Associations Between COVID-19 ARDS Treatment, Clinical Trajectories and Liberation From Mechanical Ventilator - an Analysis of the NorthCARDS Dataset

Overview

The mortality rates associated with COVID-19 related ARDS (COVIDARDS) have varied from observational reports from around the world. This has ranged from 44% (28 day mortality) in the UK to 36% (28 day mortality from ICU admission) in Italian studies, to 32% (all-cause 28 day mortality) in Spain. Predictive models have identified risk factors for COVID-19 hospitalized patients' mortality to include male sex, obesity, age, obesity, comorbidities including chronic lung disease and hypertension, as well as biomarkers including high levels of D-Dimer, LDH and CRP. In addition, practice patterns, such as drugs that were administered, timing of mechanical ventilation and adherence to established lung protective ventilation protocols are known to be variable across sites and have changed over time. The investigators propose to analyze outcomes for patients with COVIDARDS within the NorthCARDS dataset (a dataset of over 1500 patients with COVID-19 related ARDS across the Northwell Health System in the NYC metropolitan region and Long Island, NY) to understand differences in hospital survival and in the time to liberation from mechanical ventilation, specifically looking at the associations between baseline patient factors, changes in biomarkers, respiratory function and hemodynamics over time, and treatments administered. The analyses will be based on three hypotheses: H.1. Worsening trajectories of: oxygenation index (OI), respiratory system compliance (C), and inflammatory markers will be associated with lower hospital survival. H.2. Higher duration of deep sedation and paralytics will be associated with greater time to liberation from mechanical ventilation. This risk will be increased in patients with worsening trajectories of OI, C, and inflammatory markers over time. H.3. Type of mechanical ventilator, specifically the time on portable mechanical ventilator, is associated with hospital mortality and with inability to liberate from mechanical ventilator despite controlling for risk factors of changes in OI, C and Inflammatory markers over time, and the use of paralytics and deep sedation.

The investigators will leverage the NorthCARDS dataset for this analysis. This dataset includes over 1500 persons admitted to the Northwell Health System who had PCR positive COVID19 testing and were invasively mechanically ventilated for ARDS. Registry development was initiated in April 2020 and continues with prospective data collection for all mechanically ventilated COVID19 ARDS patients among the Northwell Health hospitals. Data structuring and engineering is informed by weekly multi-disciplinary review including frontline clinicians, data scientists, biostatisticians and data engineers within medical informatics. Random selection of patients for individual 'manual' chart review occurs for data assumptions and recording. The two outcomes to be modeled using multivariable regression analyses will be: 1. Index hospital survival and 2. Time to liberation from mechanical ventilation. Liberation from mechanical ventilation will be defined as non-palliative extubation and persistent extubation for greater than one week. Outcomes will be obtained from electronic health record queries. Patients in whom the investigators do not have outcomes data by November 30,2020 will be censored in analyses, and descriptive statistics will be summarized and presented separately. The investigators will approach this analysis using both hypothesis-driven methods wherein known risk factors for poor outcomes will be included in the multivariable regression models (logistic regression for Model 1, and Cox Proportional Hazards for Model 2); and investigators will also perform data-driven variable selection for the models. A priori defined risk factors that will be included in the models will be: Age, Gender, BMI, functional status at baseline (nursing home versus community admission), Comorbidities (coronary artery disease, Chronic Kidney Disease, Neurologic disorders, COPD, Diabetes, Active cancer, Hypertension); Inpatient treatments (for continuous values will be (max, median, trajectory)) including PEEP levels, Driving Pressure, FiO2, hypoxemia (Pao2:Fio2), type of mechanical ventilator (portable versus not), COVID-targeted medications (e.g., azithromycin, hydroxychloroquine, corticosteroids); and end-organ damage in-hospital: liver dysfunction, Kidney dysfunction, coagulopathy, (captured via SOFA scores), cardiac dysfunction, and shock requiring vasopressor/inotrope. Calendar-time, hospital type (community versus tertiary hospital) and hospital capacity (measured as number of hospital beds filled and time from admission order in ER to being transferred to an inpatient bed) will also be included in the analyses to account for temporal and systemic influences of outcomes. The final models will include variables selected through a backward selection process, together with variables ranked highly through data-driven methods including a logistic regression model regularized by Lasso penalty and Cox Proportional Hazards Model regularized by Lasso penalty. Model performance will be assessed for Model 1 (hospital survival) using the C-statistic. Model performance for Model 2 (time to mechanical ventilator liberation) will be based on the C-statistic adapted for censored data. Missing Data management: When the outcome data is missing for Model (1) (hospital survival), if there is less than 5% of outcomes missing, complete case analysis will be used; if there is more than 5% missing, sensitivity analysis will be performed by assuming all the missing outcomes to be either expired or alive to see if the results are similar to those using complete case analysis. When the outcome data is missing for Model (2) (liberation from mechanical ventilation) the missing outcome will be considered as censored. If overall \< 5% of our cohort has missing data for any risk factors, only patients with complete values for all risk factors will be included (others will be discarded). If \> 5% of the cohort is missing data for any risk factor, the missing data will be imputed using multiple imputation. If a risk factor is missing in \> 50% of patients the variable will not be included in the analysis. Feature Engineering/Data Reduction: We will also test whether combinations of covariables considered as one covariable increases model performance. This will include COVID-19 illness index (combination of hyperinflammatory markers, PaO2:FiO2 index at the time of intubation, requiring vasopressors at the time of intubation, and Oxygenation Index) and adherence to standard ARDS treatment protocols (Driving Pressure, whether receiving less than 6-8 cc/kg predicted body weight and whether or not proned).