Development of an artificial intelligence model to predict venous thromboembolic disease in COVID-19 patients at hospital admission

Background: A lot of attention has been drawn to the identification of predictors of VTE in COVID-19 patients, and an accurate clinical prediction model is still lacking in this context. Aims: To develop a clinical prediction model using artificial intelligence techniques, to predict VTE in COVID-19 patients, using variables easily available upon hospital admission. Methods: This multicenter cohort included consecutive adult patients (≥ 18 years-old) with laboratory-confirmed COVID-19 from 37 Brazilian hospitals from 17 cities, between March and September 2020. Study data were collected from medical records using Research Electronic Data Capture (REDCap) tools. We trained multiple machine learning models on various combinations of structured and non-structured features, calibrated to reflect a probability distribution while predicting the desired clinical outcome. Subsequently, we analyzed the relationship between this model ' s predicted confidence score and the fraction of false negatives in the test sample to devise a splitting point where no false negatives would occur, thus calibrating for sensitivity over specificity. The study was approved by the National Research Ethics Commission waiving off the application of informed consent. Results: The dataset included 6421 patients (median age 61 [P25-75 48-73] years-old, 54.8% men), 4.5% of them developed venous thromboembolic disease. Patient ' s age, sex and comorbidities, as well as their list of household prescription drugs, history of recent surgery and laboratory tests were significant predictors. Given a proper confidence level, our model predicted 100% of the true positive cases while eliminating a significant portion of the true negatives (Figure 1). (Figure Presented) Conclusions: This study suggests that an ensemble of decision rules can effectively predict COVID patients with high risk of VTE. It might be possible to decrease the use of anticoagulants while still treating patients with an appreciable likelihood of thromboembolism.