Introduction: Early and valid identification of cardiac pathologies, leading causes of death worldwide, remains a challenge in modern medicine, especially in rural areas. Phonocardiography (PCG) is a widely available and cheap method for cardiovascular screening, but interpretation demands an experienced physician. Attempts to use computer algorithms for PCG interpretation were made as early as the 1970s. Machine Learning (ML) has shown great potential in interpreting medical data within the last few years. We therefore reviewed publicly available literature on ML for PCG interpretation published since 2019 and performed a meta-analysis on the obtained data.

Methods and Results: A comprehensive literature search was conducted in the “PubMed” database. Only original articles published since 2019 in peer-reviewed journals were considered for this study if the article evaluated ML to interpret PCG signals in adults with valve disease as outcome variable. Four reports were identified that described ML to interpret PCG signals with contingency tables available. If studies further discriminated abnormal heart sounds the contingency table was reduced to presence or absence of abnormal heart sounds representing potential valve diseases. If an article examined multiple ML-algorithms, the best performing one was chosen. For comparability between the articles, data includes results for the testing and validation datasets. Table 1 shows the included articles. Within the meta-analysis, principal measurement of effect size was the diagnostic odds ratio after natural logarithmic transformation (lnDOR) with 95% confidence interval (CI). All evaluated studies showed the ability of ML applied to PCG data to identify potential valve diseases with lnDORs ranging from 3.24 to 12.62 leading to a calculated meta lnDOR of 8.09 [2.24 – 13-94] with a heterogeneity I² of 85% (p<0.001). The results are summarized in figure 1.

Discussion and Conclusion: The present study confirms the potential of ML to discriminate between normal and abnormal cardiac auscultation sounds. ML might therefore be used in wide-spread population screenings for cardiac pathologies in the future identifying patients needing a cardiac evaluation. Advances in heart sound recording, for instance via smartphone microphones, further emphasize a future for ML-assisted cardiac auscultation especially regarding rural regions in a worldwide perspective.

Table 1:

Figure 1: