Materials and Methods
Yale University's Human Subject Protection Program initially approved this retrospective cohort study and waived informed consent on May 6, 2020 (institutional review board protocol no. 2000028070). The reporting of this study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.
This study was based on all patients diagnosed with COVID-19 secondary to a SARS-CoV-2 infection and treated for COVID-19–related complications in the ICUs of six hospitals affiliated with Yale New Haven Health System (i.e., Yale New Haven Hospital, Saint Raphael Campus, Greenwich Hospital, Bridgeport Hospital, Lawrence + Memorial Hospital, and Westerly Hospital). The study period was from February 13, 2020, when the first COVID-19 patient was admitted to the ICU at Yale, to March 4, 2021, when the COVID-19 ICU admission significantly declined. We included all COVID-19 patients admitted to Yale's ICUs during the study period to reflect the experience of treating critically ill COVID-19 patients throughout the first pandemic year.
Inclusion criteria for this study included an age of 18 yr or older, a diagnosis of COVID-19 (based on real-time reverse transcription–polymerase chain reaction assay targeting three regions of the SARS-CoV-2 genome, namely orf1ab, spike [S] gene, and nucleocapsid [N] gene), and treatment in one of Yale New Haven Health System's ICUs at any time during the study period. COVID-19 patients who required organ support therapies or intensive monitoring and care were eligible for ICU admission at Yale. No patients were admitted to ICU purely for isolation. The respiratory criteria for ICU admission varied over time: when there were sufficient ICU resources, patients requiring noninvasive ventilation or invasive mechanical ventilation were admitted to the ICU; however, during the case surge, when ICU resources were inadequate, only patients requiring invasive mechanical ventilation were admitted to the ICU. Exclusion criteria for this study were death within 24 h after ICU admission, age of less than 18 yr, and continued hospitalization on the last day of data censoring. Patient care was per the institutional protocols customized for COVID-19 patients and continuously updated based on the evolving evidence.
The primary outcome was in-hospital mortality, defined as all-cause death that occurred during a patient's hospitalization. Patients were regarded as survivors if they were discharged alive from the hospital or as nonsurvivors if they died during hospitalization. We included patients who were admitted to Yale's ICUs up to March 4, 2021. The relevant information of those patients who remained hospitalized on March 4, 2021, was updated based on the electronic medical records on June 1, 2021 (i.e., the last day of data censoring).
The treatments in this study were any COVID-19–related pharmacologic or organ support intervention instituted during a patient's hospitalization. The pharmacologic treatments included (1) antiviral drugs (e.g., remdesivir and hydroxychloroquine); (2) anticoagulants (e.g., enoxaparin, heparin, and apixaban); (3) antiplatelet agents (e.g., aspirin, clopidogrel, and ticagrelor); (4) steroids (e.g., dexamethasone, methylprednisolone, and hydrocortisone); (5) immunomodulators (e.g., tocilizumab); (6) immunosuppressants (e.g., tacrolimus); (7) vasopressors (e.g., norepinephrine, epinephrine, and dopamine); and (8) uncategorized drugs (e.g., azithromycin, convalescent plasma, and famotidine). Information on drug dose, timing, and duration of treatment was collected. The organ support therapies included (1) conventional oxygen therapy delivered using a regular nasal cannula or face mask; (2) high-flow nasal cannula; 3) bilevel positive airway pressure ventilation; (4) continuous positive airway pressure ventilation; (5) invasive mechanical ventilation; (6) continuous venovenous hemofiltration; and (7) extracorporeal membrane oxygenation.
The potential confounders were as follows: (1) the known risk factors for COVID-19 mortality (age, sex, and hypertension); (2) the severity of the acute illness during the first 24 h after ICU admission (Sequential Organ Failure Assessment score, Glasgow Coma Scale score, and invasive mechanical ventilation); (3) the various phases during the first pandemic year, i.e., the first phase (February 1, 2020, to May 31, 2020), the second phase (June 1, 2020, to August 31, 2020), the third phase (September 1, 2020, to November 30, 2020), and the fourth phase (December 1, 2020, to March 4, 2021), with each patient assigned to a phase based on their ICU admission date; (4) the demographics and comorbidities; and (5) the laboratory results and vital signs during the first 24 h after ICU admission.
Data Sources and Measurement
The measurements of all variables of interest were conducted in routine patient care guided by the institutional protocols customized for COVID-19 patients and continuously updated based on the evolving evidence. Patient data were extracted from the electronic medical records by the Joint Data Analytics Team at the Yale Center for Clinical Investigation. This team centralizes and coordinates clinical and research analytics and reporting across the Yale New Haven Health System and Yale School of Medicine.
Efforts were made to minimize selection bias. Our study analyzed all adult COVID-19 patients admitted to the ICUs in six hospitals affiliated with Yale New Haven Health System at any time during the study period. Yale New Haven Health System covers a significant portion of Connecticut and provides a mixture of different levels of care to state residents. As all our patients were treated in hospital settings, missing data were minimized because of standardized electronic methods for data capture and recording. All variables of interest were measured using the same methods across the healthcare system.
No statistical power calculation was conducted before the study because we planned to include all COVID-19 patients who had been treated in Yale New Haven Health System's ICUs throughout the entire first pandemic year. The sample size was based on the available cases.
We used original quantitative data collected from electronic medical records, including demographic characteristics, laboratory results, vital signs, drug doses, and treatment timing and duration. We removed data outside of the 0.5 to 99.5 percentile range for vital signs, considering that some of these measurements could be artifacts or outliers.
Continuous data are presented as means and SD or median and interquartile range, depending on the normality of distribution, assessed using histograms and Q-Q plots. Categorical data are presented as numbers and percentages. Missing data were not imputed.
Our objective was to identify treatments associated with lower mortality using a multivariable Cox proportional-hazards model. The variables entering the multivariable analysis included all COVID-19–related treatments and the potential confounders described above under "Variables." Only those treatments that were used in at least 5% of patients were included in the analysis. Demographics, comorbidities, laboratory results, and vital signs with a P value less than 0.25 in univariate analyses were included in the multivariable analysis. If two variables had an absolute Pearson's or Spearman's rank correlation coefficient greater than 0.5, we included only one variable to avoid collinearity. We excluded variables that had missing data for more than 10% of the patients. Multiple testing correction was performed using the Bonferroni method to reduce the chance of type I errors at the two-sided 0.05 α level. The hypotheses for all COVID-19–related treatments were considered as a family; therefore, the raw P value for each treatment was multiplied by the number of treatments being analyzed to derive the corrected P value. The association was estimated using hazard ratios and reported with 95% CIs. To account for clustering within hospitals, we used robust sandwich estimators to compute standard errors for the hazard ratios.
We used propensity score–matching analysis to evaluate the reproducibility of the association identified by the multivariable analysis. We divided patients into two cohorts: one cohort received the treatment, and the other cohort did not, with these two cohorts balanced at the baseline level using propensity score matching. The propensity score model included the demographic characteristics, comorbidities, pandemic phase, severity of acute illness (during the first 24 h after ICU admission), laboratory results (during the first 24 h after ICU admission), and vital signs (during the first 24 h after ICU admission). The matched pairs were identified using a one-to-one nearest neighbor caliper of 0 to 0.1 width. The balance between matched pairs was assessed using a standardized 10% difference. Survival was estimated using the product-limit Kaplan–Meier estimator, and the log-rank statistic was used to compare the survival curves. A stratified Cox proportional-hazards model was used in the analysis of the matched pairs.
We additionally explored the factors that could have modified the association identified by the multivariable analysis and evaluated by the propensity score–matching analysis. The method of analysis depended on the characteristics of the treatment associated with lower COVID-19 mortality. If a drug was associated with lowering mortality significantly, we presented the relevant data by dividing the patients into subgroups with different drug doses when feasible. When feasible, we also split the matched pairs derived from the propensity score matching into subgroups with different drug doses to explore the potential factors that might have modified the association.
A data analysis and statistical plan was written and filed with a private entity (institutional review board) before the data were accessed. During the peer-review process, significant modifications were requested and implemented. No minimum clinically meaningful effect size was defined before data access. The propensity score-matched analyses were planned post hoc. For a two-tailed hypothesis test, the significance level for each general hypothesis was 0.05. All analyses were performed in R software (version 3.5.3, R Foundation for Statistical Computing, Austria), with packages including sqldf, dplyr, sandwich, survival, survminer, arsenal, mltools, MatchIt, stddiff, and tableone.
Anesthesiology. 2021;135(6):1076-1090. © 2021 American Society of Anesthesiologists | Lippincott Williams & Wilkins