes, which were hampered by the limited number of available studies, were not performed in the current meta-analysis.

Assessment of Risks of Bias for the Included Studies
The risk of bias for each study was independently reviewed by two authors in accordance with the six domains of the Quality in Prognostic Studies (QUIPS) tool, namely, study participation, outcome measurement, study attrition, prognostic factor measurement, adjustment for other prognostic factors, and statistical analysis and reporting [23]. For each domain, the risk of a study was classified as low, unclear, or high. The overall risk of bias of a study was deemed low when all or most of the domains were judged to be low (or low to moderate) in that study [24].

Data Synthesis and Analysis
To assess the association between PNI and the prognostic outcomes, three different approaches were applied. First, the preoperative PNI was compared between the survival and non-survival groups; second, the association between PNI and the risk of mortality was investigated, with the PNI values serving either as a binary variable (i.e., high vs. low) or a continuous parameter. The same approach was used to assess the association between PNI and disease severity, where appropriate. Because the present meta-analysis was based on observational studies, an overall OR produced using a random-effects model served as the main summary measure of effect size. We evaluated statistical heterogeneity of effect size with I 2 statistics and defined the substantial heterogeneity as an I 2 over 50% [25]. Sensitivity analysis through omitting one study in turn was conducted to examine the reliability and robustness of the available evidence. Potential publication bias was detected through the inspection of a funnel plot and examination with Egger's tests regarding a particular outcome reported in 10 or more studies. The statistical analyses were conducted with the comprehensive Meta-Analysis (CMA) V3 software (Biostat, Englewood, NJ, USA).
To evaluate the accuracy of PNI for predicting mortality or disease severity, we calculated the pooled estimates of sensitivity and specificity based on the bivariate model [26]. After generation of a hierarchical summary receiver operating characteristic (hsROC) curve, the area under the curve (AUC) was used to determine the overall accuracy according to the summary receiver operating characteristic (sROC) curve. Forest plots of pooled sensitivity and specificity, sROC curve, and Deeks' funnel plot for assessing the publication bias were generated using the MIDAS command in Stata 15 (StataCorp LLC., College Station, TX, USA). A p value of <0.05 was considered statistically significant in the current meta-analysis.

Study Selection
The flow diagram depicting the process of the study selection is shown in Figure 1. Of a total of 182 records retrieved on the initial database search, 164 were preserved after excluding 18 duplicates. Following the further exclusion of 144 records during screening of their titles and abstracts, six more articles were excluded after full-text assessment because of being a review article (n = 1), abstract (n = 3), lack of available outcomes (n = 2), or focusing on disease progression instead of severity or mortality (n = 1) [27]. Finally, 13 studies involving 4204 patients published between 2020-2021 were eligible for quantitative syntheses [11,[19][20][21][28][29][30][31][32][33][34][35][36]. detected through the inspection of a funnel plot and examination with Egger's tests regarding a particular outcome reported in 10 or more studies. The statistical analyses were conducted with the comprehensive Meta-Analysis (CMA) V3 software (Biostat, Englewood, NJ, USA).
To evaluate the accuracy of PNI for predicting mortality or disease severity, we calculated the pooled estimates of sensitivity and specificity based on the bivariate model [26]. After generation of a hierarchical summary receiver operating characteristic (hsROC) curve, the area under the curve (AUC) was used to determine the overall accuracy according to the summary receiver operating characteristic (sROC) curve. Forest plots of pooled sensitivity and specificity, sROC curve, and Deeks' funnel plot for assessing the publication bias were generated using the MIDAS command in Stata 15 (StataCorp LLC., College Station, TX, USA). A p value of <0.05 was considered statistically significant in the current meta-analysis.

Study Selection
The flow diagram depicting the process of the study selection is shown in Figure 1. Of a total of 182 records retrieved on the initial database search, 164 were preserved after excluding 18 duplicates. Following the further exclusion of 144 records during screening of their titles and abstracts, six more articles were excluded after full-text assessment because of being a review article (n = 1), abstract (n = 3), lack of available outcomes (n = 2), or focusing on disease progression instead of severity or mortality (n = 1) [27]. Finally, 13 studies involving 4204 patients published between 2020-2021 were eligible for quantitative syntheses [11,[19][20][21][28][29][30][31][32][33][34][35][36].

Primary Outcome-Association of PNI with Mortality
The association between the risk of mortality and PNI was investigated in three different approaches; while seven studies compared the PNI between the mortality and nonmortality groups, the correlation between PNI and the risk of mortality was assessed with the former serving either as a continuous parameter or as a binary variable (i.e., high vs. low) in seven and six studies, respectively.
First, based on the PNI in the mortality and non-mortality groups in seven studies [19,21,28,29,32,34,36], the merged results revealed a lower mean PNI among patients in the mortality group (MD: −8.65, 95% CI: −11.81 to −5.49, p < 0.001, I 2 = 95.3%) compared to that in the non-mortality group (Figure 3a). Second, when PNI was used as a continuous measure to predict the risk of mortality in seven studies [11,[19][20][21]29,34,36], the pooled results demonstrated that a higher PNI was related to a lower risk of mortality (OR: 0.84, 95% CI: 0.79 to 0.9, p < 0.001, I 2 = 67.3%) (Figure 3b) with the demonstration of a correlation between an additional increase in one unit in PNI and an 16% decrease in the odds of mortality. Third, an analysis of the link between PNI as a binary variable (i.e., high vs. low) and the risk of mortality in six studies [11,20,28,29,32,36] also revealed an association of a lower PNI with a higher mortality risk (OR: 7, 95% CI: 3.44 to 14.23, p < 0.001, I 2 =

Primary Outcome-Association of PNI with Mortality
The association between the risk of mortality and PNI was investigated in three different approaches; while seven studies compared the PNI between the mortality and non-mortality groups, the correlation between PNI and the risk of mortality was assessed with the former serving either as a continuous parameter or as a binary variable (i.e., high vs. low) in seven and six studies, respectively.

Secondary Outcome-Association of PNI with Disease Severity
The merged results from five studies with information on PNI in the severity and non-severity groups [19,21,33,35,36] demonstrated a lower mean PNI among patients in the severity group (MD: −5.19, 95% CI: −6.89 to −3.49, p < 0.001, I 2 = 73.4%) compared to that in the non-severity group (Figure 4a). Using another approach, in which PNI served as a continuous measure to predict disease severity, the pooled results from five studies [19,30,31,35,36] also demonstrated a negative correlation between PNI and disease severity (OR: 0.84, 95% CI: 0.77 to 0.92, p < 0.001, I 2 = 83%) (Figure 4b). The sensitivity analysis demonstrated the robustness of the results from the two approaches.

The Use of PNI for Predicting Mortality and Disease Severity: Pooled Estimates of Sensitivity/Specificity and sROC
The pooled sensitivity and specificity of using PNI for the prediction of mortality  [11,19,20,21,28,32,34,36], and (b) hierarchical summary receiver operating characteristic (hsROC) curves of using PNI for the prediction of in-hospital mortality in patients with COVID-19. SENS: Sensitivity, SPEC: Specificity, SROC: Summary receiver operating characteristic, and AUC: Area under the curve.

Discussion
With the staggering increase in the number of patients diagnosed with COVID-19, the minimization of disease mortality and severity as well as a rational allocation of medical resources have become priorities in medical systems worldwide. Identification of patients at high risks of mortality and complications can guide medical decisions to facilitate timely implementation of individualized therapeutic strategies [38]. Our results not only supported the use of PNI as a predictor of mortality with a pooled sensitivity of 0.76 and specificity of 0.71 (AUC of sROC: 0.79) but also demonstrated a negative association between PNI and disease severity in patients with COVID-19. Taking into consideration the simplicity of calculation, our findings suggested that PNI may be a cost-effective indicator for medical resource allocation during the pandemic.
Regarding the factors associated with COVID-19 disease severity and mortality, a previous meta-analysis has underscored an increased susceptibility to severe diseases in those with cardiovascular diseases as well as those with chronic respiratory and renal illnesses [39,40]. Other reported indicators of disease progression and mortality also included serological biomarkers such as C-reactive protein (CRP), neutrophil-lymphocyte ratio (NLR), ferritin, troponin, and lymphocyte count [41,42]. On the other hand, although several studies demonstrated an association between a poor nutritional status and an elevated risk of in-hospital death in patients infected with COVID-19 [11,12], the pooled evidence between the PNI and prognosis of COVID-19 was rarely investigated. Our finding demonstrated that a low PNI correlated with a seven-fold increased risk of mortality in hospitalized patients with COVID-19.
There were several potential confounding factors for the current investigation. Because an alleviation of disease severity and mortality through an intensive COVID-19 vaccination program may be a confounder of our outcomes, the exclusive inclusion of studies conducted in 2020 when the vaccination rates in the studied countries were still low may help in minimizing the impact of this factor in the current meta-analysis. Besides vaccination, the overall mortality of COVID-19 has also been demonstrated to be different between men and women, with the former being 2.3 times higher than the latter. Moreover, ethnically, the mortality and disease severity rates have been reported to be lower in Asians, as compared with other populations [43,44]. Therefore, in addition to the known influences of gender and geographical location on immune responses (e.g., cytokine expressions) [6], the inclusion of studies from a limited number of countries (e.g., China and Turkey) with a comparable proportion of males and females in most studies (no relevant information in one study) may offer a high homogeneity to minimize the related biases in the current meta-analysis.
In an attempt to identify prognostic markers for COVID-19 severity and mortality, a previous study suggested a better predictive outcome when combining indicators of inflammation with those of nutritional status than that from inflammation biomarkers alone [31]. The association of malnutrition with a poor prognosis in patients with COVID-19 has been reflected in the finding of a previous study, in which patients with COVID-19 who were admitted to ICU were found to be at risk for moderate (69.9%) to severe (12.3%) malnutrition [12]. Several factors have been reported to contribute to malnutrition in patients infected with COVID-19, including elevated catabolism from fever and exertion of respiratory muscles as well as endocrinological perturbations that lead to accelerated gluconeogenesis, protein breakdown, and lipid oxidation [11]. PNI, which is readily obtainable through routine peripheral blood examinations, comprises both nutritional and inflammatory components to serve this purpose. Interestingly, although most studies in the current meta-analysis enrolled non-critical hospitalized patients, we still found a significantly negative association between PNI and prognosis that may suggest its usefulness as an early marker of disease progression before an overt deterioration of nutritional status.
The finding of a significant negative correlation between PNI and COVID-19 severity despite the adoption of different definitions for disease severity in our included studies highlighted a potential role of PNI in COVID-19 severity prediction. With an explosive increase in the number of patients with a confirmed COVID-19 diagnosis, patient triage is of the utmost importance. As the symptoms of COVID-19 vary widely, ranging from no symptoms to critical illness requiring urgent medical attention [45,46], an accurate prediction of disease progression is essential to guiding medical resource allocation to reduce mortality [47]. For instance, assigning patients with a relatively minor condition (e.g., mild pneumonia) to ambulatory care and those with a severe disease to the intensive care unit (ICU) [48]. In countries where medical resources are limited, a simple feasible indicator of disease severity may be critical for the effective allocation of medical supplies and manpower.
Notwithstanding the known adverse impact of malnutrition on disease progression and mortality of COVID-19, age may be a significant confounder. A large-scale prospective cohort study on 20,133 inpatients contracting COVID-19 identified advanced age and obesity as independent risk factors of COVID-19-related mortality [49]. Besides, the susceptibility to malnutrition in older people has been demonstrated in previous studies that exhibited a prevalence of malnutrition as high as 35-65% in aged hospitalized patients and 25-60% in older institutionalized adults [50,51]. These findings, together with the elevated mortality rate in the aged population infected with COVID-19, underscored the importance of age as a confounder factor in the present study. Nevertheless, a previous study has still demonstrated PNI as an independent predictor of COVID-19 severity after adjustment for gender, body mass index, and age [35], highlighting the applicability of this indicator in different clinical scenarios.
Although our results demonstrated robust evidence on the negative association of PNI with mortality and disease severity, several issues remain to be clarified. First, as a variety of vaccines are currently available, the correlation between PNI and prognosis in patients receiving vaccine prophylaxis should be re-evaluated. Second, as malnutrition may be a modifiable risk factor, the impacts of nutrition supplementation on mortality and disease severity in patients with COVID-19 need to be assessed. Third, although PNI has previously been used to predict the prognostic outcomes in patients with various cancers [52,53], our included study did not specifically focus on patients with cancer. Therefore, further studies are required to investigate the predictive value of PNI regarding mortality and disease severity in this subgroup of patients after being infected with COVID-19.
There are several limitations in the current study. First, the major shortcoming of our meta-analysis was the retrospective nature of data collection that rendered our results susceptible to various confounding factors. For instance, no information about viral load and treatment options, which are known determinants of the duration of hospital stay and disease severity, was available for analysis [54]. Second, although previous studies reported a superior accuracy of mortality prediction in patients with COVID pneumonia by using a combination of different prediction indices compared to the use of a single indicator [31,55], our meta-analysis could not address this issue. In addition, we also did not compare the efficacy of PNI in mortality prediction with that of other predictive tools. Third, considering the significant differences in mortality among countries and phases of the pandemic [4], our inclusion of studies mainly from China and Turkey may limit the extrapolation of our findings to other ethnic groups and geographical locations. Fourth, because previous pooled evidence has already suggested associations of other confounding factors (e.g., age and comorbidities) with mortality and disease severity in patients contracting COVID-19, the predictive values of such confounders were not evaluated in the present study. In addition, although some other health-related indicators such as history of alcohol consumption, smoking, physical activities, sleeping, working environment and social routine among hospitalized patients could also have influenced our study outcomes, there were only three studies that included information about smoking [11,20,35], in which only two mentioned alcohol consumption [11,20]. Moreover, the lack of data about quantity and duration in those studies precluded the conduction of a subgroup analysis to elucidate the significance of their effects on our results. Finally, our results may be biased by variations in the definitions of disease severity and cut-off values of PNI for mortality prediction among our included studies.

Conclusions
In conclusion, this meta-analysis of 13 cohort retrospective studies on hospitalized patients with COVID-19 supported the use of a prognostic nutritional index as a promising index in predicting mortality and disease severity. Further large-scale studies are warranted to investigate the potential benefits of incorporating this index into clinical practice to improve prognostic outcomes in patients diagnosed with COVID-19.