| 摘 要: 抑郁症在大学生群体中尤为高发，利用移动端收集每日情绪和日常行为数据进行抑郁评估成为重要研究方向。现有机器学习的方法大多仅简单整合数据，忽略了不同数据间的内在关联，也难以捕获长时序依赖关系；还未充分考虑其他精神障碍因素对抑郁症状的影响，为此提出了一种周期性划分驱动的时空图卷积网络（Per-STGCN）。通过傅里叶变换计算时序数据的周期，将长时序数据划分为不同周期序列，通过捕获周期内及周期间关系以建模长时序依赖。在周期内利用共享时空图卷积网络（ST-GCN）捕捉特征的动态演变规律；周期间则通过时序卷积网络（TCN）挖掘时序相关性，从而提取出富含时空信息的多周期关联特征。随后设计了压力、孤独感专用编码器对多周期关联特征进一步处理，借助对应标签用监督模型学习富含压力和孤独感信息的相关特征，再融合日常行为特征进行抑郁相关特征提取以实现抑郁评估。实验结果表明，Per-STGCN模型在StudentLife数据集上的抑郁评估性能优于现有方法，均方根误差（RMSE）降低至0.09。 |
| 关键词: 抑郁症评估 ST-GCN 周期性 多标签学习 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金资助项目（62236006，62306172，62376215）；陕西省重点研发计划资助项目(2025CY-YBXM-191) |
|
| Depression Assessment Method Based on Periodic Partitioning Driven Spatio-Temporal Graph |
|
chenhaifeng1,2, changyake1,2, wusiyan1,2, lijian1,2
|
1.Shanxi University of Science &2.Technology
|
| Abstract: Depression is particularly prevalent among college students, making the use of mobile devices to collect daily mood and behavioral data for depression detection an important research direction. Most existing machine learning methods simply integrate data without considering the intrinsic relationships between different types of data, and they struggle to capture long-term temporal dependencies. Additionally, they often fail to adequately account for the influence of other mental disorder factors on depressive symptoms. To address these issues, a Periodic Partition driven Spatio-Temporal Graph Convolutional Network (Per-STGCN) is proposed. The periodicity of the time-series data is calculated using Fourier transform, dividing long-term sequences into different periodic segments. This allows modeling long-term dependencies by capturing intra-period and inter-period relationships. Within each period, a shared Spatio-Temporal Graph Convolutional Network (ST-GCN) is used to capture the dynamic evolution patterns of features, while a Temporal Convolutional Network (TCN) is employed to mine temporal correlations across periods, thereby extracting multi-period associative features rich in spatio-temporal information. Subsequently, dedicated encoders for stress and loneliness are designed to further process the multi-period associative features. Using corresponding labels, a supervised model learns relevant features enriched with stress and loneliness information. These are then fused with daily behavioral features to extract depression-related characteristics for depression detection. Experimental results show that the Per-STGCN model outperforms existing methods in depression detection on the StudentLife dataset, reducing the Root Mean Square Error (RMSE) to 0.09. |
| Keywords: Depression assessment ST-GCN Periodic Multi label learning |
