An efficient lung sound separation algorithm base on GIHO-VMD

Lung sounds (LS) serve as a critical source of pathological information for diagnosing lung diseases. However, during clinical auscultation, LS are often contaminated by heart sounds (HS) and other noises, significantly compromising the accuracy of pathological assessment. The primary challenge lies in the spectral overlap between LS and HS, coupled with the non-stationary nature of both signals. Furthermore, abnormal LS exhibit intricate frequency components and ambiguous band boundaries, causing conventional separation methods to yield incomplete or contaminated results, which hinders precise diagnosis.

To overcome these challenges, we propose a Guided-Improve Hippopotamus Optimization based Variational Mode Decomposition (GIHO-VMD) algorithm for LS separation. Specifically, a VMD model is first established with Fuzzy Entropy (FE) as the objective function. Critically, a novel complexity constraint is incorporated into this model to effectively restrict decomposition outcomes. The Hippopotamus Optimization (HO) algorithm is then employed to iteratively optimize the VMD parameters based on this objective function. To prevent the optimization from converging to local optima, a Disruption Operator (DO) is integrated into HO. Subsequently, a Guided Learning Strategy (GLS) is introduced to enhance the guiding influence of superior agents during iteration, thereby improving convergence accuracy. Finally, a dual-process dynamic threshold mechanism is implemented within GLS to moderate the frequency of guidance, thereby stabilizing the exploration-exploitation balance throughout the optimization process.

To evaluate the proposed method, we apply it to separate pre-simulated mixed signals of HS and LS. The SNR and NCC between the separated LS and the original ones are evaluated. Experimental results indicate that the method achieves an average SNR of 30.27 dB for normal lung sounds and an average SNR ranging from 28.60 dB to 30.11 dB for abnormal lung sounds, with an average NCC of 63.15%.

Our method demonstrates robust capabilities in lung sound signal separation, effectively addressing the challenge of extracting LS from heart and lung sound(HLS), and shows significant potential for enhancing the generalization, quality, and accuracy of lung sound signal decomposition.