Reducing Energy Consumption Using Bidirectional Long Short Term Memory-ECG Compression

This paper presents a neural network-based processor that enhances computation efficiency for achieving multi-class pulse recognition in wearable devices. Using a combination of convolutional neural networks (CNN) and bi-directional long short-term memory (BLSTM) is recommended for achieving high accuracy with a small network size. Hardware architecture achieves a significant level of compute reuse by using the similarity between consecutive heartbeats to reduce the energy consumption of the categorization process. Neural network compression techniques are used during the inference phase to save hardware resources. Based on the simulation, this processor operates with a 1.1 V power supply and has a frequency of 100 MHz, achieving an average power efficiency of 8147mW. The suggested processor achieves a significant improvement in energy efficiency, with a 61.27% decrease in energy dissipation per categorization. Additionally, it demonstrates a substantial performance enhancement, with a 2.58x increase in performance compared to the architecture without compute reuse. This research is a significant investigation into the use of neural networks for continuous monitoring of abnormal heart rhythms in daily activities.