Individualized Treatment Effects of Steroids in Veterans Hospitalized With COVID-19

BACKGROUND Treatment with corticosteroids is recommended for hospitalized patients with COVID-19 based on randomized controlled trials. These trials showed heterogeneity of treatment effect (HTE) for dexamethasone, with patients requiring supplemental oxygen benefitting more than those on room air. Recently, causal machine learning has successfully identified HTE for various critical care interventions and can provide individualized predictions regarding the degree of benefit or harm from an intervention. RESEARCH QUESTION We hypothesize causal machine learning can identify HTE with corticosteroid treatment and provide more accurate, personalized predictions for a national cohort of Veterans admitted with COVID-19. STUDY DESIGN AND METHODS Veterans admitted to wards or step-down units in Veterans Health Administration hospitals between March 2020 and March 2022 with a positive COVID-19 lab test or a strong positive billing diagnosis were eligible. A causal forest model predicted the individualized treatment effect of receiving ≥6 mg dexamethasone or equivalent versus no steroid administration in the first 24 hours on in-hospital mortality. Demographic variables, the worst value in the first 24 hours for vital signs, laboratory results, and nursing documentation were included. The model was trained in 70% of hospitals, and performance was assessed using the Qini coefficient in the remaining 30%. RESULTS Included Veterans were admitted to 114 hospitals, totaling 58,535 encounters. Model predictions ranged from an 8.5% reduction to a 7.8% increase in in-hospital mortality when given steroids. The Qini coefficient indicated a statistically significant reduction of in-hospital mortality if the individualized treatment effect (ITE) predictions were used compared to random treatment (P=0.003). In contrast, treatment allocation using FiO2 resulted in a non-significant Qini coefficient (P=0.206). INTERPRETATION A causal forest model accurately predicted the degree of benefit or harm from corticosteroids in Veterans hospitalized with COVID-19. Further research is needed to determine if individualized treatment recommendations could improve outcomes beyond usual care.