Three-dimensional face reconstruction for emotional dynamics: a novel approach to distinguishing bipolar disorder from major depressive disorder.

Abstract

Major depressive disorder (MDD) and bipolar disorder (BD) are frequently misdiagnosed in clinical practice due to their overlapping symptoms, resulting in delayed treatment and an increased burden on patients. Reliable biomarkers for early differential diagnosis are currently lacking. To address this, we developed a clinical facial video dataset of patients with BD and MDD and introduced an adaptive-shape vision-to-language model (AsmV2L), a vision-language-guided fine-grained three-dimensional (3D) face reconstruction framework that better preserves affect-related facial cues from monocular videos. Using valence-arousal (VA) representations estimated from these reconstructed 3D faces, we identified distinct affective patterns that differentiated the two disorders. In the subject-level downstream evaluation, the emotional dynamic features extracted by our framework achieved 88.9% accuracy in distinguishing BD from MDD. Additionally, experiments on public benchmark datasets demonstrated that AsmV2L maintained competitive 3D reconstruction accuracy while preserving emotion-related semantic information more effectively.