We propose new cardiac disorder detection and classification methods by heart sound signals using hidden Markov model (HMM) and murmur score.
First, we evaluate the baseline system using the cepstral features. In the baseline system, we segment input heart sound signals into the HMM states for all cardiac disorder models and compute the log-likelihood (score) for each state of the models. To reflet the spatio-temporal characteristics of heart sound and murmur signals, we compute the mel-frequency cepstral coefficients (MFCC) from a single period of heart sound signals. Then we use pattern classifiers to determine the cardiac disorders. In computer experiments, the baseline system using the MFCC features achieves the detection accuracy of 96.9%, 96.6% and 93.2%, and the classification accuracy of 73.1%, 76.3% and 75.6% using the MLP-, SVM-, and HMM-based classifiers, respectively.
Second, in order to consider the temporal position characteristics of murmur signals, we add three kinds of temporal features to the conventional MFCC features: Heart sound envelope, murmur probability, and murmur amplitude variation. In cardiac disorder classification experiments, the added murmur probability is shown to achieve the detection accuracy of 97.5%, 97.5% and 98.2%, and the classification accuracy of 79.4%, 80.0% and 81.3% for the MLP-, SVM-, and HMM-based classifiers, respectively.
Third, to reduce the classification errors occurring when heart sound signals have similar murmur positions but different spectral characteristics, we extract the state-based HMM scores and the murmur scores. We divide the input signals into two subbands and compute the murmur probability in each subband for each frame, and obtain the murmur score assigned to every HMM state by using the Viterbi segmentation algorithm. Then the SVM classifies the cardiac disorder category by using the HMM state scores and the murmur scores for all cardiac disorder models as the input vector of the SVM. In cardiac disorder detection experiments, the proposed method yields the detection accuracy 98.2%, the false rejection rate 1.1%, and the false alarm rate 2.5%. In cardiac disorder classification experiments, the proposed method achieves the relative improvement of 20.4% compared with the HMM-based classifier with only the conventional cepstral features.