Identifying Normal, AF and other Abnormal ECG Rhythms using a Cascaded Binary Classifier

Abstract

In this paper, we present a methodology for classifying normal, atrial fibrillation (AF), non-AF related other abnormal heart rhythms and noisy recordings by analysing single lead ECG signal of short duration. In a two layer binary cascaded approach proposed in our methodology, an unlabelled recording is initially classified into one of the two intermediate classes ('normal+others' and 'AF+noisy') at the first layer before actual classification at the second layer. The Physionet Challenge 2017 dataset containing more than 8500 ECG recordings are used for creation of training models and interval validation. The proposed methodology yields an average F1-score of 0.91, 0.79 and 0.77 respectively in classifying normal, AF and other rhythms on the training dataset using 5-fold cross validation. Results also show that, the said methodology, when applied on a hidden test set maintained by the challenge organisers yields F1-score values of 0.92, 0.86 and 0.74 in classifying the same. IntroductionAtrial Fibrillation (AF) is a common type of heart disease that leads to stroke, heart failure or other complications. Millions of people get affected by AF every year and the prevalence of the disease is likely to increase. Noninvasive detection of AF is a popular area of research for quite a long time. Irregularities in heart beat is considered to be the most common symptom of AF and can be traced in an ECG. However, being an episodic event an accurate detection of AF is not always trivial. Conventional AF detectors ([1], [2], [3]) are mostly of atrial activity analysisbased or ventricular response analysis-based methods. The absence of P waves or the presence of f waves in the TQ interval are searched in atrial activity analysis-based AF detectors. On the other hand, time, frequency and morphological features are extracted from RR intervals to identity the heart beat irregularity in ventricular response analysisbased methods.However, the prior art methods have certain limitations regarding real time deployments. 1) Most of them are validated on clinically accepted 12 lead ECG signals, recorded for a relatively longer duration. 2) Algorithms are mostly applied on carefully selected clean data. However, in practical scenario, ECG signals are often noisy in nature. 3) Size of the test dataset are often not adequate for making a conclusion. 4) Most prior arts perform binary classification between AF and normal recordings only. However, there are many non-AF abnormal rhythms (like tachycardia, bradycardia, arrhythmia etc) which exhibits heart beat pattern similar to AF. Considering them in the in the dataset makes the classification task more challenging. In this paper we propose a robust algorithm for classifying normal, AF, other abnormal rhythms and noisy ECG recordings. The diverse ECG dataset, provided in Physionet challenge 2017 [4] is used for internal performance evaluation and creating the training models. Information regarding individual recordings are not available regarding the other rhythms in the dataset as...

A Wearable Electrocardiogram Telemonitoring System for Atrial Fibrillation Detection

“…In comparison with our previous method, the CatBoost model proposed in this work achieved an overall F 1 score of 0.91 on AFDB-2017, which increased by 0.04, while using less features. It is also worth noting that the algorithms of Datta et al [21], Zabihi et al [25] and Hong et al [22] achieved high performance on the training set but much lower performance on the test set. The reason may be the overfitting problem was not prevented when training the models.…”

“…AF detection methods using ECG signals are usually investigated from two aspects, the absence of P waves or the presence of f-waves [13,14], and irregularity of RR intervals [15][16][17][18][19]. In the 2017 PhysioNet/Computing in Cardiology (CinC) Challenge of AF classification (termed Challenge) [20], the official algorithms [21][22][23][24][25][26][27] based on machine learning, especially deep learning methods, have achieved excellent performance on AF detection.…”

Strong deep learning baseline for single lead ECG processing

“…On the other hand, real-life samples often pose too much noise and high variance that could mislead handcrafted rules, and yet stateof-the-art approaches are still relying heavily on feature engineering for AF detection. Accordingly, three of the four winners of the CinC Cardiology 2017 challenge combined only medically relevant feature-extractors and did not incorporate any neural network-based features [7,8,9].…”

A Robust Detection Method of Atrial Fibrillation

“…Considering that RR interval irregularity is an important indicator of AF detection [10], we developed three types of AF prior features related to RR interval and deviation of RR interval (dRR), including: 1) sample entropy, coefficient and density histogram of sample entropy, and threshold of median absolute deviation of RR interval; 2) standard deviation, ratio of standard deviation to mean, density, approximate entropy, sample entropy and coefficient of sample entropy of dRR; 3) interval, duration, amplitude, position, slope and area of the 10s slice, the average RR interval within the 10s slice, the number of samples with a difference in the slice exceeding 0.1mV within the 10s slice, and the normalized power spectral density of the 10s slice; 4) normalized power spectrum at intervals of 0-0.05 Hz, 0.05-0.15 Hz, and 0.15-0.5 Hz of RR .…”