The study was registered with PROSPERO (CRD42020224255) and conducted in adherence with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) Statement [20].
Search strategy, information sources, and study selection
Two authors (AS and SA) independently searched the publicly available COVID-19 living systematic review [21]. It is updated daily to provide a dynamic database of research papers related to COVID-19 that are indexed by PubMed, EMBASE, MedRxiv, and BioRxiv. Studies were extracted between December 1st 2019 and February 28th 2021, using the search terms “frail”, and “frailty” within the title and the abstract columns of the systematic review list. Due to the rapidly evolving pandemic, pre-print studies that were yet to be peer-reviewed were included to capture as much data as possible. These terms were combined with the Boolean operator “OR”.
Eligibility Criteria
The corresponding authors of eligible studies [22–38] (Supplementary Table 1) were invited to participate and share their original individual patient data. We included studies that reported on adults aged ≥ 18 years with laboratory-confirmed symptomatic COVID-19 patients, a documented CFS score and admitted to ICU. Only the patients with hospital outcome were included in the final analysis.
Data extraction
Data collection included patient demographics (age, sex, comorbidities, ethnicity, ICU admission source, smoking status), CFS score, ICU organ supports, such as the need for mechanical ventilation (MV), non-invasive ventilation, vasopressors, and/or renal replacement therapy); medical treatment limitation order, ICU and hospital mortality, and ICU and hospital length of stay (LOS).
Explanatory variable - frailty
In the Canadian Study of Health and Aging, CFS based on a 9-point judgement-based categorical scale was used for frailty measurement [39]. This scale has demonstrated validity and reliability in frailty assessment in ICU patients and other populations [8, 39]. This scale includes CFS = 1 (very fit), 2 (well), 3 (managing well), 4 (vulnerable), 5 (mildly frail), 6 (moderately frail), 7 (severely frail), 8 (very severely frail), and 9 (terminally ill) [39]. The modified eight-category CFS is the most utilised frailty assessment in the critically ill [40]. Frailty scores were also dichotomised as non-frail (CFS = 1–4) or frail (CFS = 5–8) according to accepted definitions [8], with the frail cohort further considered in terms of mild/moderate frailty (CFS 5–6) and severe/very severe frailty (CFS 7–8).
Ethical Issues
The individual patient data meta-analysis was exempt from ethics approval because we obtained de-identified data from previously published and ethically approved individual studies.
Other covariates
Exposure variables such as age, sex, chronic respiratory disease, chronic kidney disease, ischemic heart disease, admission source, SOFA and APACHE 2 scores were investigated as risk factors for hospital mortality in patients with COVID-19.
Main outcome(s): This was a one-stage individual patient data meta-analysis to assess continuous covariates (CFS and age). The primary aim was to evaluate whether frailty scores predicted outcomes for patients with COVID-19 admitted to ICU, including ICU mortality, hospital mortality, and discharge destination after adjusting for age and gender. The primary outcome was hospital mortality. We examined the following secondary outcomes: organ support within the ICU (mechanical and non-invasive ventilation, renal replacement therapy, vasoactive infusion, and extracorporeal membrane oxygenation), length of ICU and hospital stay, ICU bed-days, and discharge destination.
Missing data
There was minimal missing data for the primary outcome (0.2%). However, there were missing data with illness severity scores (42.9%), comorbidities (11.4%), presenting symptoms (8.2%), biochemistry within 24 hours of ICU admission (10.5%), use of non-invasive ventilation (37.8%), use of invasive mechanical ventilation (IMV, 33.2%), hospital length of stay (0.7%) and discharge destination (31.3%). We did not perform any imputation.
Statistical Analysis
In this retrospective study, data were initially checked for completeness and validity with queries directed back to the contributing institutions. Normality was assessed in continuous data by employing both normal quantile (probit) plots and the Shapiro-Wilk test. Normally distributed data were reported using the mean (standard deviation [SD]). Non-normal, categorical, and dichotomous data were reported using either the median (interquartile range [IQR]) or number (frequency [%]) respectively. The final dataset was gathered from seven discrete institutions and all analyses were clustered by the institution. An initial analysis was conducted between survivors and non-survivors to identify the independent predictors of mortality in critically ill patients with COVID-19. When specifically analysing from a clinical frailty point of view, the binary CFS categories non-frail/frail, data were compared using either a standard t-test for normally distributed data, the Wilcoxon rank-sum test for categorical data or the Fisher’s exact test for dichotomous data. Comparisons were conducted on demographic, co-morbidity, symptomatology, illness severity score, available within the first 24 hours. ICU and hospital mortality were examined using a logistic model in the primary analysis. Our primary analysis included the CFS scale (1–8) with subsequent adjustment for age and sex. Secondary multivariable logistic models were constructed to examine the effects of age, CFS, obesity measured as body mass index less or greater than 30 kg.m− 2, presence or absence of co-morbidities including active cancer, dementia, and the neutrophil-lymphocyte ratio as a marker of chronic inflammation on ICU mortality with the predictors selected from the results of the univariate analysis described above. Two-way interactions were also tested between the CFS and the significant predictors. Sequential deletion of non-significant predictors with potential misspecification tested using the linktest was conducted at each step of model development. Post-estimation checks for model specification and presence of collinearity were conducted using the link test and variance inflation factor respectively. Results were reported as the odds ratio of death with its 95% confidence interval and p-value. A competing risk analysis was performed next to examine the marginal probability of death using both the presence of ventilation and CFS. The method of Fine and Gray was used to generate the cumulative incidence function [41]. The significant predictors from the logistic model were used and results expressed as sub-hazard ratios and their 95% confidence intervals. Youden’s Y statistic was calculated for each CFS score thus yielding individual sensitivity and specificity results. All analyses were conducted using STATA™ (version 16.0) with the level of significance set at α < 0.05.