Speech-based differentiation of unipolar and bipolar depression across scripted reading and interview questions

Differentiating bipolar depression (BPD) from unipolar depression (UPD) remains clinically challenging. We examined whether acoustic and linguistic speech features could support this distinction and whether their informativeness varied across scripted reading and interview questions.

This study included 109 healthy controls (HCs), 52 patients with UPD, and 53 patients with BPD experiencing depressive episodes. Participants completed a scripted reading task and an interview-based speech task comprising 16 questions. Acoustic features were extracted from all audio recordings, and linguistic features were derived from interview transcripts. Group differences were examined using robust analysis of covariance with global false discovery rate correction. Two-level stacking models with fold-wise sex adjustment of acoustic features evaluated classification performance and importance of speech features and interview-based speech elicitation prompts.

Multiple acoustic and linguistic features showed significant group differences in interview responses, whereas no acoustic features differed significantly during scripted reading. The best-performing models achieved F1 scores of 0.6655 for three-class classification (HC, UPD and BPD) and 0.6430 for binary classification (UPD and BPD). Within interview-based speech, statistical and machine learning analyses identified Average Pause Duration, Phonation Ratio, and quantity-related expressions as particularly informative features. Two interview questions about (hypo-)manic experiences contributed most to differentiating BPD from UPD.

Acoustic and linguistic speech features may help differentiate BPD from UPD during depressive episodes, with their informativeness varying across speech tasks and interview questions. Our findings suggest that targeted interview-based speech assessment may have potential as a non-invasive adjunct to conventional clinical evaluation.