Development of a Machine Learning Predictive Model for Stroke Among Patients with Non-Valvular Atrial Fibrillation Receiving Oral Anticoagulant Treatment

OBJECTIVES: Objective was to develop a Machine Learning (ML) predictive model for stroke among patients with NVAF receiving Oral Anticoagulants (OAC) that allows to predict the individual patient level risk of stroke given patient features and select the best OAC for a NVAF patients.

METHODS: Study database consisted of patients with NVAF, receiving OAC treatment with maximum gap between prescriptions of 60 days, and at least 1-year history medical records, extracted from Spanish IQVIA’s Longitudinal Patient Database. Follow-up period was set at 120, 180 and 365 days after first OAC prescription, giving 10,024, 8,028 and 4,628 patients in each subsample and a number of stroke events of 299 (3.0%), 256 (3.2%) and 156 (3.3%) respectively. Gradient boosted trees (Xgboost) followed by imbalance correction (weighting) and Platt’s calibration as well as ML uplift modelling (S-/T-/X-/R-learners) with and without Propensity Score Matching were used to build the model. Patient risk of stroke under current treatment was compared with optimal treatment as determined by the model and patient sub-populations for particular treatments were proposed.

RESULTS: ML model for 180 days (n=8,028 patients) showed an AUROC of 0.87 with a test recall on 50% precision of 0.79 and a test precision on 50% recall of 0.80. The same model for 365 days (n=4,628 patients) showed an AUROC of 0.81.

Important predictors found were previous stroke history, age, use of other antithrombotic agents, hypertension, BMI, use of lipid modifying agents, extrapyramidal disorders, gastritis and duodenitis, urine creatinine level, use of corticosteroids, use of drugs for peptic ulcer and gastro-oesophageal reflux disease and glucose level.

The model can identify patients that have been initiated on suboptimal OAC treatment: those whoser risk of stroke would be lower under different/optimal treatment.

CONCLUSIONS: The ML predictive model developed showed good performance and can define subgroups that might benefit from OAC switching.