Early depression detection can enable timely intervention. Automatic depression detection has relied on features extracted from individual-level data, which may be too coarse to support effective detection. Existing detection models have largely overlooked interview questions commonly used in clinical depression assessment. This research proposes a two-layered multi-modal model for depression detection, which not only extracts features from responses at a level of individual interview questions, but also identifies semantic categories of those questions. The evaluation results demonstrate that the proposed model outperforms the state-of-the-art methods for depression detection. The research findings have broad and cross-disciplinary implications.

早期发现抑郁症可以及时进行干预。自动压低检测依赖于从单个级别的数据中提取的特征，这些特征可能太粗糙而无法支持有效的检测。现有的检测模型在很大程度上忽略了临床抑郁评估中常用的访谈问题。这项研究提出了一种用于抑郁症检测的两层多模式模型，该模型不仅可以从单个访谈问题级别的响应中提取特征，还可以识别这些问题的语义类别。评估结果表明，提出的模型优于用于抑郁症检测的最新方法。研究结果具有广泛和跨学科的含义。