Multi-source features and support vector machine for heart rhythm classification

The PhysioNet/CinC Challenge 2017 provides over 8,528 short single channel ECG recordings for the classification of rhythms as normal sinus rhythm, AF, other rhythm, or too noisy. We present a support vector machine (SVM)-based heart rhythm classifier that leverages features based on rhythm, morphology and arrhythmia characteristics of the ECG. Four benchmark signal quality features, eleven rhythm features, fifteen morphology features, six arrhythmia features and four novel abnormality indices, were presented to the libSVM software package with 10 fold cross-validation. The SVM parameters were tuned using the modified cuckoo search (MCS) algorithm during the training. Since two of the authors were involved in the challenge, this is an unofficial entry. Without the MCS parameter optimization, the results for the mean F1 measures from the 10 fold cross validation on the training set were 0.911, 0.826, 0.728 and 0.852 for normal sinus rhythm, AF rhythm, other rhythm and noise respectively, resulting in a F 1 score of 0.822. With MCS parameter optimization, the mean F 1 measures on the training set increased to 0.921, 0.835, 0.739 and 0.870 respectively, resulting in a F 1 score of 0.832. The final F 1 measures on the test set were 0.914, 0.805, 0.691 and 0.737 for normal sinus rhythm, AF rhythm, other rhythm and noise respectively, resulting in a final F1 score of 0.803.