Early detection of high blood pressure from natural speech sounds with graph diffusion network

Abstract

This study presents an innovative approach to cuffless blood pressure prediction by integrating speech and demographic features. With a focus on non-invasive monitoring, especially in remote regions, our model harnesses speech signals and demographic data to accurately estimate blood pressure. We found a strong correlation between our predictive model and early-stage high blood pressure, highlighting its potential for early detection. Central to our investigation is the Graph Diffusion Network (GDN) model, achieving exceptional performance with an R² score of 0.96 and a Pearson correlation coefficient (PCC) of 0.98. In early-stage hypertension detection, the GDN model achieved an F1-Score of 0.8735 ± 0.10 and accuracy of 0.8896 ± 0.11. Additionally, without considering demographic features, the model still performed well, with an R² of 0.740 and PCC of 0.764 when used alone. These results emphasize the value of combining speech and demographic features, offering a promising, non-invasive solution for blood pressure monitoring.