Pulmonary Hypertension Detection From Heart Sound Analysis

Area of Science:

• Biomedical Engineering
• Cardiology
• Artificial Intelligence in Medicine

Background:

• Pulmonary Hypertension (PH) detection via computer analysis of digitized heart sounds offers a low-cost, non-invasive approach for early screening.
• Traditional methods for PH detection can be invasive or costly, highlighting the need for accessible screening tools.

Purpose of the Study:

• To evaluate the effectiveness of digital auscultation technologies for detecting Pulmonary Hypertension across different animal models and auscultation methods.
• To introduce and assess PH-ELM, a novel, resource-efficient model for PH detection based on extreme learning machine.

Main Methods:

• An extensive cross-domain evaluation methodology was employed, utilizing data from humans and porcine animals with both phonocardiography (PCG) and seismocardiography (SCG) signals.
• The performance of the PH-ELM model was compared against a leading deep network, focusing on resource efficiency (parameters, power) and accuracy, particularly on out-of-distribution testing.
• Analysis included evaluating the interchangeability of SCG and PCG signals, the utility of porcine heart sounds for human PH detection, and the impact of heartbeat data volume.

Main Results:

• Digital auscultation demonstrates significant promise for PH detection.
• Seismocardiography (SCG) and phonocardiography (PCG) signals were found to be interchangeable for training PH detectors.
• The PH-ELM model showed improved median accuracy (0.09 auROC) on out-of-distribution testing compared to a deep network, while being smaller, more energy-efficient, and faster.
• Porcine heart sounds effectively supported PH detection in human heart sounds, retaining 88% of baseline performance.
• Accurate PH detection performance could be maintained with as few as 10 heartbeats, with performance improving up to approximately 200 heartbeats per subject.

Conclusions:

• Digital auscultation, particularly using the PH-ELM model, presents a viable, efficient, and accurate method for early Pulmonary Hypertension detection and screening.
• The study validates the use of cross-species data (porcine) and reduced heartbeat counts for robust PH detection, enhancing the practicality of the technology.
• PH-ELM's superior efficiency and accuracy make it a promising tool for widespread, non-invasive cardiovascular health monitoring.