Breathing and speech patterns for predictive modelling of exacerbations of asthma and COPD

BACKGROUND: COPD and asthma cause significant morbidity and mortality across the globe. The development of non-invasive telemonitoring solutions offers the potential to improve patient care. OBJECTIVE: To examine whether breathing and speech patterns extracted from audio recordings enhance model generalizability of distinguishing stable from exacerbation periods in COPD and asthma. METHODS: Patterns related to breathing and speech were extracted from a dataset that combined daily audio speech recordings with a patient-reported outcome measure (EXACT). Medical experts categorized patient health status as stable or exacerbation. A tree-based model was investigated for exacerbation prediction. The model was trained on signal-related features, breathing and speech patterns, and combination of both. Results were examined for variability among all people and feature importance. RESULTS: The study included 21 people (mean age 62.7 years; 66.7% female), 61.9% diagnosed with COPD and 38.1% with asthma. Minimum breath duration and speech duration differed between stable and exacerbation states (p < 0.05). The model, trained on combined features sets achieved Sensitivity of 0.78, 0.10 SD exceeding models trained on acoustic-only (0.70, 0.17 SD) and patterns-only sets (0.68, 0.13 SD). Inconsistencies were observed in patient-level results across exacerbation severity. Speech Duration, Spectral Contrast, Harmonic to Noise Ratio, Breath Group Duration and 26 Mel-Frequency Cepstral Coefficient were found to have the highest importance for the best performing model. CONCLUSION: Combining signal derived features with breathing and speech patterns helps models to achieve higher results. Detected clusters within the study group indicate patient-level variability. Presented results position breathing and speech patterns as valuable tool for remote patient screening.