Accurate and Low-Cost Cardiac Disorder Detection from Wearable Phonocardiogram Signals Using Hybrid Feature Selection...

Area of Science:

• Biomedical Engineering
• Cardiology
• Signal Processing

Background:

• Early cardiac disorder identification from Phonocardiogram (PCG) signals is crucial for clinical decisions.
• Wearable biosensors offer potential for continuous cardiac monitoring.
• Current methods may lack accuracy or require computationally intensive deep learning (DL) architectures.

Purpose of the Study:

• To propose a multi-stage hybrid feature selection-classification approach for accurate cardiac disorder identification.
• To enhance diagnostic accuracy without relying on computationally expensive DL models.
• To evaluate the clinical applicability of feature selection strategies with conventional classifiers.

Main Methods:

• Utilized mRMR, ReliefF, and Kruskal-Wallis for feature selection.
• Applied Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO) for feature space optimization.
• Tested selected features with k-nearest neighbor (k-NN), Support Vector Machines (SVMs), and Bagged Tree (BT) classifiers.

Main Results:

• Achieved high accuracy (up to 99.80%) and F1-score (99.50%) with Kruskal-Wallis+k-NN and ReliefF+k-NN.
• Hybrid models with PSO and ACO also demonstrated strong performance (99.60% accuracy).
• The proposed method showed significant improvements in model robustness and generalizability compared to DL approaches.

Conclusions:

• Well-designed feature selection strategies with conventional classifiers provide high accuracy for cardiac disorder detection.
• The framework offers enhanced clinical applicability and computational efficiency.
• This approach is a viable candidate for smart stethoscope-based early screening solutions.