Prediction of incident atrial fibrillation in post-stroke patients using machine learning: a French nationwide study

Background: Targeting ischemic strokes patients at risk of incident atrial fibrillation (AF) for prolonged cardiac monitoring and oral anticoagulation remains a challenge. Clinical risk scores have been developed to predict post-stroke AF with suboptimal performances. Machine learning (ML) models are developing in the field of AF prediction and may be used to discriminate post-stroke patients at risk of new onset AF. This study aimed to evaluate ML models for the prediction of AF and to compare predictive ability to usual clinical scores.

Methods: Based on a French nationwide cohort of 240,459 ischemic stroke patients without AF at baseline from 2009 to 2012, ML models were trained on a train set and the best model was selected to be evaluate on the test set. Discrimination of the best model was evaluated using the C index. We finally compared our best model with previously described clinical scores.

Results: During a mean follow-up of 7.9 ± 11.5 months, 14,095 patients (mean age 77.6 ± 10.6; 50.3% female) developed incident AF. After training, the best ML model selected was a deep neural network with a C index of 0.77 (95% CI 0.76-0.78) on the test set. Compared to traditional clinical scores, the selected model was statistically significantly superior to the CHA₂DS₂-VASc score, Framingham risk score, HAVOC score and C₂HEST score (P < 0.0001). The ability to predict AF was improved as shown by net reclassification index increase (P < 0.0001) and decision curve analysis.

Conclusions: ML algorithms predict incident AF post-stroke with a better ability than previously developed clinical scores. AF: atrial fibrillation; DNN: deep neural network; IS: ischemic stroke; KNN: K-nearest neighbors; LR: logistic regression; RFC: random forest classifier; XGBoost: extreme gradient boosting.