Aggregate Index of Systemic Inflammation (AISI), Disease Severity, and Mortality in COVID-19: A Systematic Review and Meta-Analysis

Combined indices of different haematological cell types appear to be particularly promising for investigating the link between systemic inflammation and coronavirus disease 2019 (COVID-19). We conducted a systematic review and meta-analysis to assess the aggregate index of systemic inflammation (AISI), an emerging index derived from neutrophil, monocyte, platelet, and lymphocyte counts, in hospitalized COVID-19 patients with different disease severity and survival status. We searched electronic databases between the 1st of December 2019 and the 10th of June 2023 and assessed the risk of bias and the certainty of evidence. In 13 studies, severe disease/death was associated with significantly higher AISI values on admission vs. non-severe disease/survival (standard mean difference (SMD) = 0.68, 95% CI 0.38 to 0.97, p < 0.001). The AISI was also significantly associated with severe disease/death in five studies reporting odds ratios (4.39, 95% CI 2.12 to 9.06, p ˂ 0.001), but not in three studies reporting hazard ratios (HR = 1.000, 95% CI 0.999 to 1.002, p = 0.39). The pooled sensitivity, specificity, and area under the curve values for severe disease/death were 0.66 (95% CI 0.58 to 0.73), 0.78 (95% CI 0.73 to 0.83), and 0.79 (95% CI 0.76 to 0.83), respectively. Our study has shown that the AISI on admission can effectively discriminate between patients with different disease severity and survival outcome (PROSPERO registration number: CRD42023438025).


Introduction
There is very good evidence that immune dysregulation and a state of excessive systemic inflammation predispose patients with coronavirus disease 2019 (COVID- 19) to homeostatic alterations in various organs and tissues and an increased risk of severe clinical manifestations, prolonged hospitalization, and in-hospital mortality [1][2][3][4][5][6][7][8][9][10]. Although these observations have been instrumental in the identification of safe and effective immunomodulatory and anti-inflammatory treatment strategies [11,12], they have also stimulated the search for robust biomarkers of excessive inflammatory response for early risk stratification and appropriate management in this patient group [13,14]. Concomitantly, abnormalities in specific blood cell types, particularly, neutrophilia, lymphopenia, and thrombocytopenia, have led to studies reporting significant associations between derived indices, particularly the neutrophil-to-lymphocyte ratio (NLR), and severe disease and mortality [15][16][17][18]. The potential clinical use of another haematological index, the aggregate index of systemic inflammation (AISI, calculated by multiplying the neutrophil, monocyte, and platelet count, and dividing the product by the lymphocyte count), originally developed by Paliogiannis et al. in surgical patients [19], has been increasingly investigated in other disease states as well as COVID-19 [20][21][22][23][24][25].
Given the significant temporal changes in viral variants, pharmacological treatments, and vaccines during the COVID-19 pandemic, we critically appraised, by means of systematic review and meta-analysis, the potential utility of the AISI in discriminating between hospitalized COVID-19 patients with different severity and survival outcomes. We hypothesised that severe disease or in-hospital mortality were associated with higher AISI values at the time of admission. We also investigated possible associations between the effect size of the AISI values and various study and patient characteristics.

Materials and Methods
We searched Scopus, Web of Science, and PubMed for articles published from the 1st of December 2019 to the 10th of June 2023 using the following terms: (1) "AISI", (2) "aggregate index of systemic inflammation", (3) "COVID-19", (4) "2019-nCoV", (5) "SARS-CoV-2", and (6) "coronavirus disease 2019". We also hand searched the reference lists of each article for additional studies. The inclusion criteria were as follows: (a) investigating hospitalized patients with COVID-19 with different severity and survival outcomes; (b) reporting continuous data on AISI values in COVID-19 patients; (c) reporting multivariate adjusted odds ratio (OR) or hazard ratio (HR) with 95% confidence intervals (CI) for disease severity and/or mortality; (d) reporting prognostic accuracy (area under the receiver operating characteristic curve, AUROC, with 95% CIs); (e) full-text available; (f) age of participants ≥18 years; and (g) English language used. Two investigators independently reviewed the abstracts and full-text articles, with a third involved in case of disagreement.
The following information was extracted from each article: participant age and sex, publication year, sample size, study design, study country and continent, clinical endpoint studied (measures of disease severity and/or survival status), AISI, AUROC, sensitivity, specificity, cut-off, true positive (TP), false positive (FP), false negative (FN), and true negative (TN) values.
The Joanna Briggs Institute Critical Appraisal Checklist for case-control studies was used to assess the risk of bias (studies addressing <50%, ≥50% and <75%, and ≥75% of checklist items were adjudicated as having a high, moderate, and low risk, respectively [26]). The certainty of evidence was evaluated with the Grades of Recommendation, Assessment, Development and Evaluation (GRADE) Working Group system [27]. We complied with the PRISMA 2020 statement (Supplementary Tables S1 and S2) [28], and registered our study in the International Prospective Register of Systematic Reviews (PROSPERO registration number: CRD42023438025).

Statistical Analysis
Forest plots of continuous AISI values were generated using standardized mean differences (SMDs) and 95% CIs to assess differences between patients with non-severe disease/survivors (NSDS) and those with severe disease/non-survivors (SDNS) (p < 0.05 for statistical significance). Additional forest plots were generated using the ORs or HRs and 95% CIs of the multivariate associations between the AISI and disease severity and survival status. Heterogeneity was assessed using the Q statistic (p < 0.10 for statistical significance) and random-effect models were used accordingly [29]. Sensitivity analyses assessed the stability of the effect size [30]. Publication bias was assessed using (1) Begg's adjusted rank correlation test, (2) Egger's regression asymmetry test (p < 0.05 for statistical significance) [31,32], and (3) the "trim-and-fill" method [33]. Univariate meta-regression and subgroup analyses investigated associations between the effect size and age, sex, number of study participants, study continent, publication year, and study endpoint.
We generated a summary receiver operating characteristic (SROC) curve [34], used empirical Bayes estimates, and calculated the pooled sensitivity and specificity. The HSROC model was also used to account for heterogeneity [34][35][36][37]. Publication bias was assessed using the Deeks method [38]. The relationship between prior probability, likelihood ratio, and posterior test probability was assessed using the Fagan's nomogram plot [39]. Analyses were performed using Stata 14 (StataCorp LLC, College Station, TX, USA).
hood ratio, and posterior test probability was assessed using the [39]. Analyses were performed using Stata 14 (StataCorp LLC, Co
Using random-effects models (I 2 = 89.8%, p < 0.001), the AISI was not associated with mortality (HR = 1.000, 95% CI 0.999 to 1.002, p = 0.39; Figure 10). Assessment of publication bias, meta-regression, and subgroup analyses was not possible because of the relatively small number of studies. The level of certainty was downgraded to extremely low (rating 0, ⊝⊝⊝⊝) after considering the high heterogeneity, the relatively high imprecision, and the lack of assessment of publication bias.  [42,[44][45][46][47].
The level of certainty remained low (rating 2, ⊕⊕ ) after considering the low risk of bias in all studies, the high but partly explainable heterogeneity, the lack of indirectness, the relatively low imprecision, the relatively large effect size (OR = 4.39; upgrade one level), and the lack of assessment of publication bias (downgrade one level).
Using random-effects models (I 2 = 89.8%, p < 0.001), the AISI was not associated with mortality (HR = 1.000, 95% CI 0.999 to 1.002, p = 0.39; Figure 10). Assessment of publication bias, meta-regression, and subgroup analyses was not possible because of the relatively small number of studies. The level of certainty remained low (rating 2, ⊕⊕⊝⊝) after considering the low risk of bias in all studies, the high but partly explainable heterogeneity, the lack of indirectness, the relatively low imprecision, the relatively large effect size (OR = 4.39; upgrade one level), and the lack of assessment of publication bias (downgrade one level).
Using random-effects models (I 2 = 89.8%, p < 0.001), the AISI was not associated with mortality (HR = 1.000, 95% CI 0.999 to 1.002, p = 0.39; Figure 10). Assessment of publication bias, meta-regression, and subgroup analyses was not possible because of the relatively small number of studies. Figure 10. Forest plot of studies examining the association between AISI and COVID-19 severity/survival using hazard ratio [40,48,50].
The level of certainty was downgraded to extremely low (rating 0, ⊝⊝⊝⊝) after considering the high heterogeneity, the relatively high imprecision, and the lack of assessment of publication bias. Figure 10. Forest plot of studies examining the association between AISI and COVID-19 severity/survival using hazard ratio [40,48,50].
The level of certainty was downgraded to extremely low (rating 0, ) after considering the high heterogeneity, the relatively high imprecision, and the lack of assessment of publication bias.

Diagnostic Accuracy for Prediction of Severe Disease or Death
Eleven studies (including 15 patient groups) reported the diagnostic accuracy of the AISI towards severe disease or death in 7427 COVID-19 patients (56% males, mean age 58 years) (Table 3) [25,[40][41][42][43][45][46][47][48][49][50]. Six studies were conducted in Europe [25,[40][41][42]45,47], four in Asia [43,[48][49][50], and one in Africa [46]. Endpoints included mortality (eight study groups) [25,40,43,[47][48][49][50], admission to the intensive care unit (two study groups) [42,46], prolonged hospital stay (one study group) [41], acute limb ischemia (one study group) [42], invasive mechanical ventilation (one study group) [45], deep vein thrombosis (one study group) [47], and acute pulmonary embolism (one study group) [47]. The risk of bias was low in all studies (Supplementary Table S3) [25,[40][41][42][43][45][46][47][48][49][50]. The pooled sensitivity and specificity of the AISI for severe disease/death were 0.66 (95% CI 0.58 to 0.73) and 0.78 (95% CI 0.73 to 0.83), respectively ( Figure 11). The AUC was 0.79 (95% CI 0.76 to 0.83), with sensitivity of 0.66 and specificity of 0.78 ( Figure 12). The empirical Bayes estimates in HSROC analysis are shown in Figure 13. The midas command was used to evaluate the quantile plot of residual-based goodness-of-fit, the Chisquared probability plot of squared Mahalanobis distances for assessment of the bivariate normality assumption, the spikeplot for checking for particularly influential observations using Cook's distance, and a scatter plot for checking for outliers using standardized predicted random effects (Figure 14). The analysis identified one outlier [49]. Its removal resulted in an AUC value of 0.80, a sensitivity of 0.69, and a specificity of 0.79. predicted random effects ( Figure 14). The analysis identified one outlier [49]. Its removal resulted in an AUC value of 0.80, a sensitivity of 0.69, and a specificity of 0.79.   predicted random effects ( Figure 14). The analysis identified one outlier [49]. Its removal resulted in an AUC value of 0.80, a sensitivity of 0.69, and a specificity of 0.79.     The Fagan's nomogram, generated to assess the potential clinical utility of the AISI, showed that, assuming a pre-test probability of 25% for adverse outcomes, the post-test probability was 50% with relatively high AISI values and 13% with relatively low AISI values ( Figure 15).  The Fagan's nomogram, generated to assess the potential clinical utility of the AISI, showed that, assuming a pre-test probability of 25% for adverse outcomes, the post-test probability was 50% with relatively high AISI values and 13% with relatively low AISI values ( Figure 15). The Fagan's nomogram, generated to assess the potential clinical utility of the AISI, showed that, assuming a pre-test probability of 25% for adverse outcomes, the post-test probability was 50% with relatively high AISI values and 13% with relatively low AISI values ( Figure 15). There was no significant publication bias (p = 0.38) according to the Deeks funnel plot asymmetry test (Figure 16). The HSROC curve ( Figure 12) was symmetric, given the negative correlation coefficient between logit transformed sensitivity and specificity (HSROC model; −0.634, 95% CI −0.897 to −0.432), and the non-significant (p = 0.50) symmetry parameter β (0.209, 95% CI −0.394 to 0.812; data not reported in tables or graphs). This suggests the absence of between-study heterogeneity [34,36]. However, the visual representation of SROC ( Figure  12) suggests moderate heterogeneity (95% CI 0.76 to 0.83). Using midas, the pooled sensitivity and specificity showed an inconsistency (I 2 ) of 90.48 and 92.07%, respectively (Figure 11). Using the bivariate boxplot with logit_Se and logit_Sp (Figure 17), four studies fell outside the circles [41,45,49,50], which indicates the presence of heterogeneity across studies. None of the investigated parameters exhibited significant associations with the effect size for sensitivity or specificity ( Figure 18). The HSROC curve ( Figure 12) was symmetric, given the negative correlation coefficient between logit transformed sensitivity and specificity (HSROC model; −0.634, 95% CI −0.897 to −0.432), and the non-significant (p = 0.50) symmetry parameter β (0.209, 95% CI −0.394 to 0.812; data not reported in tables or graphs). This suggests the absence of between-study heterogeneity [34,36]. However, the visual representation of SROC ( Figure 12) suggests moderate heterogeneity (95% CI 0.76 to 0.83). Using midas, the pooled sensitivity and specificity showed an inconsistency (I 2 ) of 90.48 and 92.07%, respectively ( Figure 11). Using the bivariate boxplot with logit_Se and logit_Sp (Figure 17), four studies fell outside the circles [41,45,49,50], which indicates the presence of heterogeneity across studies. None of the investigated parameters exhibited significant associations with the effect size for sensitivity or specificity ( Figure 18).

Discussion
In our study, the AISI values on admission were significantly higher in hospitalized SDNS vs. NSDS COVID-19 patients. The AISI between-group differences were significant when expressed either as SMDs or as ORs, whereas the lack of significant differences in studies reporting HRs is likely to be secondary to their limited number (n = 3). Importantly, the AISI exhibited good diagnostic performance towards severe disease or mortality, with an AUC of 0.79. Sensitivity analysis showed that the results of the meta-analysis were stable. There were non-significant associations in meta-regression between the effect size and either age, sex, publication year, or sample size. In particular, the absence of significant associations with the year of publication suggests that the capacity of the AISI to discriminate between COVID-19 patients with different disease severity and survival status persisted during different phases of the COVID-19 pandemic, regardless of study participants with different vaccines, viral variants, and treatments. In subgroup analysis, there were differences in effect size with study continent, particularly for ORs, indicating possible ethnicity-related effects on the association between the AISI and COVID-19.
Compared to other indexes of inflammation derived from haematological parameters, e.g., NLR and the platelet-to-lymphocyte ratio (PLR), the AISI is calculated using information from four types of blood cell that are involved in inflammation, i.e., neutrophils, monocytes, platelets, and lymphocytes. The AISI was initially investigated in 2018 to predict outcomes in surgical patients [19]. Since then, studies have assessed the potential clinical utility of the AISI in patients with other disease states characterized by a systemic pro-inflammatory state, such as macular degeneration [20], idiopathic pulmonary fibro-sis [21,23], and cancer [24]. Notably, in patients with idiopathic pulmonary fibrosis, the capacity of the AISI to predict for four-year survival was superior to white blood cell, neutrophil, monocyte, lymphocyte, and platelet count taken singly, and the NLR and the PLR, further supporting its potential clinical utility [23]. Over the last three years, an increasing number of studies have also investigated the AISI in patients with COVID-19, given the critical pathophysiological and prognostic role of excess systemic inflammation in terms of severe clinical manifestations, multiorgan involvement, requirement for intensive care, prolonged hospitalization, and in-hospital mortality [1][2][3][4][5][6][7][8][9][10].
The comprehensive appraisal of the available evidence strongly supports the concept that assessing the AISI at the time of admission can provide useful information to rapidly direct subgroups of patients with COVID-19 to different treatment and monitoring pathways. This proposition is further supported by the observed pooled AUC value and the Fagan's nomogram, which indicated a wide separation of the post-test probability of severe disease/death compared to the pre-test probability, according to whether the admission AISI values were relatively low or relatively high [52]. However, it should be emphasised that further studies are required before routinely using the AISI to assess COVID-19 patients. Specifically, prospective studies should investigate whether the AISI can predict, with or without other clinical parameters, ethnicity, and specific comorbidities, disease severity and mortality in this group, in a similar way to other conditions [53].
One limitation of our study is the moderate-to-high between-study heterogeneity. However, we identified specific sources of heterogeneity when investigating the SMD (study continent) and the OR (study continent). Furthermore, publication bias could not be assessed with the OR and the HR because of the limited number of relevant studies. In contrast, significant strengths are the comprehensive assessment of the significance of the AISI with meta-regression and subgroup analysis, SROC, and Fagan's nomogram.

Conclusions
In our study, higher AISI values on admission were significantly associated with severe disease/death in COVID-19. Further studies in patients with a wide range of comorbidities and ethnic backgrounds are warranted to determine whether the AISI can routinely assist with discriminating between COVID-19 patients with different severity and outcomes in order to optimize management and outcomes.