University of Twente Student Theses

Login

Compressed Convolutional Neural Networks for the Classification of Electrocardiogram Signals on Embedded Devices

Birton, Laurentiu (2021) Compressed Convolutional Neural Networks for the Classification of Electrocardiogram Signals on Embedded Devices.

[img] PDF
378kB
Abstract:To determine abnormal conditions of the cardiac muscles, medics usually rely upon electrocardiography, a process that maps the electrical activity of the heart onto an electrocardiogram. Classic methods used to read and classify ECG's can be sensitive to the difference between patients or require deep specific knowledge and can be time-consuming. As an alternative, Machine learning is often used nowadays to automatize the classification of ECG signals, mostly for the detection of cardiovascular diseases. However, this comes with a large computational cost, usually requiring large and powerful machines to perform the classification task. This can be very limiting, especially in situations where high mobility or low costs are desired. Since research into the application of efficient machine learning methods for use in low powered devices has greatly increased in the last years, here we show an investigation into the compression of Convolutional Neural Networks for use in embedded devices for the classification of ECG signals. We describe the data we used, the employed compression techniques, what type of models we have analysed and finally show the results of our experiments together with possible ways of improving the performance of these models in the future. By the end of this paper, we'll have shown how a CNN can be compressed to sizes of down to under 1 Megabyte and still obtain very good accuracy in heartbeat classification.
Item Type:Essay (Bachelor)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science BSc (56964)
Link to this item:http://purl.utwente.nl/essays/87323
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page