Sequence-to-Graph Learning for Depression Detection with Clinical Interview Video Recording

Clinical interview serves as an essential and effective tool for depression recognition. However, traditional interview procedures rely on clinicians, rendering the process subjective and time-consuming. Although automatic depression detection (ADD) methods have been developed, their performance is limited by the long and complex nature of interview videos, where spatiotemporal cues are irregularly distributed. In this paper, we propose a Sequence-to-Graph Network (S2GNet), which jointly models sequential progression and spatiotemporal contextual correlations in interview recordings to establish discriminative depression representations. Specifically, we first design a facial graph (FG) module that captures dynamic interactions among key facial regions to enhance behavior pattern modeling. Based on facial embeddings, we further propose two collaborative graph paradigms for better video representation: a temporal chain (TC), which com poses local temporal patterns within each question-wise clip and integrates them into a global evolutionary chain of depression-specific signals; and a hierarchy-guided graph (HG), where each node aggregates question/topic-relative depression cues within the graph based on the interview structure. Notably, we collect a self-constructed dataset including start/end timestamps of symptom-related questions, interview structure, and labels provided by psychologists from 138 subjects. Comprehensive experiments on the AVEC2013, DAIC-WOZ, LMVD and a newly constructed clinical interview dataset demonstrate the effectiveness and robustness of S2GNet.