Predicting the Risk of Ischemic Stroke in Patients Treated with Novel Oral Anticoagulants: A Machine Learning Approach

Abstract

Introduction: The aim of this cohort study was to estimate the predictors of ischemic stroke in patients treated with non-vitamin K antagonist oral anticoagulants (NOACs) in a large database containing data from general practitioners in Germany using machine learning methods. Methods: This retrospective cohort study included 39,652 patients with a diagnosis of atrial fibrillation (AF) and an initial prescription of NOAC in 1,278 general practices in Germany between January 2011 and December 2018. Of 39,652 patients, 2,310 (5.8%) receive the first stroke or TIA diagnosis during the follow-up time (average follow-up time 2.5 [SD: 1.8] years). Sub-Population Optimization and Modeling Solutions (SOMS) tool was used to identify subgroups at a higher risk of stroke compared to the overall population receiving NOAC based on 37 different variables. Results: Using SOMS, a total of 9 variables were considered important for the stroke prediction. Age had 59.1% of prediction importance, following by ischemic heart diseases (10.6%), urinary tract infections (4.6%), dementia (3.5%), and male sex (3.5%). Further variables with less importance were dizziness (2.2%), dorsalgia (1.5%), shoulder lesions (1.1%), and diabetes mellitus (1.1%). Discussion/Conclusions: The stroke risk in AF patients treated with NOAC could be predicted based on comorbidities like ischemic heart diseases, urinary tract infections, and dementia additionally to age and male sex. Knowing and addressing these factors may help reduce the risk of stroke in this patient population.