PCG Signal Acquisition and Classification for Heart Failure Detection: Recent Advances and Implementation of Memory-Efficient Classifiers for Edge Computing-Based Wearable Devices

Conventional diagnostic tools for cardiovascular diseases usually employ expensive instrumentation and require specialized medical staff. An inexpensive and non-invasive alternative is the phonocardiogram (PCG). This paper presents the development of classifiers for binary (Normal/Pathological) and multiclass classifications of PCG signals. The latter discerns between a subset of heart diseases (mitral valve prolapse (MVP), coronary disease (CAD), and benign murmurs (Benign)). Two balanced datasets were created from the Physionet 2016/CinC database, consisting of 10104 and 13136 5-s frames. A custom preprocessing chain includes denoising, normalizing, and splitting the PCG signals, making them suitable to extract the scalar features set, constituting the training and test set. Several ML/DL models (e.g., SVMs (Support Vector Machines), k-NNs (k-Nearest Neighbors), and NNs (Neural Networks)) were trained and tested to classify the PCG signals. For binary classification, three different NNs have reached 96.0%, 95.9%, and 93.4% accuracy, and 95.9%, 96.0%, and 93.3% F1-scores, respectively. However, k-NN classifiers provide higher accuracy (up to 98.7%) than NNs but require much larger memory (up to 11 MB). As for the multiclass classification, three custom NNs have achieved 96.0%, 95.8%, and 94.7% accuracy with 735 kB max memory occupation. The developed classifiers provide a good balance between complexity and performance, with the latter not dependent on signal quality. In the feature engineering phase, the heart sound segmentation was not performed to make the classifiers suitable for resource-limited platforms.