Biomedical Introduction

Electrocardiograms (EKGs) are tests that measure the electrical conductivity through the heart and are a common method of identifying abnormal heart rhythms (arrhythmias). Atrial fibrillation (AF) is one of the most common types of arrhythmia characterized by rapid and irregular beating of the atria. AF affects about 3% of the population of North America and Europe and resulted in about 200,000 deaths in 2015. The lack of symptoms and access of many underdeveloped communities to cardiologists contributes to the high number of undiagnosed cases. Our objective is to develop a model that can classify AF or any other type of arrhythmia from a normal rhythm using EKGs.

A normal EKG beat usually consists of three segments: the P wave, QRS complex and T wave. The P wave is the first bump of the beat representing atrial depolarization. The normal amplitude is usually between 0.05 to 0.25mV while the duration is normally 0.06 to 0.11 seconds. The QRS complex is a large upward deflection following the P wave that represents ventricular depolarization. The interval usually ranges from 0.08 to 0.10 seconds. Finally the T wave is a bump following the QRS complex and represents ventricular repolarization. It is characterized as being small, upward and round. A normal ECG typically consists of 12 different leads. Leads I, II and III are known as the bipolar limb leads, leads aVR, aVL and aVF are the augmented voltage leads and leads V1-V6 are the periodical leads.

Atrial fibrillation (AF) is the most common type of sustained arrhythmia. Incidents of AF have been increasing over the past few years. AF has been associated with heart failure, dementia and strokes. On an EKG, AF is classified as being an irregularly irregular rhythm meaning that the variation of the interval between each beat is random with no identifiable pattern. AF also lacks the presence of a P wave, instead having many fibrillatory waves and an irregular R-R interval. There are four other types of irregularly irregular rhythms which are premature atrial complexes (PACs), premature ventricular complexes (PVCs), wandering atrial pacemaker (WAP) and multifocal atrial tachycardia (MAT). PACs and PVCs are known as ectopic beats meaning they do not follow a rhythm but instead can appear at any given time. PACs are characterized by an abnormal P wave morphology and a premature heartbeat arising from an abnormal initiation in the atria. PVCs are characterized by an abnormal QRS complex as well as a premature heartbeat arising from an initiation in the Purkinje fibers. WAP is classified as having three different P wave morphologies arising from three independent nodes of depolarization in the atria. MAT constitutes the same features as WAP but also has a rapid heart rate of more than 100 bpm. There are many other arrhythmias which can mainly be identified from abnormalities of the placement, duration, amplitude or morphology of one or more of the three segments constituting an EKG.

Our training data was obtained from the Physionet website, which contains multiple datasets of medical tests. In particular we used the dataset of 8528 EKGs, part of a challenge to classify AF known as "Physionet Challenge 2017". The EKGs were measured using the AliveCor device which consists of two electrodes on a mobile case. The device produces a lead I or lead II EKG based on placement of the electrodes. The resulting EKGs were labeled by medical professionals as one of four categories: normal, AF, other, noisy. Normal meant a normal hearth rhythm, AF meant atrial fibrillation, other meant any other type of arrhythmia and noisy meant the signal was too noisy to classify. We developed a multinomial logistic regression model that could classify a lead I or lead II EKG as normal, AF, other or noisy using the Physionet Challenge 2017 data as our learning data.