Predicting exacerbation state of COPD patients with voice analysis: findings of the TACTICAS study

Novel technologies for the early detection of disease deterioration in COPD patients are warranted. We hypothesized that deteriorations in health status can be detected by changes in speech. Our aim was to train a machine learning model to capture early signs of such deterioration using longitudinal voice changes. <bold>Method:</bold> Voice recordings from COPD patients, collected over 12 weeks with mobile app, were linked to daily EXACT scores. Regression models were trained on spectral and phonation features extracted from recordings to predict daily EXACT-score deviation of at least 5 points from baseline, enabling the classification into three classes (improvement, stable, deterioration) as in Fig. 1. <fig><object-id>erj;64/suppl_68/PA792/F1</object-id><object-id>F1</object-id><object-id>F1</object-id><graphic></graphic></fig> <bold>Results:</bold> We analyzed 2808 sustained vowel samples of 27 COPD patients. Support vector regression (SVR) model trained on spectral features showed the best performance with total accuracy and recall of 69%, F1 score 65%. Highest accuracy for single-class prediction was 84% for “stable” and 72% for “deterioration” state (Fig. 2). <bold>Conclusion:</bold> The study shows promising results of ML models to detect changes in respiratory health status of COPD patients. <fig><object-id>erj;64/suppl_68/PA792/F2</object-id><object-id>F2</object-id><object-id>F2</object-id><graphic></graphic></fig>