Heart sounds analysis using probability assessment

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Automated analysis of heart sounds is crucial for diagnosing cardiac conditions.
• The Physionet Challenge 2016 aimed to develop algorithms for classifying heart sound recordings.

Purpose of the Study:

• To develop and evaluate an automated method for discriminating between normal, abnormal, and unsure heart sound recordings.
• To classify heart sound recordings based on automated feature extraction and probability assessment.

Main Methods:

• Heart sounds (S1 and S2) were detected using amplitude envelopes across multiple frequency bands.
• A total of 53 features were extracted, focusing on statistical properties and symmetry of averaged sound envelopes.
• A machine learning approach using logical rules and probability assessment was employed.

Main Results:

• The method achieved a training score of 0.903 and a testing score of 0.841 on 1277 hidden records.
• A revised method improved the test score to 0.853 during the challenge's follow-up phase.
• The system demonstrated high sensitivity and specificity in classifying heart sounds.

Conclusions:

• The developed automated heart sound classification method ranked 7th out of 48 entries in the Physionet Challenge 2016.
• The PROBAfind software was introduced as a tool for probability assessment in heart sound analysis.