Speech-based monitoring of exacerbations in chronic respiratory diseases

<bold>Introduction:</bold> Deteriorations in respiratory health, also called exacerbations, are important events in the natural history of individuals living with chronic respiratory diseases such as COPD and asthma. <bold>Aim:</bold> Voice is known to change during respiratory distress. We hypothesized that exacerbations of COPD and asthma can be detected by AI models using voice as input. The aim was to train a machine-learning model for the detection of COPD and asthma exacerbations solely using speech recordings. <bold>Methods:</bold> We used the TACTICAS dataset which contains speech recordings from Dutch COPD and asthma patients, who had their voice collected daily via a mobile application for 3 months. Besides voice, an EXACT questionnaire was completed daily. Acoustic features, specifically Mel-frequency cepstral coefficients (MFCC), were extracted from speech recordings. Subsequently, a K-Nearest Neighbors (KNN) algorithm was developed using 10-fold cross-validation to predict the EXACT event. <bold>Results:</bold> The TACTICAS dataset contained 5460 speech recordings from 38 participants (18 COPD and 20 asthma), 1670 recordings could be classified as EXACT events and 3786 recordings were classified as non-EXACT events. The developed EXACT event classification model achieved an accuracy of 88.7% (SD: ±0.02), a sensitivity of 85.4% (SD: ±0.04), a specificity of 90.1% (SD: ±0.02), and an AUC score of 87.8% (SD: ±0.02) averaged across all folds. <bold>Conclusions:</bold> Exacerbations of asthma and COPD can be accurately detected using only the voice, hence opening up possibilities for patient monitoring through daily speech recordings. <bold>Funding:</bold> the TACTIAS study was sponsored by an unrestricted research grant from AstraZeneca and Boehringer Ingelheim.