Transfer learning based heart valve disease classification from Phonocardiogram signal

Physiological conditions that prevent heart valves from functioning precisely to ensure proper blood circulation are known as heart valve disorder (HVD). Detection of HVD is critical as untreated heart valve disease often develops life-threatening cardiac diseases. Typical HVD detection methods, like echocardiography, MRI, and cardiac CT, are costly, complex, and require robust healthcare infrastructure. Although, by simple non-invasive listening to heart sound irregularities, an expert physician can anticipate the signs of HVD from ancient times. Contemporary development suggests that with machine learning-based algorithms, a graphical representation of heart sound, known as the phonocardiogram (PCG), can effectively predict the anomaly in the valvular activity. In recent studies, deep learning-based strategies showed promising results in the PCG classification task but demand extensive resources and training data. This work investigates the merits of transfer learning (TL) using pre-trained convolution neural networks for the automatic PCG classification when data is scarce. With standard time–frequency representations (i.e., spectrogram, log-Mel spectrogram, and scalogram) as input features, audio and image-based pre-trained lightweight models are fine-tuned to categorize the PCG. The proposed YAMNet-based TL method classifies four types of HVD data collected from public heart sound databases and achieves overall accuracy, sensitivity, and specificity of 99.83%, 99.59%, and 99.90%, respectively. Alongside, it classifies the PhysioNet/CinC Challenge 2016 dataset into binary classes with 92.23% accuracy. The study achieves high classification metrics despite data scarcity. It also investigates the proposed method’s computational efficiency and robustness against practical noise contamination for performance evaluation in a possible real-life scenario.