Introduction: This thesis explores the use of machine learning to automatically detect abnormalities in heart sounds, known as phonocardiograms (PCGs). These sounds, recorded using digital stethoscopes, carry vital information about the mechanical function of the heart. The main objective of this work is to develop an accurate and interpretable system that supports early diagnosis of heart conditions in a practical and clinically relevant way.

Related Work: Previous studies have used both traditional machine learning and deep learning for heart sound classification. Traditional models often relied on basic features like zero-crossing rate or spectral centroid, but lacked clinical relevance. Deep learning models improved accuracy but acted as black boxes, offering limited interpretability. Most prior work did not incorporate explainability, which is essential in medical contexts. This thesis addresses these gaps by comparing a feature-based method with a deep learning baseline, focusing on transparency and clinical usefulness.

Method: Two approaches were implemented and evaluated. First, a one-dimensional convolutional neural network (ID CNN) was used as a pilot study to assess classification performance and examine the limitations of black-box models. Although Grad-CAM was used to highlight signal regions influencing predictions, the specific clinical relevance such as which features made the model label a heart sound as abnormal remained unclear. To address this, a second, feature-based pipeline was developed. It involved signal denoising, S1/S2 detection, and extraction of clinically relevant features like heart rate, cardiac cycle duration, MFCCs, entropy, and energy. These features were used to train interpretable models including Random Forest, SVM, and XGBoost. LIME was then applied to visualize how each feature influenced the prediction, making it possible to understand what aspect of the heart sound such as frequency patterns in MFCCs or timing irregularities may be contributing to an abnormal classification. All experiments used 3,240 PCG recordings from the PhysioNet database.

Results and Analysis: Both approaches successfully classified heart sounds as normal or abnormal. However, the feature-based models offered more transparent reasoning be hind their decisions. With LIME, it became pogible to see which clinically relevant features such as changes in MFCCs or heart rate variability were contributing to a specific prediction, helping us understand what aspect of the heart sound indicated an abnormality.

Discussion: This study shows that combining clinical signal features with interpretable machine learning models results in systems that are not only accurate but also understandable and practical for clinical use. It also reveals that while deep learning models like ID CNNs can achieve good accuracy, they lack the transparency needed for medical decision-making unless paired with effective explainability tools.

Conclusion: By focusing on explainable classification of PCG recordings, this thesis contributes toward building machine learning systems that support early detection of heart abnormalities in a way that clinicians can trust and act upon with confidence.