Biomarker potential of real-world voice signals to predict abnormal blood glucose levels

Background: Voice signal analysis is an emerging noninvasive technique to examine health conditions, and is implemented in various real life applications and devices. The purpose of this study was to evaluate the association of voice signals with blood glucose levels in healthy individuals. The study aimed to investigate the longitudinal stabilities of voice signals and identify voice biomarkers to predict abnormal blood glucose levels. Methods: We created voice profiles composed of 17,552,688 voice signals from 44 participants and their 1,454 voice recordings. From each voice recording, 12,082 voice features were extracted. Longitudinal stabilities of voice-features were quantified using linear mixed effect modelling. Voice-features that showed significant difference between different blood glucose levels, strong intra-stability and the ability to make distinct choice in decision trees were selected as voice biomarker. Voice biomarkers were fed into a multi class random forest classifier to predict high, normal, and low blood glucose levels. Findings: In total, 196 voice biomarkers were characterized. Results showed a predictive model with an overall accuracy of 78.66%, overall AUC of 0.83 (95% confidence interval is 0.80 to 0.85), and 0.41 of Matthews Correlation Coefficient (MCC) to discriminate three different blood glucose levels in an independent test set. Interpretation: Our voice biomarkers could serve as a noninvasive and conventional surrogate of blood glucose monitoring in daily life as well as a screening tool to estimate potential risk of poor glycemic control.