Allocation of scarce resources during the COVID-19 pandemic

Resources are insufficient to treat patients affected by COVID-19, particularly those who face imminent death without immediate treatment with mechanical ventilation. Medical ethicist Ezekiel Emanuel states that “we cannot leave these decisions to the frontline clinicians fighting the virus, forcing them to make well-intentioned, but ad hoc choices under extreme pressure”. Likewise, Lisa Rosenbaum states in a New England Journal of Medicine commentary that “the first and most important [priority] is to separate clinicians providing care from those making triage decisions”. Thus, data-based triage rules are required. \n \n Ventilator triage during the COVID-19 pandemic, like any medical policy decision, is a fundamentally causal problem. Any triage rule must be recommended on the basis of predictions of counterfactual population outcomes (e.g. all cause mortality) that would be observed were the rule actually implemented compared to if alternative rules were implemented. Machine learning algorithms that simply predict outcomes under current practice, even with unlimited data, are not sufficient. The determination of optimal strategies for the allocation of scarce resources would ideally be evaluated in randomized controlled trials (RCTs), comparing the relevant outcomes (e.g., all-cause mortality) from different allocation algorithms under real-world constraints. However, in this time of crisis, conducting such RCTs would be neither timely nor feasible. Herein, we propose to apply causal inference techniques to observational data from intensive care units (ICUs) that have provided care to patients with COVID-19 to emulate the RCTs that would ideally be performed.