Machine learning-based phonocardiographic analysis for pulmonary hypertension screening

Introduction

The detection of pulmonary hypertension (PH), particularly by non-specialized personnel, remains a significant challenge in clinical practice. Developing noninvasive, accessible, and cost-effective solutions to aid in PH diagnosis is critical, especially in resource-limited settings. Automated analysis of cardiac auscultation and electrocardiogram (ECG), combined with advancements in telemedicine and artificial intelligence (AI), holds potential to improve the early identification of PH and make diagnostic tools more widely accessible. With this work, we aimed to evaluate the feasibility of using a bimodal stethoscope integrated with a machine learning (ML) algorithm for PH detection.

Methods

Phonocardiogram (PCG) data were collected from 12 patients using the Rijuven Cardiosleeve, a bimodal stethoscope capable of recording both heart sounds and electrocardiogram signals. As a reference standard, mean pulmonary artery pressure (mPAP) was measured via right heart catheterization (RHC). PH suspicion was assessed using echocardiography by measuring the maximum velocity of tricuspid regurgitation (TR). A velocity exceeding 2.8 m/s was considered suspicious for PH.

A machine learning (ML) algorithm was applied to PCG data collected at the pulmonary auscultation site, focusing on the analysis of the S2 fundamental heart sound to differentiate patients with elevated pressures (mPAP > 20 mmHg, 10 subjects) from those with normal pressures (mPAP ≤ 20 mmHg, 2 subjects).

Results

The automated PCG analysis, when compared to the gold standard RHC measurements, achieved an average area under the Receiver Operating Characteristic (ROC) curve (AUC) of 0.80, demonstrating a promising ability to differentiate between elevated and normal pulmonary pressures.

In the same cohort, echocardiographic analysis identified at least a moderate probability of PH (TR vmax > 2.8 m/s) in 6 of the 10 elevated pressure cases but failed to do so in 4 out of 10 cases (TR vmax ≤ 2.8), yielding a recall of 0.60.

Conclusion

This study highlights the potential of AI-driven analysis of cardiac auscultation and ECG as a noninvasive and accessible method for detecting PH. Its ease of use and ability to be performed by non-specialized personnel make it a promising tool for early PH identification, particularly in resource-constrained or telemedicine settings. In the future, this approach could also be combined with echocardiographic evaluation to enhance the accuracy of PH estimation. Further validation with larger, more representative datasets is required to confirm these findings and enhance clinical applicability.