Atrial Fibrillation Detection Algorithm with Ratio Variation-Based Features

A two-layer analysis approach of the atrial fibrillation episode detection algorithm tested in the MIT-BIH atrial fibrillation database (MIT-BIH AFDB) is proposed in the paper. We use several methodologies, including gradient varying weighted filter, template matched filter, adaptive threshold, and sliding window to accurately extract the locations and amplitudes of P, Q, R, S, and T peaks, P wave width, and QS width in an ECG complex as basic features. On the other hand, most existing works utilize features of RR intervals, a difference of RR intervals, or amplitude of P wave for AF episode detection. In the proposed algorithm, we exploit the ratio concept to transform basic features into ratio-based features with relative relations because those features are much easier to measure the irregularity of RR intervals and P wave absence precisely in atrial fibrillation episodes. Furthermore, we apply the innovative definition of ratio variation-based features to generate robust and qualitative feature extraction sets. Finally, a rule-based ratio variation hypothesis classifier with techniques of weighted coefficient function, product-form score function, Gini index function, and Gini splitting function is adopted. The performance result of the proposed algorithm, trained and tested in the MIT-BIH AF database, achieves an average sensitivity value of 99.272% and an average specificity value of 98.495%, respectively. The accuracy is superior to that of other various AF episode detection algorithms.