AI-Driven Cardiac Diagnostics: Leveraging Vision Transformers for PCG Signal Classification

Abstract

work presents a comprehensive approach to classifying cardiac sound (PCG)

signals into normal and abnormal categories using hybrid deep learning and ma- chine learning techniques. Our goal is to develop a solution based on artificial

intelligence techniques. The study leverages the PhysioNet/CinC Challenge 2016 dataset, which consists of recordings from multiple environments and patient cases, introducing significant variability in signal quality, duration, and class balance. A

robust preprocessing pipeline was implemented to normalize and segment the sig- nals by patient. Time-domain and frequency-domain features were extracted to

improve representation. n our work, we proposed a hybrid architecture that inte- grates multiple models, long short-term memory (LSTM) networks to model tem- poral context, and an attention mechanism to weight the most informative heart- beats. The most relevant segments were selected using attention weights, forming

the final input to a random forest classifier. Experimental results demonstrate that the proposed system achieves promising classification performance, proving its effectiveness in handling PCG variability and imbalanced real-world data. This

work highlights the potential of hybrid deep learning methods in automated car- diac listening systems