A Heart Sound Diagnosis Processing Unit Based on LSTM Neural Network

Cardiovascular disease is the deadliest disease in the world, so prevention and diagnosis of cardiovascular disease are essential. Manual auscultation cannot meet the demand for heart sound auscultation, and computer automatic heart sound diagnosis offers a new method. In recent years, wearable auscultation devices are receiving increasing attention. However, the low power consumption and high-performance requirements limit wearable device development. In this work, an LSTM-based (Long Short-Term Memory) low-power heart sound diagnostic processing unit (HSDPU) is proposed. Considering the differences between the actual heart sounds and the open-source heart sound dataset, we develop an FPGA system for heart sound acquisition. Data augmentation is used to extend the dataset in response to the imbalance between the collected dataset and the open-source dataset. We develop the heart sound diagnosis system and achieve an accuracy of 96.9%. Then the hardware implementation of the HSDPU is finished and verified by RTL simulation. Finally, we develop the FPGA prototype verification and layout design of the HSDPU. The post-simulation results show that the power consumption of the HSDPU is $289\mu\mathrm{W}$.