P34 Identifying Interventions That Reduced COVID-19 Mortality in Long-Term Care Facilities: A Causal Inference Analysis

Abstract

ObjectivesBecause of a lack of experimental evidence to treat COVID-19, we applied causal inference (CI) analysis to longitudinal health record data of 4,091 long-term care high-risk patients with COVID-19 to ascertain those interventions that directly improved health outcomes.MethodsCOVID-19 patient data collected from January through October of 2020. Directed Acyclic Graphs (DAGs) were built to model assumed treatment cause and effect relationships and to eliminate bias from latent variables. Orthogonal Random Forests were used to generate individual heterogeneous treatment cause and effect models of all concurrent pharmacotherapy and a propensity score was calculated for each. Treatment-specific logistic-regression models were used to determine average treatment effects across the entire patient population. Mortality within 120 days of a Covid -19 diagnosis was the primary endpoint. Analysis was conducted using Python 3.9.0 in conjunction with the EconML and Scikit-Learn packages.ResultsPatients with higher body temperatures, oxygen saturation below 90% during their COVID-19 infection, and low platelet counts were at a significantly increased risk of death within 120 days. We determined that the average treatment effect across all COVID-19 positive residents on mortality risk was caused directly by omeprazole, enoxaparin cholecalciferol, apixaban, prednisone, sennosides and guaifenesin by showing counterfactually that mortality risk increased without those medicines.ConclusionsCausal inference combines graphical models of causal relationships with the calculation of counterfactuals (“what-if” questions) to identify actionable causal factors that lead to improved health. Additionally, as demonstrated, interventional models derived from causal inference can be used to extrapolate “findings across domains (i.e., settings, populations, environments) that differ both in their distributions and in their inherent causal characteristics.” Future application of causal inference should demonstrate that such models can dynamically update treatment to optimize health as patient personal and clinical conditions change. ObjectivesBecause of a lack of experimental evidence to treat COVID-19, we applied causal inference (CI) analysis to longitudinal health record data of 4,091 long-term care high-risk patients with COVID-19 to ascertain those interventions that directly improved health outcomes. Because of a lack of experimental evidence to treat COVID-19, we applied causal inference (CI) analysis to longitudinal health record data of 4,091 long-term care high-risk patients with COVID-19 to ascertain those interventions that directly improved health outcomes. MethodsCOVID-19 patient data collected from January through October of 2020. Directed Acyclic Graphs (DAGs) were built to model assumed treatment cause and effect relationships and to eliminate bias from latent variables. Orthogonal Random Forests were used to generate individual heterogeneous treatment cause and effect models of all concurrent pharmacotherapy and a propensity score was calculated for each. Treatment-specific logistic-regression models were used to determine average treatment effects across the entire patient population. Mortality within 120 days of a Covid -19 diagnosis was the primary endpoint. Analysis was conducted using Python 3.9.0 in conjunction with the EconML and Scikit-Learn packages. COVID-19 patient data collected from January through October of 2020. Directed Acyclic Graphs (DAGs) were built to model assumed treatment cause and effect relationships and to eliminate bias from latent variables. Orthogonal Random Forests were used to generate individual heterogeneous treatment cause and effect models of all concurrent pharmacotherapy and a propensity score was calculated for each. Treatment-specific logistic-regression models were used to determine average treatment effects across the entire patient population. Mortality within 120 days of a Covid -19 diagnosis was the primary endpoint. Analysis was conducted using Python 3.9.0 in conjunction with the EconML and Scikit-Learn packages. ResultsPatients with higher body temperatures, oxygen saturation below 90% during their COVID-19 infection, and low platelet counts were at a significantly increased risk of death within 120 days. We determined that the average treatment effect across all COVID-19 positive residents on mortality risk was caused directly by omeprazole, enoxaparin cholecalciferol, apixaban, prednisone, sennosides and guaifenesin by showing counterfactually that mortality risk increased without those medicines. Patients with higher body temperatures, oxygen saturation below 90% during their COVID-19 infection, and low platelet counts were at a significantly increased risk of death within 120 days. We determined that the average treatment effect across all COVID-19 positive residents on mortality risk was caused directly by omeprazole, enoxaparin cholecalciferol, apixaban, prednisone, sennosides and guaifenesin by showing counterfactually that mortality risk increased without those medicines. ConclusionsCausal inference combines graphical models of causal relationships with the calculation of counterfactuals (“what-if” questions) to identify actionable causal factors that lead to improved health. Additionally, as demonstrated, interventional models derived from causal inference can be used to extrapolate “findings across domains (i.e., settings, populations, environments) that differ both in their distributions and in their inherent causal characteristics.” Future application of causal inference should demonstrate that such models can dynamically update treatment to optimize health as patient personal and clinical conditions change. Causal inference combines graphical models of causal relationships with the calculation of counterfactuals (“what-if” questions) to identify actionable causal factors that lead to improved health. Additionally, as demonstrated, interventional models derived from causal inference can be used to extrapolate “findings across domains (i.e., settings, populations, environments) that differ both in their distributions and in their inherent causal characteristics.” Future application of causal inference should demonstrate that such models can dynamically update treatment to optimize health as patient personal and clinical conditions change.