Registry features and data acquisition
We did a multinational registry analysis of the use of hydroxychloroquine or chloroquine with or without a macrolide for treatment of COVID-19. The registry comprised 671 hospitals located in six continents (
appendix p 3). The Surgical Outcomes Collaborative (Surgisphere Corporation, Chicago, IL, USA) consists of de-identified data obtained by automated data extraction from inpatient and outpatient electronic health records, supply chain databases, and financial records. The registry uses a cloud-based health-care data analytics platform that includes specific modules for data acquisition, data warehousing, data analytics, and data reporting. A manual data entry process is used for quality assurance and validation to ensure that key missing values are kept to a minimum. The Surgical Outcomes Collaborative (hereafter referred to as the Collaborative) ensures compliance with the US Food and Drug Administration (FDA) guidance on real-world evidence. Real-world data are collected through automated data transfers that capture 100% of the data from each health-care entity at regular, predetermined intervals, thus reducing the impact of selection bias and missing values, and ensuring that the data are current, reliable, and relevant. Verifiable source documentation for the elements include electronic inpatient and outpatient medical records and, in accordance with the FDA guidance on relevance of real-world data, data acquisition is performed through use of a standardised Health Level Seven-compliant data dictionary, with data collected on a prospective ongoing basis. The validation procedure for the registry refers to the standard operating procedures in place for each of the four ISO 9001:2015 and ISO 27001:2013 certified features of the registry: data acquisition, data warehousing, data analytics, and data reporting.
The standardised Health Level Seven-compliant data dictionary used by the Collaborative serves as the focal point for all data acquisition and warehousing. Once this data dictionary is harmonised with electronic health record data, data acquisition is completed using automated interfaces to expedite data transfer and improve data integrity. Collection of a 100% sample from each health-care entity is validated against financial records and external databases to minimise selection bias. To reduce the risk of inadvertent protected health information disclosures, all such information is stripped before storage in the cloud-based data warehouse. The Collaborative is intended to minimise the effects of information bias and selection bias by capturing all-comer data and consecutive patient enrolment by capturing 100% of the data within electronic systems, ensuring that the results remain generalisable to the larger population. The Collaborative is compliant with the US Agency for Healthcare Research and Quality guidelines for registries. With the onset of the COVID-19 crisis, this registry was used to collect data from hospitals in the USA (that are selected to match the epidemiological characteristics of the US population) and internationally, to achieve representation from diverse populations across six continents. Data have been collected from a variety of urban and rural hospitals, academic or community hospitals, and for-profit and non-profit hospitals. The data collection and analyses are deemed exempt from ethics review.
Study design
We included all patients hospitalised between Dec 20, 2019, and April 14, 2020, at hospitals participating in the registry and with PCR-confirmed COVID-19 infection, for whom a clinical outcome of either hospital discharge or death during hospitalisation was recorded. A positive laboratory finding for SARS-CoV-2 was defined as a positive result on high-throughput sequencing or reverse transcription-quantitative PCR assay of nasal or pharyngeal swab specimens, and this finding was used for classifying a patient as positive for COVID-19. COVID-19 was diagnosed, at each site, on the basis of WHO guidance.
11
Patients who did not have a record of testing in the database, or who had a negative test, were not included in the study. Only one positive test was necessary for the patient to be included in the analysis. Patients who received either hydroxychloroquine or a chloroquine analogue-based treatment (with or without a second-generation macrolide) were included in the treatment group. Patients who received treatment with these regimens starting more than 48 h after COVID-19 diagnosis were excluded. We also excluded data from patients for whom treatment was initiated while they were on mechanical ventilation or if they were receiving therapy with the antiviral remdesivir. These specific exclusion criteria were established to avoid enrolment of patients in whom the treatment might have started at non-uniform times during the course of their COVID-19 illness and to exclude individuals for whom the drug regimen might have been used during a critical phase of illness, which could skew the interpretation of the results. Thus, we defined four distinct treatment groups, in which all patients started therapy within 48 h of an established COVID-19 diagnosis: chloroquine alone, chloroquine with a macrolide, hydroxychloroquine alone, or hydroxychloroquine with a macrolide. All other included patients served as the control population.
Data collection
Patient demographics, including age, body-mass index (BMI), sex, race or ethnicity, and continent of origin were obtained. Underlying comorbidities (based on International Classification of Diseases, tenth revision, clinical modification codes) present in either the inpatient or outpatient electronic health record were collected, which included cardiovascular disease (including coronary artery disease, congestive heart failure, and history of a cardiac arrhythmia), current or previous history of smoking, history of hypertension, diabetes, hyperlipidaemia, or chronic obstructive pulmonary disease (COPD), and presence of an immunosuppressed condition (steroid use, pre-existing immunological condition, or current chemotherapy in individuals with cancer). We collected data on use of medications at baseline, including cardiac medications (angiotensin converting enzyme [ACE] inhibitors, angiotensin receptor blockers, and statins) or use of antiviral therapy other than the drug regimens being evaluated. The initiation of hydroxychloroquine or chloroquine during hospital admission was recorded, including the time of initiation. The use of second-generation macrolides, specifically azithromycin and clarithromycin, was similarly recorded. A quick sepsis-related organ failure assessment (qSOFA) was calculated for the start of therapy (including a scored calculation of the mental status, respiratory rate, and systolic blood pressure) and oxygen saturation (SPO2) on room air was recorded, as measures of disease severity.
Outcomes
The primary outcome of interest was the association between use of a treatment regimen containing chloroquine or hydroxychloroquine (with or without a second-generation macrolide) when initiated early after COVID-19 diagnosis with the endpoint of in-hospital mortality. The secondary outcome of interest was the association between these treatment regimens and the occurrence of clinically significant ventricular arrhythmias (defined as the first occurrence of a non-sustained [at least 6 sec] or sustained ventricular tachycardia or ventricular fibrillation) during hospitalisation. We also analysed the rates of progression to mechanical ventilation use and the total and intensive care unit lengths of stay (in days) for patients in each group.
Statistical analysis
For the primary analysis of in-hospital mortality, we controlled for confounding factors, including demographic variables, comorbidities, disease severity at presentation, and other medication use (cardiac medications and other antiviral therapies). Categorical variables are shown as frequencies and percentages, and continuous variables as means with SDs. Comparison of continuous data between groups was done using the unpaired
t-test and categorical data were compared using Fisher's exact test. A p value of less than 0·05 was considered significant. Multiple imputation for missing values was not possible because for disease and drug variables, there were no codes to indicate that data were missing; if the patient's electronic health record did not include information on a clinical characteristic, it was assumed that the characteristic was not present.
Cox proportional hazards regression analysis was done to evaluate the effect of age, sex, race or ethnicity (using white race as a reference group), comorbidities (BMI, presence of coronary artery disease, presence of congestive heart failure, history of cardiac arrhythmia, diabetes, or COPD, current smoker, history of hypertension, immunocompromised state, and history of hyperlipidaemia), medications (cardiac medications, antivirals, and the treatment regimens of interest), and severity of illness scores (qSOFA <1 and SPO2 <94%) on the risk of clinically significant ventricular arrhythmia (using the time from admission to first occurrence, or if the event did not occur, to the time of discharge) and mortality (using the time from admission to inpatient mortality or discharge). Age and BMI were treated as continuous variables and all other data were treated as categorical variables in the model. From the model, hazard ratios (HRs) with 95% CIs were estimated for included variables to determine their effect on the risk of in-hospital mortality (primary endpoint) or subsequent mechanical ventilation or death (composite endpoint). Independence of survival times (or time to first arrhythmia for the ventricular arrhythmia analysis) was confirmed. Proportionality between the predictors and the hazard was validated through an evaluation of Schoenfeld residuals, which found p>0·05 and thus confirmed proportionality.
To minimise the effect of confounding factors, a propensity score matching analysis was done individually for each of the four treatment groups compared with a control group that received no form of that therapy. For each treatment group, a separate matched control was identified using exact and propensity-score matched criteria with a calliper of 0·001. This method was used to provide a close approximation of demographics, comorbidities, disease severity, and baseline medications between patients. The propensity score was based on the following variables: age, BMI, gender, race or ethnicity, comorbidities, use of ACE inhibitors, use of statins, use of angiotensin receptor blockers, treatment with other antivirals, qSOFA score of less than 1, and SPO2 of less than 94% on room air. The patients were well matched, with standardised mean difference estimates of less than 10% for all matched parameters.
Additional analyses were done to examine the robustness of the estimates initially obtained. Individual analyses by continent of origin and sex-adjusted analyses using Cox proportional hazards models were performed. A tipping-point analysis (an analysis that shows the effect size and prevalence of an unmeasured confounder that could shift the upper boundary of the CI towards null) was also done. All statistical analyses were done with R version 3.6.3 and SPSS version 26.
Role of the funding source
The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author and co-author ANP had full access to all the data in the study and had final responsibility for the decision to submit for publication.