Cardiovascular diseases (CVDs) are the main cause
of global mortality, and heart rhythm disorders pose major
diagnostic challenges, especially in environments limited by
resources. To address this problem, we introduced a hybrid deep
learning architecture that combines the Convolutional Neural
Network (CNN) and the Gated Recurrent Unit (GRU), and
classified a PCG heartbeat sound into normal and abnormal
categories. CNN components extract spatial features from the
sound of the heartbeat, while GRU models temporal patterns
within the heart cycle. This fusion allows the system to effectively
capture the structural and sequential characteristics of
heartbeat signals. The data sets include 1,000 PCG recordings
evenly divided into five classes: aortic stenosis, ventricular
regurgitation, ventricular stenosis, ventricular valve collapse,
and normal. In the binary classification (normal or abnormal),
the weight loss function was used to solve class imbalances.
Further techniques such as stratified partitioning, dropout, early
stopping, and real-time validation have been used to improve
model training stability and generalization. Methods of data
augmentation, including noise addition, volume change, time
expansion and spectral filtering, were implemented to simulate
the variability of heart sounds in the real world. CNN-GRU
achieved 97% percent test accuracy and 98% percent F1-
scores, with 198 of 203 test samples correctly identified. The
high sensitivity and specificity of the model highlight its ability
to learn key cardiac features from raw PCG data. These
results confirm the clinical feasibility of this model, which offers
robust and scalable tools for detecting and improving diagnostic
accuracy both in well-equipped and low-resource environments.