Deep-Learning Based Segmentation of In-Ear Cardiac Sounds.

Area of Science:

• Biomedical Engineering
• Cardiology
• Signal Processing

Background:

• Cardiovascular disease is a major global health concern.
• Heart sound segmentation is vital for diagnosing valve irregularities but requires expertise.
• Current automated methods rely on specialized medical devices, limiting continuous use.

Purpose of the Study:

• To explore the potential of ear-based devices for continuous heart sound segmentation.
• To address the signal differences between in-ear microphones (IEM) and traditional phonocardiographs (PCG).
• To develop and evaluate a deep learning model for IEM-based heart sound segmentation.

Main Methods:

• Analysis of temporal and frequency characteristics distinguishing IEM and PCG signals.
• Development of a U-Net deep learning model specifically for in-ear heart sound segmentation.
• Implementation of a rigorous evaluation metric to assess segmentation accuracy.

Main Results:

• The proposed U-Net model achieved 84% accuracy in heart sound segmentation using IEM data.
• The model significantly outperformed existing baseline methods.
• The study highlights the feasibility of using readily available earable devices for cardiac monitoring.

Conclusions:

• Earable devices offer a promising avenue for portable, continuous cardiovascular monitoring.
• The developed U-Net model is effective for heart sound segmentation from in-ear recordings.
• This approach could facilitate wider accessibility to cardiac diagnostics and remote patient monitoring.