arXiv Preprint Introduces Vision Transformer System for Heart Murmur Classification

A new research paper proposes an automated system for classifying systolic heart murmurs, which are additional sounds during the heart's contraction phase that can signal cardiac abnormalities. The method uses a feature extraction module that applies complex orthogonal matching pursuit to murmur segments. These segments are projected onto a redundant dictionary built from multiresolution complex Gabor basis functions. The resulting projection weights are then transformed into variable-resolution time-frequency feature matrices. A classification model processes these matrices. The approach handles multiple segments from a single recording by using a shared dictionary, which helps address the inherent variability of murmurs by learning specific weights for each segment while enforcing consistency. Systolic murmurs arise from turbulent blood flow and vary in intensity, pitch, and quality, making their precise identification crucial for accurate cardiac diagnosis. The study is documented in the preprint arXiv:2604.16563v1, which was announced as a cross-type submission.