EffSCG: An Efficient Framework for Real-Time Seismocardiogram Denoising on Resource-Constrained Edge Devices

Most deep learning models for wearable devices are still implemented on centralized servers due to memory and computational power limitations under the assumption that these servers will communicate with the wearable units. However, enabling complex physiological computing models to run directly on such edge units is beneficial, as it reduces latency and enhances data privacy. In this paper, we address this problem under the setting of a denoiser for a particular type of cardiovascular signal originating from the mechanical movements of the heart and blood - the seismocardiogram. Specifically, we demonstrate that by improvements in multiple aspects, the model can be made applicable to resource-constrained edge devices by improving the computational speed and the memory footprint without substantive accuracy reduction. Our optimization framework achieves a lower inference time and memory footprint by using 1. structured pruning of the weights of the neural network, 2. quantization of the weights to integer representations, 3. incorporation of a faster ordinary differential equation solver, and 4. application of progressive distillation. Compared to the baseline, our framework achieves a 47.2× speedup (97.9% latency reduction) and a 4.94× smaller memory footprint (79.8% memory reduction), while maintaining 92.88% of the original accuracy (7.12% deviation). In addition, we conduct an extensive analysis of multiple parameters to demonstrate their impact on computational speed, memory footprint, and accuracy. The presented framework will be applicable to models with diffusion-based architectures to enable their use on scarce-resourced edge devices to deploy these models in everyday life.