Abstract
BACKGROUND: Depression affects more than 300 million people worldwide and is a leading contributor to the global disease burden. Traditional diagnostic methods, such as structured clinical interviews, are reliable but impractical for frequent or large-scale screening. Self-report tools like the Patient Health Questionnaire-8 (PHQ-8) require disclosure and clinician oversight, limiting accessibility. Recent artificial intelligence-based approaches leverage multimodal behavioral cues (linguistic, acoustic, and visual) for automated depression detection but remain constrained by limited adaptability, scarce annotated data, weak emotional expression in real-world settings, and the high computational cost of deployment of socially assistive robots (SARs). OBJECTIVE: This study introduces Depression Social Assistant Robot (DEPRESAR)-Fusion, a lightweight multimodal depression detection framework designed for natural interactions with emotion-aware SARs. The objective of this study was to enhance detection accuracy in everyday conversations while addressing the challenges of data scarcity, weak emotional cues, and computational efficiency. METHODS: DEPRESAR-Fusion integrates acoustic, linguistic, and visual features with an emotion-aware response module powered by large language models to adapt conversational strategies dynamically. To stimulate richer emotional expression, participants were exposed to emotionally evocative videos before SAR interactions. To overcome data scarcity, we augmented training with (1) public depression-related social media corpora and (2) synthetic samples generated via large language models. The proposed multimodal fusion architecture was evaluated on benchmark clinical datasets for both binary depression classification and PHQ-8 regression tasks. Performance was compared against prior multimodal baselines using root mean square error, mean absolute error, and standard classification metrics. RESULTS: Participants who viewed emotional stimuli before interacting with SARs exhibited significantly higher emotional expressiveness, leading to improved model performance. Regression tasks showed lower root mean square error and mean absolute error, while classification tasks achieved significantly higher accuracy than the nonstimulus condition. DEPRESAR-Fusion outperformed prior multimodal baselines across multiple benchmark datasets, achieving state-of-the-art performance in both binary classification and PHQ-8 regression. The system maintained a lightweight architecture suitable for real-time deployment on SARs. CONCLUSIONS: DEPRESAR-Fusion demonstrates that integrating emotion induction, data augmentation, and lightweight multimodal fusion can enable accurate and scalable depression detection in naturalistic SAR interactions. By bridging the gap between structured clinical assessments and everyday conversations, this approach highlights the potential of SAR-based systems as nonintrusive, artificial intelligence-driven tools for proactive mental health support.