A Support Vector Machine approach for reliable detection of atrial fibrillation events

There is a need for accurate and reliable detectors of asymptomatic atrial fibrillation (AF). Several ECG-based algorithms have been described in the literature, but no open comparison of features on out of-sample data has been published. Therefore, ten R-peak related features were selected from detectors available in literature and their classification performances were assessed both univariately and when combined using a Support Vector Machine (SVM). The MIT-BIH AFDB was used as the training set, and the MIT-BIH Normal Sinus Rhythm Database (NSRDB) and the MIT-BIH Arrhythmia Database were used for out-of-sample test performance assessment. During the training phase, the optimal number of beats for accurate detection was determined using cross validation. The SVMs hyper-parameters were optimized with a grid search. On the training set the SVM had a Sensitivity (Se) of 99.07% and a Positive Predictive Value (PPV) of 98.27%. During independent testing on the MIT-BIH NSRDB the SVM had a Sp=99.72% which was superior to any single feature or previous detector. The SVM also provided a Sp=99.70% on series 100 of the MIT-BIH Arrhythmia Database and a Sensitivity of 100% on series 200 of the same datase. A good Specificity (82.00%) and Accuracy (85.45%) were also obtained. Results are superior to any previously reported, for both training and testing and robust across multiple databases.