Abstract Background Mobile health (mHealth) technologies offer a promising avenue for non-invasive heart failure monitoring through vocal biomarkers. The relationship between feature changes in anonymous and standardized voice recordings of heart failure patients and the NYHA scores of the same patients is assessed, aiming to identify vocal biomarkers as delicate indicators of chronic noncommunicable disease progression and/ or fluctuation. Methods In a cardiology outpatient setting, 55 chronic heart failure patients provided voice recordings via a mobile app, performing specific vocal tasks prospectively. We analyzed waveforms visually via Mel spectrograms and extracting features by audio processing, focusing on chroma centroids and tonnetz measurements across 147 clinical and laboratory parameters (Mean NYHA: 2.62, SD: 0.56). Applying Ordinary Least Squares (OLS) test, we assessed the relevance of spectral bandwidth mean (0.04), spectral bandwidth std (0.00), spectral roll-off std (0.011) between NYHA scores, given each p-score < 0.05. Results 147 voice recordings taken from 55 patients were analyzed to examine whether there were any correlations between changes in the voice recordings and NYHA scores. OLS test underscored the significance of spectral bandwidth mean (0.04), spectral bandwidth std (0.00), spectral roll-off std (0.011) between NYHA scores with p-value < 0.05. These findings show the potential role of multiple vocal features in reflecting heart failure progression and/or fluctuation. Analysis based on the clinical status progression from NYHA 2 to NYHA 3 over time showed notable changes in chroma centroids (mean = 0.006, std = 0.174, skew = 0.179, kurtosis = 0.205) and tonnetz (mean = 0.023, std = 0.131, skew = 0.023, kurtosis = 4.247). Conclusion This investigation highlights vocal biomarkers' potential in heart failure management, offering insights into non-invasive monitoring techniques and shows the importance of standardised recording procedures and patient engagement for data integrity. Future directions include quantifying vocal changes' correlation with clinical status and integrating additional parameters (e.g., edema, weight) to enhance predictive models for personalised care.Vocal features according to NYHA scores
Read full abstract