Interpretable speech biomarkers of psychological resilience in major depressive disorder via expert-guided large language models

Abstract

Abstract Background Major depressive disorder (MDD) is highly prevalent and recurrent, but current clinical care focuses mainly on symptom reduction, overlooking the individual’s capacity for psychological resilience which is a key factor for long-term recovery. Traditional resilience assessments rely on subjective self-reports, lacking objective, scalable alternatives. Methods This study developed a Multi-Topic Emotion Interview (MTEI) paradigm and collected multimodal data from 248 MDD patients and 101 healthy controls between December 2024 and November 2025. A 4-week longitudinal follow-up ( n = 106) was conducted to evaluate model generalizability. We introduced an expert-guided Large Language Model (LLM) to extract deep Semantic Resilience Index (SRI) and three-dimensional emotional features (VAD). A Voting Regressor ensemble was employed to integrate these LLM-derived features with momentary Self-Assessment Manikin (SAM) state assessments to predict continuous Connor-Davidson Resilience Scale (CD-RISC) scores. Results The optimal multimodal fusion model (SRI + VAD + SAM) demonstrated favorable predictive efficacy ( R 2 = 0.555, Concordance Correlation Coefficient [CCC] = 0.705) and maintained high stability in the temporal validation set ( R 2 = 0.557, CCC = 0.701). Crucially, residual analysis verified that the model captures specific resilience traits independent of depression severity, explaining 9.6% of the variance in pure resilience residuals. SHAP analysis revealed that LLM-derived agency & control, cognitive nuance, and contextual emotional dominance serve as the most critical predictors. Traditional acoustic features failed to provide independent incremental validity during late fusion. Conclusions Integrating expert-guided LLMs for deep semantic parsing with momentary emotional assessments offers a promising, interpretable digital biomarker for psychological resilience. This framework successfully decodes cognitive coping strategies independent of depressive mood, providing potentially actionable targets for personalized psychiatric interventions. Registered institution Chinese Clinical Trial Registry, ChiCTR2500100966, registration date: April 17, 2025.

The authors' abstract, as published at the source. BMC Psychiatry, 2026 · DOI ↗