Heart disease remains the leading cause of death globally, and early detection plays a vital role in reducing mortality. Phonocardiograms (PCGs), which record heart sounds via digital stethoscopes, offer a portable and cost-effective diagnostic modality. However, manual PCG analysis is time-consuming and requires expertise. In this study, we present a machine learning-based pipeline to classify heart sounds as normal or abnormal for early disease detection, crucial for cardiac diagnostics. We propose a twofold approach: i) a deep learning model using a one-dimensional convolutional neural network (1D-CNN) trained directly on raw PCG recordings, and ii) traditional machine learning models, support vector machine (SVM), random forest (RF), and eXtreme gradient boosting (XGBoost), trained on clinically meaningful features. The dataset comprises 3,240 PCG recordings from the PhysioNet Challenge 2016. The models were evaluated using k-fold cross-validation, and performance was assessed via accuracy, precision, recall, and F1-score. Explainability techniques such as gradient-weighted class activation mapping (Grad-CAM) and local interpretable model-agnostic explanations (LIME) were applied to improve clinical interpretability. Our results show that the feature-based models outperform the CNN in transparency and time efficiency, highlighting the value of interpretable artificial intelligence (AI) systems in healthcare diagnostics.