LF-NET: A Low-Frequency Self-Supervised Model for Coronary Artery Disease Detection From Phonocardiogram Signals

Coronary artery disease (CAD) remains the leading cause of cardiovascular morbidity and mortality worldwide. As myocardial infarction or stroke maybe the first manifestation of CAD, there is a critical need for early, non-invasive pre-screening. Heart sounds, recorded as phonocardiogram (PCG) signals, provide a low-cost diagnostic avenue. However, CAD-related murmurs are low in both frequency and intensity, rendering them inaudible during traditional auscultation. Deep learning has shown promise for PCG-based CAD classification, yet limited dataset sizes have driven reliance on fine-tuning foundation models. Models such as Wav2Vec 2.0, HuBERT, and EnCodec are pre-trained primarily on speech and environmental audio, while others like OPERA utilise respiratory sounds. Consequently, their feature representations are biased toward higher-frequency content and may be suboptimal for low-frequency PCG analysis. To address this, we propose LF-NET and LF-NET2D; self-supervised frameworks pre-trained on low-frequency ocean sound recordings that share similar frequency range with PCG signals. An energy-based noise removal algorithm is integrated into the proposed framework to extract the cleanest PCG segments from recordings obtained in real clinical environments with typical hospital noise. After fine-tuning, these lightweight models achieve comparable single-channel accuracy to existing foundation models while using fewer parameters and outperform them in multichannel configurations. Fusing LF-NET with OPERA-CE yields a state-of-the-art subject-level accuracy of 80.88% on the current 296-subject practically challenging dataset, approximately 3% higher than standalone existing foundation models.