Heart Sound Classification for Early Detection of Cardiovascular Diseases Using XGBoost and Engineered Acoustic...

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Early detection of cardiovascular diseases (CVDs) through heart sound analysis is crucial for patient outcomes.
• Traditional acoustic analysis and modern deep learning embeddings offer potential for heart sound classification.

Purpose of the Study:

• To compare the efficacy of traditional acoustic features versus Wav2Vec embeddings for multi-class heart sound classification.
• To evaluate hybrid CNN+LSTM and XGBoost models using these different feature sets.

Main Methods:

• Extracted ten engineered acoustic features (e.g., MFCC, DWT, ZCR) and Wav2Vec embeddings.
• Trained and evaluated CNN+LSTM and XGBoost models on five heart sound categories.
• Utilized hyperparameter tuning and cross-validation on a held-out test set.

Main Results:

• Models trained with engineered acoustic features consistently outperformed those using Wav2Vec embeddings.
• XGBoost achieved the highest accuracy (99%), outperforming the hybrid CNN+LSTM model (98%).

Conclusions:

• Domain-specific feature engineering remains highly effective for heart sound classification.
• XGBoost demonstrates robust performance, offering a promising tool for early cardiovascular disease detection.