Spectrogram-Driven Heart Sound Classification using YOLOv8 for Mobile Health Monitoring

This paper presents a novel approach using the classification of abnormal or normal phonocardiogram (PCG) recordings by transforming them to spectrograms and applying YOLOv8 model for real-time classification. Unlike traditional models that work directly with raw audio or time-series data, our approach adopts the use of spectrograms to exploit the strengths of deep learning-based classifiers resulting in better accuracy. The model was trained on the PhysioNet/CinC 2016 dataset with a remarkable accuracy of 94.5%, F1-score of 0.95, recall of 0.97, precision of 0.95, and MCC of 0.83. Compared to traditional models such as LSTM or ResNet-18, YOLOv8 had a clear advantage in handling class imbalance within the dataset given. Although the LSTM models give promising results, they have larger computational overheads, which makes them unsuitable for real-time applications. Future work will investigate the optimization of YOLOv8 for deployment on microcontroller-based systems to improve efficiency within mobile health applications by minimizing inference time and resource usage. From our results, we believe that this method holds great promise to be implemented towards improving portable health diagnostics as a more accurate and efficient means of cardiovascular monitoring.