Vocal Fundamental Frequency and Generalized Anxiety Disorder: Insights from a Large Multi-Country Cohort in Africa

Voice-based biomarkers have been proposed as objective tools for anxiety assessment, yet existing research relies on small, demographically homogeneous samples and shows inconsistent associations between fundamental frequency (F₀) and anxiety. Using a large, diverse, multi-country African cohort, we evaluated whether F₀ measures extracted from uncontrolled audio recordings are associated with generalized anxiety beyond demographic and self-reported health factors.

Speech recordings and survey data were collected from 1788 adults in Botswana, Ghana, Nigeria, and Tanzania using a web-based platform. Participants completed neutral readings of a short sentence and a connected-speech sample and completed the generalized anxiety disorder 7-item (GAD-7) questionnaire. Additionally, self-reported age and health data were collected. F₀ statistical measures (mean, standard deviation, minimum, maximum) were extracted from audio data using standardized acoustic procedures. Associations with high-anxiety status were analyzed using generalized estimating equations logistic models to account for within-participant clustering, with L1-penalized logistic regression as a robustness check. As a complementary nonlinear assessment, an extreme gradient boosting classifier was fit and interpreted using SHapley Additive exPlanations.

Across all models, better self-reported health and older age were associated with lower odds of high anxiety, whereas female sex and country differences (highest in Botswana and Ghana) were associated with higher odds. In contrast, F₀ features showed weak associations and did not improve discrimination (area under the receiver operating characteristic curve ≈ 0.70). In the machine learning analysis, health status and age dominated predictive contributions, and among acoustic variables, F₀ min ranked highest but remained modest relative to demographic and health factors.

F₀ features derived from uncontrolled audio data provide limited explanatory or predictive value for GAD assessment. These findings highlight the importance of recording conditions and underscore the need for richer acoustic and linguistic features and potentially stress-inducing protocols in future work.