Case fatality rate in COVID-19: a systematic review and meta-analysis

Chanaka N Kahathuduwa; Chathurika S Dhanasekara; Shao-Hua Chin

Introduction pressure of oxygen < 300 mmHg, requiring hospitalization. Critical illness was defined as having 114 respiratory failure, circulatory shock, end-organ failure or any combination of the above 115 requiring intensive care. In addition, the following variables were extracted as potential 116 covariates of the above outcomes. Central tendency (i.e. mean or median) and dispersion (i.e. 117 SD, SE, 95%CI, IQR or range) of age were extracted. When not reported, study level means and 118 standard deviations for age were imputed from the available statistics (i.e. median, IQR or 119 range). 12 Proportions of the following variables within a study sample were extracted: age < 18 120 years, age > 60 years, female sex, diabetes mellitus, hypertension, heart disease, chronic liver 121 disease, chronic kidney disease, chronic obstructive pulmonary disease, malignancy, 122 immunosuppression (e.g. HIV), smoking and pregnancy. Proportions of patients with specific 123 presenting symptoms (i.e. fever, cough, sore throat, shortness of breath, headache, diarrhea), 124 asymptomatic cases, specific laboratory parameters (i.e. positive nucleic acid test for COVID-19, 125 leukopenia, leukocytosis, thrombocytopenia, lymphopenia, elevated lactate dehydrogenase 126 (LDH), elevated C-reactive protein (CRP), elevated erythrocyte sedimentation rate (ESR), high 127 procalcitonin and high D-dimer based on reference ranges considered in each study) and 128 radiographic features (i.e. no lesions on CT, patchy consolidation, ground glass opacities, 129 peripheral distribution, and bilateral lung involvement or involvement of > 3 lobes). 130 is also known to ensure analytic consistency and minimize bias. 14 Consistency of the findings of 137 the meta-analyses were confirmed by leave-one-out sensitivity analyses. 15 Given that under- 138 reporting and publication bias could result in biased (i.e. smaller) prevalence estimates, 139 publication bias was examined using funnel plots and the effect-sizes were imputed for estimated 140 missing (i.e. unpublished / unreported) studies via the trim-and-fill method. 16, 17 The meta-141 analyses were repeated including the effect-size estimates of these potentially missing studies in 142 order to obtain unbiased estimates of the three primary outcomes. Heterogeneity of effect-sizes 143 was quantified by calculating the Higgins' I 2 statistic for each meta-analysis. 18, 19 To explain the 144 heterogeneity of the studies 20 , exploratory univariate random-effects meta-regression analyses 145 were performed to examine the moderator effects of each of the covariates described above.