Optimizing speech elicitation tasks for machine learning-based depression assessment

Abstract

THEORETICAL BACKGROUND: The field of machine learning-based speech analysis may provide unobtrusive, time-efficient and cost-effective ways of automated depression assessment. Systematically optimizing speech elicitation tasks may further improve the accuracy of this approach. We hypothesized that machine learning-based depression classification would perform better if trained on recordings of individuals reading anti-depressive statements with the instruction to intone them as convincingly as possible compared to readings of anti-depressive statements without instructions regarding intonation.METHODS: To test this hypothesis, we recruited a sample of 48 clinically depressed individuals, 48 sub-clinically depressed individuals, and 48 non-depressed individuals. Participants from each group were randomly allocated to either the experimental or the control condition. In both conditions, participants read aloud scripted anti-depressive self-statements. Participants in the experimental condition received instructions to heighten the prosodic expression of conviction in their voice, whereas participants in the control condition received no such instructions. Separate classification models aimed at detecting current depression were trained for each condition and with a selection of different machine learning methods. RESULTS: We found that models trained on data from the experimental condition were more accurate and reliable than those trained on data from the control condition. While the former models reached balanced accuracies between 65–76%, the latter only reached balanced accuracies between 36–61%.DISCUSSION: Our results suggest that features of speech elicitation tasks have substantial influence on model performance for automated depression classification. The present findings highlight that speech elicitation tasks including voice modulation instructions can improve the validity and reliability of machine learning-based depression classification.

Keywords: Depression Assessment, Machine Learning, Voice, Speech, Major Depressive Disorder

DOI: 10.54941/ahfe1005694