Using valence and arousal scoring in speech to detect depressive symptoms: A study on family caregivers.

BACKGROUND: Depression is a prevalent mental health condition that significantly affects individuals' well-being and daily functioning, making early depressive symptom detection crucial. OBJECTIVES: This study explores the potential of advancing speech analytics with dimensional sentiment analysis, using valence and arousal scoring to enhance depressive symptom detection in family caregivers. METHODS: We recruited 116 family caregivers and assessed their depression levels using the Beck Depression Inventory-II. Participants provided verbal responses to 12 open-ended questions related to family resilience. The extracted words from the responses were analyzed for sentiment using the Chinese Valence-Arousal Words dictionary to determine their valence and arousal ratings. For each participant, we computed various summary statistics (e.g., mean, variance, quantiles) and frequency-domain statistics (e.g., spectral entropy, dominant frequency components) of valence and arousal distributions. These features were then used as predictors in nine different machine learning models evaluated under repeated 5-fold nested cross-validation with cost-sensitive cutoff selection, enabling robust classification of depression levels while prioritizing the identification of true positive cases. RESULTS: Our analysis showed that participants with higher depressive symptoms used words with more extreme valence and arousal scores than those with lower depressive symptoms. The models demonstrated strong predictive utility of these emotional dimensions: the quadratic discriminant analysis (QDA) model achieved an out-of-sample sensitivity of 0.725, a specificity of 0.563, and an AUC of 0.7. These findings support that valence and arousal analysis provides meaningful discriminative power for identifying depressive symptoms. Importantly, the high sensitivity achieved by these models ensures that individuals with elevated symptoms are more reliably detected, thereby reducing the risk of missed cases. Overall, the results suggest strong potential for automated depressive symptom detection based on emotional signal analysis, with cost-sensitive optimization reducing false negatives while maintaining clinically relevant specificity. CONCLUSIONS: This study offers proof-of-concept evidence that emotional sentiment analysis of speech using valence and arousal scoring can capture linguistic markers associated with depressive symptoms among family caregivers. With further refinement and validation, this approach lays an important foundation for the development of more robust machine learning models and ultimately contributes to non-intrusive mental health screening tools.