Multimedia Systems Компьютерные сети и коммуникации
Adaptive peer-referenced multimodal learning for robust depression assessment under population heterogeneity
Аннотация
Automated depression assessment using multimodal behavioral signals has shown promising potential for scalable and objective screening. However, achieving reliable performance remains challenging under population heterogeneity, primarily due to two intertwined issues: substantial inter-individual variability in behavioral expression patterns across demographic groups and modality imbalance, where dominant modalities may overshadow subtle yet informative signals from other channels. To address these challenges, we propose Adaptive Peer-Referenced Multimodal Learning, a framework that explicitly models fine-grained population structure for depression severity estimation. Specifically, we introduce a demographically adaptive graph convolutional network (DA-GCN) that constructs individualized peer-reference graphs by integrating behavioral similarity and demographic comparability through learnable weighting mechanisms. Furthermore, we incorporate a RollingQ dynamic attention mechanism to dynamically adjust multimodal contributions during fusion. Experiments on DAIC-WOZ and E-DAIC demonstrate the effectiveness of the proposed framework. On DAIC-WOZ, our method achieves a Concordance Correlation Coefficient (CCC) of 0.690 and a Mean Absolute Error (MAE) of 3.42, achieving the highest CCC and lowest MAE among the compared methods. On E-DAIC, the proposed method achieves a CCC of 0.670 and an MAE of 3.95, demonstrating effective depression severity estimation under a more heterogeneous evaluation setting.