Cardiac arrhythmias classification using photoplethysmography database

Abstract

Worldwide, Cardiovascular Diseases (CVDs) are the leading cause of death. Patients at high cardiovascular risk require long-term follow-up for early CVDs detection. Generally, cardiac arrhythmia detection through the electrocardiogram (ECG) signal has been the basis of many studies. This technique does not provide sufficient information in addition to a high false alarm potential. In addition, the electrodes used to record the ECG signal are not suitable for long-term monitoring. Recently, the photoplethysmogram (PPG) signal has attracted great interest among scientists as it provides a non-invasive, inexpensive, and convenient source of information related to cardiac activity. In this paper, the PPG signal (online database Physio Net Challenge 2015) is used to classify different cardiac arrhythmias, namely, tachycardia, bradycardia, ventricular tachycardia, and ventricular flutter/fibrillation. The PPG signals are pre-processed and analyzed utilizing various signal-processing techniques to eliminate noise and artifacts, which forms a stage of signal preparation prior to the feature extraction process. A set of 41 PPG features is used for cardiac arrhythmias' classification through the application of four machine-learning techniques, namely, Decision Trees (DT), Support Vector Machines (SVM), K-Nearest Neighbors (KNNs), and Ensembles. Principal Component Analysis (PCA) technique is used for dimensionality reduction and feature extraction while preserving the most important information in the data. The results show a high-throughput evaluation with an accuracy of 98.4% for the KNN technique with a sensitivity of 98.3%, 95%, 96.8%, and 99.7% for bradycardia, tachycardia, ventricular flutter/fibrillation, and ventricular tachycardia, respectively. The outcomes of this work provide a tool to correlate the properties of the PPG signal with cardiac arrhythmias and thus the early diagnosis and treatment of CVDs.

Figure 1

An example of raw PPG signal.

Figure 2

A block diagram outlining the pre-processing procedure for PPG signals.

Figure 3

Example for (a) Valleys' detection, and (b) one segment of the PPG signal.

Figure 4

The principles of the morphological features of the PPG signal.

Figure 5

The classification accuracy for different techniques with different PCA components.

Figure 6

The accuracy of different classification techniques with the number of features.

Figure 7

The classification sensitivity of different cardiac arrhythmias; namely AF, Bradycardia, Normal, and Tachycardia by the utilization of different machine learning techniques (a) DT, (b) SVM, (c) KNN, and (d) Ensemble.