Abstract / Summary
Lung auscultation analysis is a fundamental and mostly used clinical method for diagnosing pulmonary disorders; however, distinguishing between chronic obstructive pulmonary disease (COPD) and Asthma remains difficult because of the overlap in physiological and structural characteristics of respiratory systems. This study introduces AVMD-ST, an IMF-guided spectrogram Transformer framework that employs FIR-conditioned VMD to preserve complementary spectral representations of non-stationary respiratory sounds, for the early detection and differential classification of lung conditions. The investigation utilizes lung sound recordings obtained from 200 participants (100 healthy, 50 with asthma, and 50 with COPD) collected via a custom Android application at AIIMS Raipur and two publically available datasets ICBHI challenge and KAUH data. Finite Impulse Response (FIR) filtering is employed for noise reduction, while Variational Mode Decomposition (VMD) is utilized to derive intrinsic mode functions (IMFs) that isolate clinically significant components such as wheezes and crackles. Mel-spectrograms generated from these IMFs are provided as input to the AST model, which leverages self-attention mechanisms to detect subtle acoustic variations and capture long-range temporal dependencies. The model is trained and evaluated on a pooled dataset combining AIIMS Raipur recordings with the publicly available ICBHI 2017 and KAUH datasets, using a subject-wise train/test split to prevent recording-level leakage. Achieving 95.75% classification accuracy, the proposed framework demonstrates strong potential as a non-invasive, scalable solution for early respiratory disease screening.