Deep Learning Hardware/Software Co-Design for Heart Sound Classification

This paper presents a software/hardware co-design for classifying three most commonly heart sounds classes: normal, murmur and extrasystole heartbeat. The detection system extracts Mel Frequency Cepstral Coefficient (MFCC)-based heart sound features to train different deep learning network architectures for multiclass classification. The software/hardware co-design for Long Short-Term Memory (LSTM) implementation indicates the multiclass classification accuracy of 85% can be achieved. The proposed heart sound classification platform has great development potential and good application prospects.