An SVM approach for identifying atrial fibrillation

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Automated analysis of electrocardiogram (ECG) signals is crucial for diagnosing various cardiac conditions.
• Distinguishing between normal sinus rhythm, atrial fibrillation (AF), noisy segments, and other arrhythmias presents a significant challenge in ECG interpretation.

Purpose of the Study:

• To develop and evaluate an automated algorithm for classifying short ECG strips into four distinct categories.
• To assess the algorithm's performance using a large, annotated dataset and validate it on an independent set of recordings.

Main Methods:

• An algorithm was designed to identify R peaks and other ECG waves, extracting 61 features including time-frequency domain characteristics, intra-beat interval variability, and beat morphology.
• A support vector machine (SVM) classifier, with and without a 2-layer neural network, was trained on these features using the PhysioNet Challenge 2017 dataset.
• The algorithm was trained on 8528 ECG recordings and validated on 3658 recordings, encompassing normal, AF, noisy, and other rhythm disturbances.

Main Results:

• The algorithm achieved a total F1 score of 0.80 on the hidden test dataset, ranking within the top 18-24 algorithms in the PhysioNet Challenge 2017.
• The addition of a neural network did not significantly improve the classification of 'other' rhythms misclassified as 'normal'.
• The algorithm demonstrated capability in classifying atrial fibrillation (AF) versus non-AF and normal versus abnormal (arrhythmia or noise) records.

Conclusions:

• The developed automated algorithm provides an effective method for classifying short ECG strips into key cardiac rhythm categories.
• The algorithm's performance is robust, as evidenced by its competitive ranking in a challenge dataset and its ability to differentiate between normal and abnormal rhythms, including AF.