Embedded 1D Convolutional Network based ECG Classification Platform for Remote Health Monitoring

International audience Remote Health (RH) offers a promising solution to broaden healthcare access, particularly in rural areas. In the heart of this RH paradigm are various types of systems, where a particular focus is on easy-to-use and low-cost embedded systems. Clinical examinations start often with heart rate monitoring and analysing of cardiac electrical activity or electrocardiogram (ECG). Recently, 1D counterparts of classic 2D convolutional neural networks (CNNs) have been introduced in the ECG processing domain, allowing to detect and classify abnormal heart rhythms with a high accuracy and precision. In this work, we present an optimized 1D temporal convolutional network (TCN) and its hardware microcontroller based implementation for ECG classification. To the best of our knowledge, the proposed embedded architecture is the lightweight 1D TCN architecture found in the literature giving the best accuracy of 97.2 % and F1-Score of 96.4 % on the ECG5000 dataset. Moreover, the obtained implementation results on microcontrollers outperform the similar works while keeping the best classification accuracy.