Unsegmented Heart Sound Classification Using Hybrid CNN-LSTM Neural Networks

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Signal Processing

Background:

• Cardiac auscultation is crucial for diagnosing cardiovascular diseases (CVDs).
• Accurate interpretation of heart sounds requires significant clinical experience.
• A deficiency in diagnostic expertise exists among early-career medical professionals.

Purpose of the Study:

• To develop an automated diagnostic tool for classifying unsegmented heart sounds.
• To aid medical practitioners in the early diagnosis of cardiovascular diseases.
• To improve the accuracy and accessibility of cardiac auscultation interpretation.

Main Methods:

• Proposed novel hybrid deep learning architectures for heart sound classification.
• Developed two methods: one with conventional feature extraction and one without.
• Utilized the Physionet dataset for training and validation.
• Introduced a mechanism for tagging uncertain predictions.

Main Results:

• The hybrid architecture with conventional feature extraction demonstrated a 1.25 absolute F-score improvement over a baseline.
• The proposed methods effectively classify heart sounds into normal and abnormal categories.
• The mechanism for tagging unsure predictions was evaluated against varying thresholds.

Conclusions:

• Novel hybrid architectures show promise for automated heart sound analysis.
• The developed tool can assist clinicians, particularly those with less experience.
• Further refinement of AI in cardiac diagnostics can enhance early disease detection and patient outcomes.