基于社交媒体中文本信息的早期抑郁症检测

诊断抑郁症的传统方法是通过面对面的评估和交谈.但是,许多患有抑郁症的患者不愿意在早期阶段就医,从而使病情恶化.为了在早期判断抑郁症患者的情况,提出一种利用社交媒体文本信息的时间序列特征和多示例学习的检测模型,考虑到抑郁症状不会立即出现,所以时序样本的使用显得非常重要,因此使用无监督LSTM提取时间序列特征,训练分类器实...

Early Detection of Depression Based on Textual Information in Social Media

The traditional method of diagnosing depression is through face-to-face assessment and conversation. However, many patients with depression are reluctant to seek medical attention at an early stage, which makes their condition worse. In order to judge the situation of patients with depression in the early stage, a detection model using time series features of social media textual information and multi-instance learning was proposed in this work. Considering that depressive symptoms will not appear immediately, the use of time series samples will be very important. Therefore, the unsupervised LSTM was used to extract time series features, binary classification was implemented by training a classifier, and a multi-instance learning model was exploited to solve the problem of unbalanced samples. Naive Bayes classifiers, random forests, multivariate social network learning and multimodal depression dictionary learning were used as benchmark methods firstly, and then the multi-instance learning with unsupervised LSTM time series functions was employed to detect depression more accurately. On the basis of the MDDL dataset, 200 survey subjects totally 7946 tweets were selected, and the training-test ratio was set as 8:2. Experimental results were following: the accuracy, precision, recall and F1 score reached 75.0%, 76.0%, 73.0%, and 74.5%, respectively, which demonstrated that it was feasible to use machine learning for early depression detection through text data in social media. In addition, a large number of ablation studies also verified that the method using time series features could achieve better performance than the traditional benchmark methods.