Stratification of Heart Sounds Morphology Through Unsupervised Learning

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Telemonitoring of hemodynamic conditions relies heavily on heart sound analysis.
• Current methods struggle with the morphological variability of first (S1) and second (S2) heart sounds, limiting component separability.
• Morphological interpretation of heart sounds remains an underexplored area.

Purpose of the Study:

• To develop and evaluate a method for stratifying S1 and S2 heart sounds based on their morphology.
• To explore the diversity and enhance the morphological interpretability of heart sounds.
• To improve the analysis of heart sounds for hemodynamic assessment in telemonitoring.

Main Methods:

• A novel unsupervised learning approach utilizing a cascade of four Self-Organizing Maps (SOMs) with decreasing dimensions.
• Clustering of heart sounds based on morphological characteristics.
• Validation on a publicly available heart sounds dataset.

Main Results:

• The proposed clustering method demonstrated robustness and consistency.
• Over 80% of heartbeats from the same patient were successfully clustered together.
• Identified heart sound templates revealed significant differences in time and energy domains.

Conclusions:

• The SOM-based stratification effectively categorizes heart sound morphologies.
• This approach enhances morphological interpretability, offering new avenues for heart sound analysis.
• The findings support the potential of this method for advanced telemonitoring applications.