IntroductionAtrial Fibrillation (AF) is one of the most common cardiac arrythmia. It occurs in 1-2% of the general population, and this number will likely triple in the next 30-50 years [1]. AF can be easily mistaken with other arrythmias, or omitted, because of its episodic occurance.Throughout the years, considerable progress has been made in the automatic detection of AF. However, current methods are not promising. Algorithms that can be found in literature are usually tested on clean data, not properly seperated from the training set, based on small amount of patients.Physionet Challenge 2017 [2] has given an opportunity for scientific community to improve AF detection, by publishing dataset of short one-lead records, containing of more than 8528 training examples. Such dataset can satisfy previous limitations. Presented work is an attempt of creating reliable, patient-independent, resistant for other arrythmias system. Related work introduced various algorithms for predicting disease and detecting different types of arrhythmia. AF symptoms and though analysis could be basically divided into two categories: based on atrial activity or ventricular response. AF detectors that could combine both features could provide an enhanced performance. Published methods include approaches based on machine learning [3,4].The paper is organized as follows: Section 2 describes feature extraction, Section 3 explains briefly the theory behind the clasifiers, Sections 4 is the explenation of trainig approach, cross-validation techniques and the configurations of proposed models. Section 5 presents the results and concludes the paper.
Features extractionOne of the most important components of proposed solution are features designed by the author. Overall, 36 features were obtained. They can be split in 3 categories: the inter-beat timing ('RR intervals') features, statistical features, frequency features, morphological features, noise features. The number of designed features in each category is respectively 8, 3, 5, 4, 16, 2.
Preprocessing and beat detectionTo remove baseline wandering and high-frequency noise, Butterworth 3rd-order filtering was performed, with bandpass frequencies between 1 and 25 Hz. The frequencies were chosen based on later cross-validation.After frequencies removal, it was necessary to detect R-peaks, in order to calculate ventricular response features. Algorithm described in [5] was used. It consists of novel nonlinear transformation of ECG signal, based on Shannon energy tresholding, and peak-finding strategy, based on the first-order Gaussian differentiator. On a popular benchmark, MIT-BIH arrhythmia database, it achieves an average sensitivity of 99.94% and a positive predictivity of 99.96%, which is a competetive score.