(SLC-048) Association between COVID-19 and thrombotic events: using Super Learner algorithm combined with doubly robust methods.

Background: Previous observational studies examining thrombotic complications from COVID-19 used parametric statistical methods, such as propensity score modeled by logistic regressions. Machine learning methods do not require any modeling assumptions and can be used to estimate causal inference quantities more flexibly.

Objectives: To compare the risk of arterial thromboembolism (ATE) and venous thromboembolism (VTE) in members hospitalized with COVID-19 vs members hospitalized with influenza, using machine learning based algorithm (Super Learner) combined with doubly robust methods.

Methods: A retrospective study was conducted among members hospitalized with COVID-19 during 12/1/2020–12/31/2021 and members hospitalized with influenza during 10/1/2018-4/30/2019, using non administrative services only health claims. Outcomes of interest were ATE and VTE within 90 days of hospitalization. The predicted values of exposure (hospitalized COVID-19) and each outcome (ATE and VTE) were separately modeled on a total of 42 demographic and clinical covariates using Super Learner, an ensemble approach combining multiple prediction algorithms, including extreme gradient boosting, random tree, generalized linear models, and simple mean. To reduce the potential for overfitting, each prediction algorithm was fit using a 5-fold cross-validation. Two doubly robust methods for parameter estimation – targeted maximum likelihood estimation (TMLE) and augmented inverse probability weighting (AIPW) were used to avoid issues caused by high-dimensional data. The effect estimation for ATE and VTE was quantified by the risk difference and 95% confidence interval (CI) comparing COVID-19 cohort vs influenza cohort.

Results: A total of 25065 members with COVID-19 and 1908 members with influenza were included. The 90-day risk of ATE was 14.4% in influenza cohort and 15.8% in COVID-19 cohort. The 90-day risk of VTE was 6.0% in influenza cohort and 12.0% in COVID-19 cohort. Using Super Learner with double robust methods, members with COVID-19 had higher risk of 90-day VTE compared to members with influenza: the risk difference was 6.0% (95 CI: 4.8-7.3%) using TMLE method and 6.2% (95% CI 4.9-7.5%) using AIPW method. The risk of ATE was not significantly different between the two cohorts.

Conclusions: Our results from this demonstration study show that machine learning based algorithm combined with doubly robust methods can be used to estimate causal associations in observational data to account for a large number of covariates without any parametric modeling assumptions. This innovative methodology has the potential to advance real-world evidence studies.