Atrial Fibrillation Detection by Means of Edge Computing on Wearable Device: A Feasibility Assessment

Cardio Vascular Diseases (CVDs) represent one of the main burden that affected world population in the last and in the current decades. The early detection by means of wide screening population-wide may represent a good path to avoid the worsening of pre-existent situation. In this arena, the use of wearable devices in combination with deep learning to deliver edge computing system seems to be the most viable pathway to follow in order to fight the CVDs burden. Despite the fact that many studies have concentrated on edge computing techniques for CVDs, there is a limited literature on Atrial Fibrillation (AF) detection directly on-device. Due to limited availability of research on this topic, the feasibility assessment of an on-device edge computing wearable system is described in this work. Starting with an examination of the features to be considered, the study progresses through the building of a Neural Network (NN), the training of the model, and the on-cloud testing process to completion. The NN is composed of 4 hidden layer made up of respectively 5, 30, 20 and 10 node. The learning rate is 0.005 and the number of training cycle is 30. The training set consists of 3362 windows, and the testing set consists of 796 windows. The findings of the test are encouraging, with an output F1-score of 0.94 for AF recognition as a result of the test. The model is then deployed on-device and evaluated offline, without the need for any additional devices or an internet connection, in order to run the inference process. Finally, the system that will be used for future human trials is presented, together with a description of the factors that led to the selection of this particular system and the major characteristic of the sensors.