Integration of Results From Convolutional Neural Network in a Support Vector Machine for the Detection of Atrial Fibrillation

Atrial fibrillation (AF) can cause a variety of heart diseases and its detection is insufficient in outside hospital. We proposed three methods for AF diagnosis in ambulatory settings. The first method is a convolutional neural network (CNN) trained on modified frequency slice wavelet transform (MFSWT) data. The second is a support vector machine (SVM) classifier trained on multiple AF features data. The third method is an SVM trained on the same feature set but extended by the predictive probability of the CNN. The proposed method (the third one) achieved the highest detection accuracy. MIT-BIH AF database was used as a training set with an accuracy of 97.87% for 30-s ECG episodes and 96.09% for 10-s ECG episodes from fivefold cross-validation. The trained model was tested on the PhysioNet/Computing in Cardiology (CinC) Challenge 2017 database, achieving an accuracy of 93.21% for 30-s episodes and 93.03% for 10-s ECG episodes. When tested on the China Physiological Signal Challenge (CPSC) 2018 database, the corresponding accuracies were 98.48% and 98.61%. The results on the wearable ECGs from a clinical AF patient were 99.21% and 97.04%. We retrained the model on the PhysioNet/CinC Challenge 2017 data set and tested on the other database to explore the generalization ability of the proposed method. Corresponding test results on the MIT-BIH AF database showed accuracies of 96.84% and 95.13%, on the CPSC 2018 database were 96.21% and 98.45%, on the wearable ECGs were 99.08% and 96.43%. The results proved that the proposed method could provide high accuracy and reliable recognition for AF events.