Multi-Class Lung Disease Classification Using Google's HeAR Foundation Model and MFCC Features: A Comparative Study of Performance and Data Efficiency

Over 454 million individuals worldwide suffer from chronic respiratory conditions, with low- and middle-income countries bearing a disproportionate burden due to inadequate diagnostic facilities. Although foundation audio models remain underexplored in this field, automated classification of respiratory sounds could support early disease detection at scale. This study evaluates Google's Health Acoustic Representations (HeAR) model for multi-class lung disease classification and directly compares it with conventional Mel-frequency cepstral coefficient (MFCC) features across five classifiers, including SVM, Gradient Boosting, and MLP. Using the Asthma Detection Dataset Version 2, audio was segmented into 3,602 two-second clips, balanced with SMOTE, and evaluated on a stratified held-out test set. Model performance was assessed using training data fractions ranging from 10% to 100% in a data efficiency experiment. The best HeAR-based model, MLP, achieved a macro F1-score of 84.5% and accuracy of 86.4%, while MFCC features combined with Gradient Boosting produced the highest overall performance, with 88% accuracy and 87% macro F1-score. HeAR embeddings consistently outperformed linear classifiers under limited data conditions. At 10% training data, HeAR SVM achieved a macro F1-score of 70%, compared to 55.8% for MFCC SVM. While MFCC features with non-linear ensembles delivered superior peak performance on controlled single-source data, HeAR embeddings produced a more linearly separable feature space, enabling stable classification with less labelled data, making them suitable for resource-limited clinical settings.