Motion-Resilient ECG Signal Reconstruction from a Wearable IMU through Attention Mechanism and Contrastive Learning

Abstract

Wearable electrocardiogram (ECG) sensors can detect dogs’ heartbeat signals and have proven useful in monitoring dogs’ welfare and predicting temperament scores in structured evaluations of potential guide dog puppies. Despite advances in the ergonomics, performance, and usability of ECG sensor technologies specifically designed for dogs, deploying those systems in the real world imposes challenges such as training human operators to ensure electrodes’ proper contact with the skin and, especially in the case of puppies, socialization to achieve comfort and reduce behavioral inhibition. Seismocardiogram signal is an alternate modality for heartbeat signals and is acquired using the Inertial Measurement Unit (IMU), which is commercially available, widely deployed, and does not require skin-contact. However, the extracted signals from IMU are subject to heavy influences from motion and other noise sources. In this paper, we present a method that enables extracting the similar physiological parameters ECG provides using easier-to-deploy IMU sensors. We propose and evaluate a machine learning framework that reconstructs ECG signals from IMU signals even under moderate to heavy movements. Our study investigated two artificial neural network architectures to overcome severe noise artifacts in the IMU signal resulting from dogs’ movements and environmental factors. The first architecture combines the attention mechanism and convolution layers to extract important features from the temporal IMU input. The second architecture adapts contrastive representation learning to the regression problem and learns a more effective embedding for the ECG reconstruction. The qualitative inspection and quantitative analysis based on F1 scores of the R-peak alignment demonstrate the effectiveness of the two proposed models in removing motion noises and reconstructing realistic ECG signals, achieving an F1 score of 0.72 in the best case compared to 0.29 from the baseline.