Improving respiratory disease detection through SSL-enhanced acoustic analysis and exercise-rest measurements

Voice analysis has emerged as a promising non-invasive approach for monitoring respiratory and systemic health conditions. However, subtle physiological alterations are often difficult to capture using recordings collected at rest. In addition, combining traditional acoustic descriptors with modern self-supervised speech representations may provide complementary information for clinical voice analysis.

This study evaluates a generalized screening model integrating stress-induced acoustic analysis with machine learning. We investigate how physical exertion and the fusion of traditional acoustic features with self-supervised learning embeddings (such as wav2vec 2.0 and WavLM) enhance the diagnostic sensitivity of vocal and respiratory signals. Post-Acute Sequelae of SARS-CoV-2 (PASC) is used as a case study to evaluate the proposed framework.

Utilizing the DICOPERIA-Voice dataset (n = 154), we collected recordings of sustained vowel phonation (/a/) and voluntary coughing at two clinical moments: resting state and following a physiological stress protocol (six-minute walk and one-minute sit-to-stand tests). We employed a dual-feature extraction strategy, combining traditional acoustic biomarkers with high-dimensional Self-Supervised Learning (SSL) embeddings from wav2vec 2.0, WavLM and HuBERT. Binary classification (PASC vs. Healthy) was performed using Logistic Regression, evaluated via stratified 5-fold cross-validation.

Physical exertion significantly improved classification performance and reduced model variability across all tasks. The fusion of acoustic features, WavLM and wav2vec 2.0 achieved peak F1-scores of 82.2% for vowel phonation and 80.8% for coughing both in post-exercise conditions. A cross-task late fusion model aggregation reached the highest overall performance, with an F1-score of 87.7%.

Incorporating Self-Supervised Learning representations into acoustic analysis improves the sensitivity of voice-based screening, while post-exercise measurements further enhance the robustness and consistency of classification. Together, these strategies provide a scalable and objective framework for detecting respiratory and vocal sequelae in chronic or post-viral conditions. With further validation, this approach could be integrated into routine functional assessments, offering a rapid, non-invasive adjunct to clinical decision-making.