Cleaning ECG with Deep Learning: A Denoiser Tested in Industrial Settings

As the popularity of wearables continues to scale, a substantial portion of the population has now access to (self-)monitorization of cardiovascular activity. In particular, the use of ECG wearables is growing in the realm of occupational health assessment, but one common issue that is encountered is the presence of noise which hinders the reliability of the acquired data. In this work, we propose an ECG denoiser based on bidirectional Gated Recurrent Units (biGRU). This model was trained on noisy ECG samples that were created by adding noise from the MIT-BIH Noise Stress Test database to ECG samples from the PTB-XL database. The model was initially trained and tested on data corrupted with the three most common sources of noise: electrode motion artifacts, muscle activation and baseline wander. After training, the model was able to fully reconstruct previously unseen signals, achieving Root-Mean-Square Error values between 0.041 and 0.023. For further testing the model’s robustness, we performed a data collection in an industrial work setting and employed our model to clean the noisy data, acquired from 43 workers using wearable sensors. The trained network proved to be very effective in removing real ECG noise, outperforming the available open-source solutions, while having a much smaller complexity compared to state-of-the-art Deep Learning approaches.