Edge Devices for Detecting Atrial Fibrillation Using Deep Learning



Journal Title

Journal ISSN

Volume Title




Cardiac arrhythmia is a condition that causes irregular heartbeat rhythms. There are several arrhythmia types, but in this thesis, we are only discussing atrial fibrillation(AF), which is classified from other rhythms that are normal and noisy. AF is defined as a ”tachyarrhythmia characterized by predominantly uncoordinated atrial activation with consequent deterioration of atrial mechanical function” [1]. AF is one of the most common cardiac arrhythmias; more than 12 million people suffer in Europe and North America. It is likely from the statics that it might get tripled in the next 30-50 years. In general, AF is recorded using an electrocardiogram (ECG/EKG) machine. Readings are taken in two different ways, first using single lead and second using 12 lead. Leads are electrodes that are attached to the limbs to measure the heart rhythm. Single lead ECGs are most commercially available for the general public. Various 12 lead are highly expensive and are used in most sophisticated laboratories. In 12 lead ECG, six leads are connected to limbs(Hands and Legs) and called limb leads. The other six leads are connected to the torso, which is also called ”Precordial Leads.” ECG is a combination of intervals, segments, waves, and one complex. Once the machine records the cardiac rhythm, the doctor should manually check if there are any irregularities present, confirming any problem. So, this tedious task of looking at every report can be automated using deep neural network models. We are using Convolution Neural Networks(CNN) to solve this problem of AF classification from other rhythms. This thesis explores aspects of CNN’s hardware implementation on various edge devices and compares powers vs. performance aspects.



Atrial fibrillation, Neural circuitry, Neural networks (Neurobiology)