Robust S1 and S2 heart sound recognition based on spectral restoration and multi-style training

Abstract Recently, we have proposed a deep learning based heart sound recognition framework, which can provide high recognition performance under clean testing conditions. However, the recognition performance can notably degrade when noise is present in the recording environments. This study investigates a spectral restoration algorithm to reduce noise components from heart sound signals to achieve robust S1 and S2 recognition in real-world scenarios. In addition to the spectral restoration algorithm, a multi-style training strategy is adopted to train a robust acoustic model, by incorporating acoustic observations from both original and restored heart sound signals. We term the proposed method as SRMT (spectral restoration and multi-style training). The experimental procedure in this study is described as follows: First, an electronic stethoscope was used to record actual heart sounds, and the noisy signals were artificially generated at different signal-to-noise-ratios (SNRs). Second, an acoustic model based on deep neural networks (DNNs) was trained using original heart sounds and heart sounds processed through spectral restoration. Third, the performance of the trained model was evaluated using the following metrics: accuracy, precision, recall, specificity, and F-measure. The results confirm the effectiveness of the proposed method for recognizing heart sounds in noisy environments. The recognition results of an acoustic model trained on SRMT outperform that trained on clean data with a 2.36% average accuracy improvement (from 85.44% and 87.80%), over clean, 20dB, 15dB, 10dB, 5dB, and 0dB SNR conditions; the improvements are more notable in low SNR conditions: the average accuracy improvement is 3.87% (from 82.83% to 86.70%) in the 0dB SNR condition.