Speech-derived acoustic biomarkers for depression: Comprehensive cross-section and longitudinal analyses in different...

Speech encodes emotional, cognitive, and motor states, offering an objective, non-invasive window into mental health. In major depressive disorder (MDD), vocal alterations are reported, yet their cross-dataset reproducibility, symptom specificity, and longitudinal stability remain uncertain, especially under naturalistic, content-variable speech tasks. We conducted a large, multi-cohort study of 1857 participants spanning a primary discovery dataset, an independent secondary clinical dataset, and an 8-week longitudinal follow-up. From standardized recordings we extracted 6373 acoustic features and examined (i) baseline case-control screening in each dataset, (ii) severity-related feature patterns across depression levels, (iii) symptom-dimension markers using stability-enhanced elastic net, and (iv) longitudinal feature changes over 8 weeks, integrating unsupervised clustering and mediation analyses, with false discovery rate control. Across cohorts, correlation-based redundancy reduction yielded a compact final set of 23 non-redundant representative features for cross-cohort reporting and interpretation. Symptom-factor analysis identified distinct, non-overlapping feature sets for HAMD-24 dimensions, with somatic and depressed mood yielding the most stable markers. Longitudinally, 38 features exhibited heterogeneous recovery trajectories and mediation patterns consistent with symptom improvement, with spectral-shape and modulation markers showing higher temporal sensitivity than energy and voice-quality features. Overall, our findings indicate that a compact set of speech-derived markers can support symptom-informed monitoring in MDD. A small subset of acoustic features is robust, symptom-specific, and temporally informative, refining assumptions of uniform vocal change and enabling targeted, symptom-informed speech biomarkers for personalized monitoring and early intervention. Future work should verify these markers using task-matched speech prompts and alternative feature representations, given potential content-related confounding in free-response speech and reported reliability limitations of some high-dimensional acoustic feature toolkits. TRIAL REGISTRATION: ChiCTR2500095151.