| 摘 要: 抑郁症在大学生群体中高发,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周期,将长序列划分为多个周期片段,分别建模周期内与周期间的时空依赖,从而提取具有多周期关联的融合特征。进一步设计专用编码器分别学习压力与孤独感的相关特征,并结合日常行为特征进行多源信息融合,以提升抑郁评估的准确性。实验结果表明,Per-STGCN在StudentLife数据集上的抑郁评估性能优于对比方法,均方根误差(RMSE)降低至0.09。 |
| 关键词: 抑郁症评估 时空图卷积网络 周期性划分 |
|
中图分类号: TP399
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(62306172);陕西省重点研发计划项目资助(2025CY-YBXM-191) |
|
| Depression Assessment Method Based on Periodic Partitioning Driven Spatio-Temporal Graph |
|
CHEN Haifeng, CHANG Yake, WU Siyan, LI Jian
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi’an 710021, China)
chenhaifeng@sust.edu.cn; changyak@163.com; 2028431249@qq.com; lijianjsj@sust.edu.cn
|
| Abstract: Depression is highly prevalent among college students, making the use of mobile devices to collect daily mood and behavioral data for depression assessment an important research direction. Most existing machine learning methods simply integrate data without fully capturing the intrinsic relationships between different types of data or modeling long-term temporal dependencies. Moreover, they often overlook the influence of other mental health factors, such as stress and loneliness, on depressive symptoms. To address these issues, this study proposes a Periodic partitioning-driven Spatio-Temporal Graph Convolutional Network (Pe-r STGCN). The method first applies Fourier transform to identify temporal cycles, dividing long sequences into multiple periodic segments. It then models both intra-cycle and inte-r cycle spatio-temporal dependencies to extract fused features with mult-i period associations. Furthermore, dedicated encoders are designed to learn features related to stress and loneliness, which are integrated with daily behavioral features through mult-i source fusion to enhance the accuracy of depression assessment. Experimental results on the StudentLife dataset demonstrate that Pe-r STGCN outperforms comparative methods, achieving a Root Mean Square Error (RMSE) as low as 0.09. |
| Keywords: depression assessment spatio-temporal graph convolutional network periodic partition |
