Study Design and Study Setting
This retrospective was conducted in the tertiary specialty referral hospitals known as King Fahad Medical City (KFMC; 1200 beds) and Prince Mohammed Bin Abdulaziz hospital (PMAH; 500 beds), both in Riyadh, Saudi Arabia. The Institutional Review Boards at KFMC and PMAH (IRB 20-200) approved the study. All methods were performed in accordance with the relevant guidelines and regulations. Institutional review boards of both KFMC and PMAH waived informed consent since it was an exempt study conducting a retrospective analysis.
Participant selection
Lists of discharged or dead COVID-19 patients were obtained from the health informatics officer and categorized into ACEI/ARB users and ACEI/ARB non-users based on their medication history. Inclusion criteria were age >18 years admitted at KFMC or PMAH from April to June 2020; being treated with ACE/ ARB inhibitors 6 months before and continued during hospital admission and after discharge for any indications such as hypertension, stroke, heart failure, myocardial infarction, diabetes mellitus, chronic kidney disease or nephrotic syndrome (criteria only for ACEI/ARB users); and SARS-CoV-2 infection confirmed by real-time polymerase chain reaction (RT-PCR) from nasopharyngeal swab for inclusion. Patients were excluded if pregnant, incomplete medical records due to patients being transferred recently from other hospitals, or unknown medications history.
Data Collection
Data were extracted manually from electronic health records (Corttex system and HIM system in KFMC, and Cerner Systems in PMAH) by a trained team and included demographic and anthropometric variables such as age, height, weight, body mass index (BMI), class of obesity; medical history of hypertension, diabetes, asthma, cardiovascular disease (including coronary artery disease, heart failure), renal disease (chronic kidney disease or nephrotic syndrome); COVID-19 treatment regimen; use of ACE/ARB; spectrum of illness severity; need for ICU admission; need for non-invasive ventilation (NIV) (e.g., face mask, nasal cannula, nasal mask, or helmet); need for invasive mechanical ventilation (MV); in hospital-death; and length of hospitalization.
COVID-19 Spectrum of Severity definitions
We defined the spectrum of disease severity according to the WHO [28]. Mild illness was defined as uncomplicated upper respiratory tract viral infection and may have non-specific symptoms such as fever, fatigue, cough, anorexia, malaise, muscle pain, sore throat, dyspnea, nasal congestion, or headache. Moderate illness was the development of non-severe pneumonia that does not require supplemental oxygen. Severe pneumonia was defined as fever plus symptoms ≥1 of the following: respiratory rate ≥30/min, dyspnea, respiratory distress, SpO2 ≤93% on room air, PaO2/FiO2 ratio <300 or lung infiltrate >50% of lung field within 24-48hr. Critical illness manifested by symptoms ≥1 of the following: acute respiratory distress syndrome (ARDS), septic shock, altered consciousness, or multi-organ Failure.
Study Outcomes
The primary outcome was COVID-19 severity classified as mild, moderate, severe, or critical according to the WHO classification in ACEI/ARB users as compared to non-ACEI/ARB users. Secondarily, we evaluated the need of ICU admission, NIV and MV, in-hospital death, and length of hospital stay in patients on ACEI/ARB as compared to non-ACEI/ARB users. We also assessed COVID-19 severity (severe or critical) in patients with (1) one of the three comorbidities of interest (diabetes, hypertension or renal disease); (2) both diabetes and hypertension; (3) diabetes only; and (4) hypertension only.
Statistical Analysis
Using R Core Team (2020) software (R Foundation for Statistical Computing, Version 4.0.1, Vienna, Austria), continuous data were expressed as mean with standard deviation (±SD) or median with interquartile range (IQR). If the normality assumption was met using normal Q-Q plot and the Shapiro–Wilk test, we used the Student’s t-test for group comparisons; if not met, we used the Mann-Whitney. Categorical data were reported as frequencies and percentages and analyzed either using the Chi-square test for nxm tables or Fisher's exact test for 2x2 tables group comparisons. To obtain odds ratios, we performed a multivariable logistic regression model adjusting for the confounders (either P<0.3 in a univariable logistic regression model or clinically important confounders) of age, sex, BMI, diabetes, hypertension, renal disease, number of comorbidities (diabetes, hypertension, cardiovascular disease [heart failure/coronary artery disease], stroke, renal disease, asthma and obesity) and COVID-19 treatment regimen. All statistical inferences were drawn with 95% confidence intervals with P<0.05. Bonferroni adjustments were applied to control for multiplicity [29].
Sensitivity analyses
Propensity score-matched analysis
Propensity score-matched (PSM) procedure was implemented to the dataset. Using MatchIt package in R [30], the treatment assignment was modelled with a multivariable logistic regression with the following specifications: nearest neighbor 1:3 with replacement and a caliper of 0.1. Multiple models were fitted and compared using the "cobalt" package to plot propensity score distributional mirror diagrams and standardized mean difference (SMD) before and after matching. The SMD check for data balance is an indicated statistical procedure to judge the quality of matched data rather than means or standard deviations of the matched subjects[31]. In our case we considered values of SMD <0.1 to be indicative of adequate balance and fruitful matching. We carefully examined models in which we evaluated different interaction terms. The final model was then selected based on the above criterion. The following covariates were included in final model: age, BMI, diabetes, hypertension, renal disease and number of comorbidities. We illustrated the propensity score distribution (prior matching was also evaluated using mirror diagram and the love plot for SMD distribution) (see Supplementary Fig. S1 and Fig. S2).
Inverse propensity score weighting (IPSW)
In addition to PSM, we considered a secondary sensitivity analysis for high risk patients (diabetes, hypertension or renal disease). We limited the analysis to the severity outcome to avoid a multiplicity problem. In this technique, we used "WeightIt" and "ipw" packages in R to calculate inverse propensity weights[32][33]. The following covariates were considered in the IPSW analysis: age, gender, diabetes, hypertension, renal disease, cardiovascular disease, body mass index, and asthma and number of comorbidities. For specifics, see Fig S3 on covariates selected and balance distribution. Again, various models were fitted and toughly tested before selecting the best model for each data subsets. Following the recommendation from Desai et al [34], the estimand was the average treatment effect. Stabilized weights were utilized in our analysis and then we examined weights with plots and summary statistics to prevent variance inflation from extreme weights. Love plots for SMD distribution presented in supplementary. Lastly, we fitted a structural causal model using "Survey" package to obtain robust variance estimation (sandwich type).