Copyright (C) 2021 ETH Zurich, Switzerland. SPDX-License-Identifier: Apache-2.0. See LICENSE file for details.

Authors: Thorir Mar Ingolfsson, Xiaying Wang, Michael Hersche, Alessio Burrello, Lukas Cavigelli, Luca Benini

This project provides the experimental environment used to produce the results reported in the paper ECG-TCN: Wearable Cardiac Arrhythmia Detection with a Temporal Convolutional Network available on arXiv. If you find this work useful in your research, please cite

@INPROCEEDINGS{9458520,
  author={Ingolfsson, Thorir Mar and Wang, Xiaying and Hersche, Michael and Burrello, Alessio and Cavigelli, Lukas and Benini, Luca},
  booktitle={2021 IEEE 3rd International Conference on Artificial Intelligence Circuits and Systems (AICAS)}, 
  title={ECG-TCN: Wearable Cardiac Arrhythmia Detection with a Temporal Convolutional Network}, 
  year={2021},
  volume={},
  number={},
  pages={1-4},
  doi={10.1109/AICAS51828.2021.9458520}}

• We developed and used the code behind ECG-TCN on Ubuntu 18.04.3 LTS (Bionic Beaver) (64bit).
• The code behind ECG-TCN is based on Python3, and Anaconda3 is required.
• We used NVidia GTX1080 Ti GPUs to accelerate the training of our models (driver version 396.44). In this case, CUDA and the cuDNN library are needed (we used CUDA 10.1).

Also the dataset ECG5000 needs to be downloaded and put into the /data folder. It is available on here

Navigate to ECG-TCN's main folder and create the environment using Anaconda:

$ conda env create -f ECG-TCN.yml -n ECG-TCN

We provide the code to quantize the ECG-TCN model in the file nemo_quantization.py there we train, quantize and deploy the network to be used by DORY. We also provide code to quantize the same model with Tensorflow in the file tensorflow_quantization.py there we train, quantize and deploy the network to be used either straight on an MCU using TFlite as a library or using X-CUBE-AI.

Under /utils you find the data loading and model making files. Please note that because of the stochastic nature of training with GPUs it's very hard to fix every random variable in the backend. Therefore to reproduce the same or similar models one might need to train a couple of times in order to get the same highly accurate models we present.

Please refer to the LICENSE file for the licensing of our code.