Backgrounds:

Artificial intelligence (AI) based on deep learning is recently developing remarkably. The aim of this study was to develop a neural network program to detect severe aortic stenosis based on the heart sounds, which can improve the efficacy of the diagnostic workflow for screening and help medical staffs in the daily clinical practice.

Methods:

We enrolled 70 patients with severe aortic valve stenosis (severe AS; sAS group) who underwent echocardiography. As controls, 70 patients without severe AS (non-sAS group; no echocardiographic sign of AS [n=28], mild AS [n=21], moderate AS [n=21]) were also enrolled. In each patient the heart sounds were recorded with 4000 Hz WAV format in the following three regions; the 2^nd intercostal right sternal border, Erb’s area, and apex. Each record was divided into 10 WAV data of 6 seconds duration, which built 4200 sound records in total. We developed 7-layer convolutional neural networks (CNN). The CNNs were designed to recognize severe AS. We adopted stratified 6-fold cross-validation method to assure the quality of CNNs which were trained by data from 100 patients (1500 sAS and 1500 non-sAS sound data) and validated by data from 20 patients (300 sAS and 300 non-sAS sound data). The average performance of the 6 trained CNNs in each cross-validation fold was evaluated with the data of 20 patients (600 test-set heart sounds) not used during the training, which had been hold out prior to the build of 6 cross-validation sets. As performance indices of CNNs we adopted accuracy, F1 value (harmonic mean of positive predictive value and sensitivity), and area under the curve (AUC).

Results:

There were no significant differences in the patients’ characteristics between the sAS and the non-sAS patients except for age and the PQ interval (Table). The average accuracy, F1 value of sAS recognition in the heart sounds by our CNN were 85.8% and 0.871, respectively. The average sensitivity, specificity and AUC were 95.0%, 76.7% and 0.953. The best performed model showed the sensitivity of 95.7%, the specificity of 82.0%, the positive predictive value of 84.2%, and the negative predictive value of 95.0% (AUC = 0.964). In the detailed analysis, as shown in Figure, the recognition accuracy was significantly higher in the patients without AS and with severe AS as compared to those with mild and moderate AS (98.6±2.6%, 95.0±1.8% vs 71.3±5.5%, 52.6±12.4%, respectively, P<0.05). In patients without AS, the recognition failure was frequently noted if the frictional noise with the skin or stethoscope tubing was recorded in the heart sound data.

Conclusions:
Our data suggest that our CNN can detect heart sound of severe aortic valve stenosis accurately. This technology may support the efficacy of daily medicine particularly where the human resource is limited. The noise reduction technique for the heart sound data and more precise classification of intermediate degree of AS should be achieved in the future projects.

Table: