E2E Multi-Scale CNN with LSTM for Murmur Detection in PCG or Noise Identification

Cardiac auscultation serves as a cost-effective method for early screening of cardiac diseases. In our pursuit of creating a cardiovascular pre-screening tool, we have developed an AI-enabled digital stethoscope known as AiSteth. Within this AI system, we introduce an innovative approach to detect heart murmurs using an end-to-end multi-scale convolutional neural network (M-CNN) in combination with long short-term memory (LSTM). This approach employs two CNN modules with varying receptive field to extract features from raw data and utilizes LSTM networks to capture temporal dependencies. The model was trained to classify input signal as noise, heart sounds with murmurs, or heart sounds without murmurs.Our evaluation, conducted on the noisy dataset from the George B. Moody PhysioNet Challenge 2022 and AiSteth recordings, showcases superior performance in comparison to the state-of-the-art hierarchical multi-scale convolutional neural network (HMS-Net) for murmur detection. The average F1 scores achieved with our model is 86.13% for identifying normal heart sounds and 83.31% for detecting murmurs.