Phonocardiogram Classification by Learning From Positive and Unlabeled Examples

The advent of deep learning has rekindled research in computer-aided auscultation for the classification of phonocardiogram (PCG) signals. Deep-learning techniques require a large labeled dataset for training. However, labeling large datasets is a formidable task. To create the large labeled dataset, the labeled PCG signal records are segmented cycle-wise, and the labels of the records are passed on to all the cycles of the record. Although this label inheritance may be appropriate for segments of the normal PCG signals, it may be inappropriate for abnormal PCGs and, therefore, may result in the wrong labeling of cycles of abnormal PCG signals. To address this issue, we propose positive unlabeled (PU) learning based on a two-step technique deep-learning model for the classification of PCG signals where PCG segments of normal records are considered positive exemplars and PCG segments of abnormal records are considered unlabeled. To attain the final results, a voting method is employed using a differential evolution (DE) algorithm and majority voting. The system was evaluated using a dataset from the 2016 Physionet/Computing in Cardiology Challenge. The proposed system achieved exceptional record classification performance with a score of about 0.95. Regardless of an imbalanced dataset, the method achieved balanced specificity and sensitivity values of about 0.94 and 0.95, respectively. Additionally, the proposed system outperformed existing human PCG binary classification systems. In conclusion, utilizing PU learning and convolutional neural network (CNN) techniques for diagnosing heart sounds can lead to effective and efficient classification.