LingoNMF: Large language model-enhanced nonnegative matrix factorization for cardiorespiratory sound separation

Auscultation is a rich source of information for diagnosing cardiorespiratory diseases. However, auscultation is challenging due to noise and mixing of heart and lung sounds. In this study, we propose a novel periodicity-based parallel non-negative matrix factorization framework (PL-NMF) for blind separation of lung and heart sounds recorded by a digital stethoscope. PL-NMF introduces a parallel multilayer structure with tunable scaling and offset parameters that exploit the relative periodic properties of cardiopulmonary signals without requiring explicit prior frequency knowledge. Building upon this framework, we further develop LingoNMF, which integrates large language models (LLMs) into the optimization process. LLaMA (Large Language Model Meta AI) is employed in two ways: (1) to provide adjunct feature-based observations of potential clinical relevance, and (2) to adaptively refine fundamental frequency estimates that enhance separation performance. We evaluated the proposed method on two datasets: 100 cases consisting of mixtures of real measurements, and 210 recordings of heart and lung sounds from a clinical manikin captured using a digital stethoscope. LingoNMF achieved maximum observed gains of up to +10.2 dB in Signal-to-Distortion Ratio (SDR), +11.5 dB in Signal-to-Artifacts Ratio (SAR), and +6.4 dB in Signal-to-Interference Ratio (SIR) compared with existing baseline methods (SDR = 22.3 dB, SAR = 25.2 dB, and SIR = 22.4 dB), demonstrating its potential to enhance medical sound analysis for disease diagnostics.