Our systematic review and meta-analysis was performed according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [14]. We searched the following electronic databases from their inception to May 5, 2020: PubMed, Embase, Cochrane Library, Web of Science, Scopus, Wanfang and China National Knowledge Infrastructure database. The following combined search terms were used: (“Coronavirus disease 2019” OR “Coronavirus 2019” OR “COVID-19” OR “COVID19” OR “Severe acute respiratory syndrome coronavirus 2” OR “SARS-CoV-2” OR “nCoV-2019” OR “2019-nCoV” OR “Novel coronavirus”) AND (“Mortality” OR “Death” OR “Dead” OR “Fatality” OR “Non-survival” OR “Non-survivors” OR “Non-survivor” OR “Prognosis” OR “Deceased”). We did not apply any restriction on language or study design. We also manually reviewed the reference lists of all included studies, major reviews and meta-analyses for further potentially relevant articles [15]. The identifier of systematic review registration was PROSPERO CRD42020184995.

Studies reporting the liver biochemical parameters in COVID-19 patients were included if they met the following criteria: (1) patients were diagnosed as COVID-19 and could be divided into survivors group or non-survivors group; (2) continuous levels or dichotomous outcomes of liver biochemical parameters were available between survivors and non-survivors group. Liver biochemical parameters included AST, alanine transaminase (ALT), total bilirubin (TBIL), albumin (ALB), LDH, globulin (GLB), gamma-glutamyl transpeptidase (GGT), direct bilirubin (DBIL), total protein, alkaline phosphatase (ALP), ALB-to-GLB ratio (A/G) and prealbumin. If there were two or more studies from the same authors or institutions, only the study with the largest sample size was chosen. To be specific, two studies [16,17] were excluded because they are from same authors and had smaller sample size, compared with these two references [18,19]. Studies were excluded if patients were asymptomatic carriers and did not fulfill the inclusion criteria.

All the records from the initial search were exported to bibliographic files and then imported into Endnote X9 to exclude any duplicate and irrelevant studies. The following data were extracted: first authors, publication date, country of origin, grouping situation, cases, age, sex and continuous levels or dichotomous outcome of liver biochemical parameters in survivors and non-survivors groups. Stratified data or interquartile range (IQR) were converted to mean (SD) based on the mathematical formulas for meta-analysis [20,21]. Newcastle-Ottawa Scale (NOS) was used for quality assessments of all eligible studies. Studies with NOS stars lower than 7 were regarded as studies with inferior quality and therefore excluded. Two authors (GD and FZ) independently selected and evaluated the included articles. Disagreements were resolved by discussion.

All the statistical analyses were carried out by STATA (Version 12.0; STATA Corporation, College Station, TX, USA) software. Weighted mean difference (WMD) with 95% confidence intervals (95% CI) was calculated for the continuous levels of liver biochemical parameters in survivors and non-survivors groups. Standard mean difference (SMD) was used to explore the consistence of the conclusion. Odd risk (OR) with 95% CI was calculated for the dichotomous outcomes of liver biochemical parameters in survivors and non-survivors groups. In the case of heterogeneity among studies (I2>50% or P<0.1), random-effect model was used and would give a more conservative estimate of the 95% CI. In other cases, fixed-effect model was adopted (I2≤50% and P≥0.1) [22]. Sensitivity analysis was performed by omitting one study each time through influence analysis to assess the stability of results. Egger test was used to evaluate publication bias and the Duval and Tweedie trim-and-fill method was implemented to adjust for this bias [15]. P<0.05 was considered statistically significant.

A total of 5848 records were initially identified by our searchers. After exclusion of duplicates, 2539 studies remained. After review of the titles and abstracts, 63 studies were considered to meet the criteria for full-text scanning. After removal of reviews, duplicate articles with smaller sample size, and studies which did not report liver biochemical parameters, 23 studies met the inclusion criteria [18,19,23–43]. Additionally, two studies were enrolled after review of the reference lists of all included studies and major review and meta-analyses [44,45]. Finally, 25 studies involving 5971 COVID-19 patients were included in our analysis [18,19,23–45] (Fig. 1).

Fig. 1.

Study selection.

(0.44MB).

The main characteristics of eligible studies were presented in Table 1. All these studies were published in 2020 and could be divided into non-survivors group and survivors group. Among them, 23 studies were from China [18,19,23–34,36–43,45], one from the United States [35] and one from Spain [44]. Among the 23 studies from China, 20 studies were written in English and three in Chinese [18,28,42]. In general, patients in non-survivors group were older than those in survivors group except for Zhang's study [42], and most of studies showed more male patients in non-survivors than survivors group. All studies were deemed of high quality with 7 or more NOS scores and details were shown in Table A.1.

Twelve studies compared the levels of AST in non-survivors and survivors groups [18,19,24–27,31,32,34,36,37,45]. Random-effect model suggested that survivors had lower level of AST than non-survivors (WMD=−16.71U/L, 95% CI=[−21.03,−12.40], P<0.001) with significant heterogeneity (I2=67.1%, P<0.001) (Fig. 2A). Random-effect model based on SMD increased the heterogeneity and did not change the conclusion (Table 2). Sensitivity analysis also arrived at a similar conclusion by omitting one study each time (Fig. 2B). Egger test did not observe the publication bias (Fig. 2C). Moreover, five studies reported the proportion of increased AST in these two groups [23–25,33,41]. With heterogeneity (I2=57.9%, P=0.050), random-effect model showed that survivors had lower proportion of increased AST than non-survivors (OR=0.25, 95% CI=[0.15,0.41], P<0.001) (Fig. 2D). Similar results were found through fixed-effect model and sensitivity analysis, and no publication bias was detected by Egger test (Fig. 2D and F).

AST. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of AST between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased AST between survivors and non-survivors.'> AST. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of AST between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased AST between survivors and non-survivors.' title='Forest plot, sensitivity analyses and publication bias assessment of AST. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of AST between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased AST between survivors and non-survivors.'/>

Fig. 2.

Forest plot, sensitivity analyses and publication bias assessment of AST. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of AST between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased AST between survivors and non-survivors.

(0.69MB).

Fourteen studies provided the levels of ALT between survivors and non-survivors group [18,19,24–27,31,32,34,36,37,39,43,45]. With heterogeneity (I2=55.7%, P=0.006), random-effect model was used and showed that survivors had lower level of ALT than non-survivors (WMD=−5.20U/L, 95% CI=[−8.00,−2.41], P<0.001) (Fig. 3A). Random-effect model based on SMD and sensitivity analysis did not change the conclusion, and Egger test did not detect the presence of publication bias (Table 2 and Fig. 3B and C). Eight studies described the proportion of increased ALT in the two groups [23–25,28,33,39,41,43]. The heterogeneity test showed low heterogeneity among these studies (I2=25.1%, P=0.229), and fixed-effect model was used for the meta-analysis. The results demonstrated that survivors had lower proportion of increased ALT than non-survivors (OR=0.49, 95% CI=[0.37,0.64], P<0.001) without publication bias, and sensitivity analysis also did not change the conclusion (Fig. 3D–F).

ALT. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of ALT between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased ALT between survivors and non-survivors.'> ALT. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of ALT between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased ALT between survivors and non-survivors.' title='Forest plot, sensitivity analyses and publication bias assessment of ALT. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of ALT between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased ALT between survivors and non-survivors.'/>

Fig. 3.

Forest plot, sensitivity analyses and publication bias assessment of ALT. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of ALT between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased ALT between survivors and non-survivors.

(0.75MB).

Ten studies described the level of TBIL between survivors and non-survivors group [25,27,31,32,34–36,38,39,45]. No significant heterogeneity was found among these studies and a fixed-model effect was selected (I2=0%, P=0.452). The results showed that survivors had lower level of TBIL than non-survivors without the presence of publication bias (WMD=−4.40μmol/L, 95% CI=[−5.11,−3.70], P<0.001). Fixed-effect model based on SMD, random-effect model, and sensitivity analysis did not change the conclusion (Table 2 and Fig. 4A–C). Additionally, three studies reported the proportion of increased TBIL [23,33,39]. Fixed-effect model showed lower proportion of increased TBIL in survivors than non-survivors (OR=0.20, 95% CI=[0.12,0.34], P<0.001) without detecting the heterogeneity (I2=0%, P=0.999) and publication bias. The conclusion was consistent through sensitivity analysis (Fig. 4D–F).

TBIL. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of TBIL between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased TBIL between survivors and non-survivors.'> TBIL. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of TBIL between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased TBIL between survivors and non-survivors.' title='Forest plot, sensitivity analyses and publication bias assessment of TBIL. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of TBIL between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased TBIL between survivors and non-survivors.'/>

Fig. 4.

Forest plot, sensitivity analyses and publication bias assessment of TBIL. (A–C) Forest plot (A), sensitivity analyses (B) and Egger test (C) of continuous levels of TBIL between survivors and non-survivors. (D–F) Forest plot (D), sensitivity analyses (E) and Egger test (F) of the proportion of abnormally increased TBIL between survivors and non-survivors.

(0.62MB).

Twelve studies reported the relationship between the level of ALB with COVID-19 mortality [24,25,27,29,31,34–36,39,40,43,45]. With heterogeneity (I2=73.9%, P<0.001), random-effect model was used and suggested that survivors had higher level of ALB than non-survivors (WMD=4.47g/L, 95% CI=[3.47,5.47], P<0.001) with the absence of publication bias. The conclusion was not changed by random-effect model based on SMD or sensitivity analysis (Table 2 and Fig. A.1A–C). What is more, as for the proportion of decreased ALB, random-effect model showed that survivors had lower proportion of decreased ALB than non-survivors (OR=0.16, 95% CI=[0.07, 0.38], P<0.001) with heterogeneity (I2=62.1%, P=0.048). Sensitivity analysis did not change the conclusion, and Egger test did not show the presence of publication bias (Fig. A.1D–F).

Twelve studies depicted the level of LDH between survivors and non-survivors group [18,19,24,25,29,31,32,34,36,43–45]. Considering the heterogeneity among studies (I2=60.6%, P=0.003), random-effect model was used and suggested that survivors had lower level of LDH than non-survivors (WMD=−252.44U/L, 95% CI=[−289.57,−215.30], P<0.001). The conclusion did not change through random-effect model based on SMD or sensitivity analysis, and no publication bias was detected (Table 2 and Fig. A.2A–C). Besides, eight studies reported the proportion of increased LDH between these two groups [23–25,30,33,41–43]. With heterogeneity (I2=77.5%, P<0.001), random-effect model revealed lower proportion of increased LDH in survivors than non-survivors group (OR=0.09, 95% CI=[0.06,0.14], P<0.001). Sensitivity analysis did not change the conclusion and Egger test did not detect the presence of publication bias (Fig. A.2D–F).

Three studies reported the levels of GLB [24,34,36] and GGT [18,25,27] between survivors and non-survivors group, respectively. With low heterogeneity (I2=23.8%, P=0.269), fixed-effect model showed that survivors had lower level of GLB than non-survivors (WMD=−4.12g/L, 95% CI=[−5.45,−2.79, P<0.001). Random-effect model based on SMD changed the conclusion while the heterogeneity became larger, suggesting that using WMD to analyze the pooled data was more recommended. Moreover, sensitivity analysis did not change the conclusion, and Egger test did not observe the publication bias (Table 2 and Fig. A.3A–C). As for GGT, random-effect model did not find the difference between survivors and non-survivors group (WMD=−8.98U/L, 95% CI=[−26.22,8.26], P=0.310) with great heterogeneity (I2=86.1%, P=0.001). When deleting Du et al.’s study [27], the heterogeneity disappeared (I2=0%, P=0.640) and fixed model changed the results (WMD=−16.10U/L, 95% CI=[−22.51,−9.69], P<0.001). No publication bias was observed among these studies (Fig. A.3D–F).

Additionally, one study on the level of DBIL [27], one study on the proportion of increased DBIL [33], one study on the level of total protein [36], one study on the level of ALP [25], one study on the level of A/G [34], one study on the level of prealbumin [36] and two study on the proportion of increased GLB [24,33], were not included in our meta-analysis due to their inadequate data.