Padrões vocais que diferenciam pessoas com e sem ansiedade

Abstract

Objective: To verify whether there are vocal and physiological variables that are capable of predicting anxiety symptoms, as well as to test the performance of global and specific models that can predict the presence of high anxiety from vocal and physiological markers. Method: This cross-sectional study involved 38 volunteers from the academic community, divided into low and high anxiety groups, based on the STAI. The collection included physiological (HR, SpO₂, temperature, skin conductance and EMG) and vocal (self-assessment, auditory-perceptual assessment and acoustic analysis) measurements. The data were analyzed through supervised machine learning, from decision trees, developing models with different compositions of variables to identify vocal and physiological markers capable of predicting high levels of anxiety. Results: The global model that combined physiological and vocal variables demonstrated robust performance and with more branches, but it has limitations in clinical routine. Models consisting of vocal variables, whether subjective or objective, have the potential to differentiate groups with low and high levels of anxiety. Isolated analysis of vocal dimensions also demonstrated predictive capacity, with emphasis on self-assessment and acoustic analysis measures. Conclusion: Supervised machine learning models indicate that physiological and vocal measures are predictors of anxiety, although they require specialized assessment. Among the isolated methods, the model composed only of acoustic measures presented the best performance, standing out as the most promising for predicting anxiety, a fact that reinforces voice as a digital biomarker.