Ethiopia is the Federal Democratic Republic with nine regional states (Afar, Amhara, Benishangul-Gumuz, Gambella, Harari, Oromia, Somali, Southern Nations Nationalities and People's Region, and Tigray) and two city administrations (Addis Ababa and Dire Dawa). The country has a total area of 1,100,000 km² and is divided into zones, which are further subdivided into districts. Districts are further subdivided into kebeles, the lowest administrative divisions (13). Ethiopia, with a population of ~112 million people, is Africa's second-most populous country (56,010,000 female and 56,069,000 male individuals in 2019) (14).

This review was reported using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guideline (15) (Supplementary material). The articles about our topic of interest were searched systematically from international and national electronic databases such as PubMed/MEDLINE, CINAHL, EMBASE, African Journals Online, Google Scholar, and ScienceDirect. In addition to the databases, articles were identified by searching the reference lists of eligible studies. Two authors (MYB and SSJ) searched independently. Endnote X9 was used to retrieve and manage studies searched out from different databases. An example of a search engine approach applied in the PubMed database searches is shown below: “(((((((COVID-19 infected [Text Word]) OR (COVID-19 patients [Text Word])) OR (COVID-19 patients [MeSH Terms])) AND (rate [Text Word])) OR (incidence (Text word)] OR (predictors (Text word)] OR (patient with COVID-19) [MeSH Terms])) AND (Ethiopia [MeSH Terms])) OR (Ethiopia [Text Word]).” The search strategy was performed from 10 to 30 March 2023.

The eligibility criteria were developed using the logic grid structure considering “Context, Condition, and Population (COCOPop).”

All cohort studies conducted on “incidence and predictors of COVID-19 mortality in Ethiopia” and published in English were included.

The titles of the articles that were similar to this study but different in the interest of outcomes (not reporting incidence and predictors) were excluded.

Two authors (MYB and SSJ) screened all titles/abstracts of the searched articles independently. To assure the transparency and validity of the included articles, the screened results of the two authors were cross-checked, and the presence of conflicting results was discussed and fixed based on scientific merit. The authors who worked on the title and abstract screening also worked on the full-length screening independently and came to a consensus on the included articles (Figure 1).

The Newcastle-Ottawa Quality Assessment Scale for cohort study (16) was used to evaluate the quality of the included studies. This guideline contains selection (representativeness of the exposed cohort, selection of the non-exposed cohort, ascertainment of exposure, and demonstration that an outcome of interest was not present at the start of the study), comparability (based on the design or analysis controlled for confounders), and outcome (assessment of outcome, was follow-up long enough for outcomes to occur, and adequacy of follow-up of cohorts). The assessment tool guides that the articles that scored 3 or 4 stars in the selection domain AND 1 or 2 stars in the comparability domain AND 2 or 3 stars in the outcome/exposure domain were declared as “good quality,” 2 stars in selection domain AND 1 or 2 stars in comparability domain AND 2 or 3 stars in outcome/exposure domain were declared as “fair quality,” and 0 or 1 star in selection domain OR 0 stars in comparability domain OR 0 or 1 stars in outcome/exposure domain were declared as “poor quality” as a result, during the quality assessment of articles, all included articles were declared as having good quality.

The data were extracted using the data extraction checklist prepared from the Microsoft Excel spreadsheet and adapted from the JBI data extraction spreadsheet. To ensure consistency, two authors (MYB and SSJ) extracted the data independently and cross-checked the quality and content of the data. The difference in the data extraction between the authors was solved through discussion. The extraction checklist has two sections, depending on the objectives. The data extracted for the first objective were region, publication year, sample size, follow-up period, first author name, sampling methods, follow-up period, incidence, and events during extraction. The data extraction checklist for predictors contains the hazard ratio, lower and upper confidence intervals, exposed and cases, exposed and non-cases, non-exposed, and non-cases of the predictors.

In this study, two outcomes were measured. The primary outcome of interest was the pooled mortality rate of COVID-19 patients in Ethiopia. It was calculated by considering the incidence of the individual studies and the standard error of the incidence of COVID-19 mortality in Ethiopia. The second outcome of interest was the pooled predictors of COVID-19 mortality in Ethiopia, which were calculated by taking the logarithm of the relative risk and the standard error of the logarithm of the relative risk using a binary meta-regression model and reporting the relative risk with 95% confidence interval (CI).

The extracted data were exported to the Stata^TM version 17.0 software from a Microsoft Excel spreadsheet. The pooled mortality rate of COVID-19 was estimated using the Metaprop Stata command. The standard errors were calculated from the reported estimates and population denominators using a binomial distribution assumption. The presence of heterogeneity among studies was assessed using Cochran's Q test and quantified using the I² statistics. Heterogeneity is classified as low (I² values: 0–25), moderate (I² values: 25–50), high (I² values: 50–75), and very high (I² values: ≥75) (17). To try to identify the source of heterogeneity, computing subgroup analysis, meta-regression, publication bias, and sensitivity analysis were used. To estimate the effect sizes, a random-effects model based on DerSimonian and Laird methods was used (18). The scatterplot in the funnel plot was assessed as the initial publishing bias evaluation, followed by Egger's linear regression test, which demonstrated that small studies had no effect on publication bias, and finally, trim and fill analysis tests were computed to obtain a conclusion (19). Finally, the findings were presented using texts, tables, and figures such as forest plots.

A total of 630 studies were searched from electronic databases such as PubMed/MEDLINE, CINAHL, EMBASE, Google Scholar, and ScienceDirect, as well as organizational records, references, and websites. Approximately 560 articles were excluded due to duplication, 60 articles were excluded due to differences in study setting/context (17, 19–23), and 10 articles were excluded due to differences in outcome of interest (20, 24–26). Finally, seven retrospective cohort studies were included in this systematic review and meta-analysis.

In Ethiopia, approximately seven studies with 31,498 children were included and followed. These studies were conducted in the Amhara region (n = 3) and the Oromia region (n = 4). All of the included studies were conducted through a retrospective follow-up method. Studies with the smallest (288) and largest (28,533) sample sizes were conducted in the Amhara region. The included studies' follow-up periods ranged from 3 to 22 days, with 3,127 to 243,887 person-days of COVID-19 mortality-free observations (Table 1).

The mortality rate of COVID-19 inpatients was 9.13 (95% CI: 5.38, 12.88) per 1,000 person-days of COVID-19 mortality-free observations. When we looked at it by region, the Oromia region had the highest COVID-19 mortality rate (ES: 11.48 (95% CI: 7.31, 15.64), and the Amhara region had the lowest (ES: 6.12 (95% CI: −0.89, 13.14).

To identify the source of heterogeneity, the subgroup analysis was conducted using region, mean sample size, and mean follow-up period. Hence, the subgroup analysis for sample size was computed using a mean sample size of 494, and heterogeneity was not explained by the sample size (<494) I² = 94.8 with a p-value of 0.000, and those conducted using a sample size greater than the mean (>494 (I²: 95.7 with a p-value of 0.00) (Figure 2). The mean follow-up period was 7 days, and heterogeneity was not explained by the follow-up period (7 days) I² = 96.2 with a p-value of 0.000, and studies conducted greater than the mean follow-up days (>7 days) I² = 97.6% with a p-value of 0.000 (Figure 3). A subgroup analysis by region was performed, but heterogeneity was not explained by it in the Amhara region (I² = 99.0% with a p-value of 0.000) or the Oromia region (I² = 89.5% with a p-value of 0.000) (Figure 4). Finally, the sources of heterogeneity were not identified through the subgroup meta-analyses.

In random-effects meta-regression, the year of publication and sample size were used as covariates. According to the findings, sample size (p = 0.77) and publication year (p = 0.43) did not affect heterogeneity (Table 2).

As the number of included articles is < 10, the Egger linear regression test was computed to objectively examine the presence of publication bias. As a result, there was no statistically significant publication bias (p = 0.000) (Table 3). As indicated in Table 3, the observed heterogeneity was not due to publication bias.

To identify the predictors of COVID-19 mortality, different predictors were considered, among which chronic diseases (e.g., chronic renal disease, diabetes mellitus, and hypertension), HIV infection, and smoking were identified as statistically significant predictors of COVID-19 mortality in Ethiopian COVID-19 patients.

Those study participants who had chronic renal disease had 2.29 (95% CI: 1.14, 4.60) times higher chance of mortality when compared with those who were free of chronic kidney disease (Figure 5). COVID-19 patients who had diabetes mellitus had 2.14 (95% CI: 1.22, 3.76) times higher chance of experiencing mortality than patients who were free of diabetes mellitus (Figure 6). The chance of experiencing mortality among patients with HIV infection was 2.98 (95% CI: 1.26, 7.03) times higher than their corresponding counterparts (Figure 7). The risk of COVID-19 mortality among patients with hypertension was 1.63 (95% CI: 1.43, 1.85) times higher than their counterparts (Figure 8). Those patients who were smokers had 2.35 (95% CI: 1.48, 3.73) times higher than their counterparts (Figure 9).