Phonocardiogram Classification Using Dynamic Mode Decomposition for Heterogeneity-Resilient Training

Medical measurements of the heart, such as a phonocardiogram (PCG), provide a non-invasive means for diagnosing valvular heart diseases. Recently, deep learning has shown great success in the classification of PCG. Current deep learning models for binary classification of PCG utilize measurements obtained in controlled or uncontrolled environments using different stethoscope sensors, each with distinct characteristics, placed at different measurement sites on the heart. Such acquisitions introduce heterogeneity in the measurements that are used for training deep learning PCG classification models. Such trained models deliver reduced balanced accuracy. Training a separate deep learning model for each sensor or each measurement site is impractical. Therefore, a unified model that can generalize across sensor and measurement variations is needed. In this work, Dynamic Mode Decomposition (DMD) in conjunction with domain-balanced training is proposed to address measurement heterogeneity and enable robust model training on mixed-source measurements. Experimental results show that the proposed heterogeneity addressing method outperforms existing approaches that ignore heterogeneity by achieving approximately 15% improvement in balanced accuracy, while 4% improvement in balanced accuracy amongst existing methods that address heterogeneity in binary classification of PCG.