Heart sound classification via sparse coding

Introduction: The aim of the Physionet/CinC Challenge 2016 is to automatically classify heart sound recordings as normal or abnormal. The Challenge provides 3,153 labeled audio recordings taken from a single precordial location, as well as Springer's state-of-the-art beat segmentation algorithm. Algorithm: Using Springer's segmentation algorithm, we divide each audio segment into an array of sub-second audio files corresponding to the four phases of the cardiac cycle. We take an N-point FFT of each audio segment and create five different data matrices: one for each sub-cycle (S1, Systole, S2, and Diastole), and one for a complete cardiac cycle. A column of the data matrix corresponds to the N-point FFT of one audio segment. Using sparse coding, we decompose the data matrix into a dictionary matrix and a sparse coefficient matrix. The dictionary matrix represents statistically important spectral features of the audio segments. The sparse coefficient matrix is a mapping that represents which features are used by each segment. Working in the sparse domain, we train support vector machines (SVMs) for each sub-cycle and for the complete cycle. We train a sixth SVM to combine the results from the preliminary SVMs into a single binary label for the entire sound recording. Results: Our algorithm achieves a cross-validation score of 0.8652 (Se=0.8669 and Sp=0.8634). The best unofficial score when tested on a subset of the unknown challenge data is 0.812 (Se=0.825 and Sp=0.799). Conclusions: We developed an algorithm to classify heart sound recordings as normal or pathological. Our results show that sparse coding is an effective way to define spectral features of the cardiac cycle and its sub-cycles for the purpose of classification. Further work will attempt to increase the sensitivity and specificity of the algorithm by exploring other classifiers while still working in the sparse domain.