Objectives: The main objective of this research is to assess the feasibility of training a machine learning model which would able to detect atrial brillation with high accuracy, precision, sensitivity, and specicity. Methodology: Deep neural networks were utilized to create high dimen- sional and non-linear decision boundaries from large datasets contain- ing atrial brillation signals. Shannon's entropy measure was used to quantify the randomness of the RR intervals. Ensembles of the best- performing models aided the construction of even better-performing mod- els. To construct the training dataset, only the electrocardiogram seg- ments that contain only one rhythm episode were selected, and oversam- pled appropriately. Data: Well-known ECG benchmark databases from the PhysioNet web- site were selected as the data source from which 30-second segments were generated. The atrial brillation databases were used for training, while the arrhythmia database was used for testing the models. Conclusion: The models trained with our approach succeeded in detect- ing atrial brillation with high accuracy, precision, and sensitivity, using only the heartbeat types and RR intervals. The best stand-alone neu- ral networks for classication achieved F1-scores of up to 86.3%. The two-stage ensemble models which we developed, that rst predict if the rhythm is normal and then recognize atrial brillation from the rest of the signal achieved an F1-score of up to 89.1% on the MITDB database.
ecg · atrial brillation · neural networks · rr intervals