QQ客服 官方微博 反馈留言

题名	基于二维步态数据的抑郁和焦虑检测
作者	敬春柯
答辩日期	2021-01
文献子类	硕士
授予单位	中国科学院心理研究所
授予地点	中国科学院心理研究所
其他责任者	朱廷劭
关键词	步态 语音 抑郁 焦虑 检测
学位名称	理学硕士（同等学力硕士）
其他题名	Detection of depression and anxiety based on two-dimensional gait data
学位专业	应用心理学
中文摘要	Depression and anxiety are the most common psychological and mental illnesses in the world. They not only seriously affect patients’ quality of life, work and interpersonal relationships, but also are important factors inducing suicide and bring a heavy burden of disease to the country and society. For a long time, depression and anxiety have mainly used subjective methods such as self-rating or others-assessment scales. These methods have been affected by uncertain factors such as the patient's feeling and expression ability, lying and social praise tendencies, and the evaluator's professional training level. In view of the significant differences in gait characteristics between depressed/anxious patients and healthy people, the use of machine learning methods combined with gait characteristics for objective and automatic diagnosis of depression and anxiety has become an effective complement to subjective diagnostic methods. Two-dimensional gait data has the more advantages in the application market such as convenient collection, low cost, and small amount of calculation. In this article, we mainly achieve the automatic recognition of depression and anxiety by using machine learning models based the two-dimensional gait data.The pre-study used two-dimensional gait data to verify its ability to predict age, gender and big five personality. In the prediction of gender, the MLP method was used to achieve the highest classification accuracy rate of 0.789; in the regression prediction of age, the MLP reached a correlation coefficient of 0.58; in the regression prediction of the big five personality, the best correlation coefficients of the five personality dimensions reached were 0.55 (extroverted), 0.52 (pleasant), 0.52 (conscientiousness), 0.56 (nervous) and 0.52 (openness). In the study of depression identification, we established some regression models based on the scores of the PHQ-9 and CES-D scales and the two-dimensional gait data, the best correlation coefficients of these regression models reached are 0.54 and 0.52 which are closed to the strong correlation level of prediction ability. Similarly, in the study of anxiety, we chose the scores of the anxiety scale GAD-7 and T-AI to establish some regression models, the best correlation coefficients of these regression models reached are 0.52 , which reached a prediction level above medium to near strong correlation. At the same time, we grouped the subjects according to gender and age, and used MLP to train different groups separately, and the results were relatively improved compared with not-grouping. Among them, the best PCC of the gender and age group all reached above 0.80, also the data showed that the male group was slightly better than the female group, and the older group over 22 years old was slightly better than younger group under 22 years old, these findings indicate that gender and age have a greater influence on the performance of the gait model. Finally, we used multimodal data which are gait and speech data to perform feature- level fusion and decision-level fusion based on Loss values and PCC coefficients on the MLP model. The fusion effect is slightly improved relative to the performance of single- source features, but not obvious.The research in this paper not only proves that the two-dimensional gait data can effectively identify age, gender and big five personality, but also can effectively detects depression and anxiety. Meanwhile it also verifies the influence of gender and age on the gait model. All these findings will promote the development of related application market and industries.
英文摘要	抑郁和焦虑是全球最普遍的心理和精神类疾病，不仅严重影响患者个体的生命 质量、工作和人际关系，同时也是诱发自杀的重要因素，给社会和国家带来了沉重 的疾病负担。长期以来，抑郁和焦虑的诊断主要是采用自评或他评量表等偏主观方 法，但这些方法受到患者的感知和表达能力、撒谎和社会赞许性倾向以及评价者的 个人训练水平等主观因素的影响，所以存在一定的不足。鉴于抑郁和焦虑患者在步 态特征中存在显著差异，所以基于客观的步态数据进行抑郁和焦虑的自动识别成为 可能，并将成为主观诊断方法的有效补充。 二维步态数据具有采集便利、成本低廉及计算量较小的优点，在实际应用中更 有优势。本文通过对二维步态数据的分析，利用机器学习模型，实现对抑郁和焦虑 的自动识别，并对模型的有效性进行检测。 预研究使用二维步态数据检验其对年龄、性别以及大五人格的预测能力。在对 性别的预测上，使用 MLP 方法实现了最高 0.789 的分类正确率;在对年龄的回归预 测中，皮尔逊相关系数(PCC)最高达到 0.58;在对大五人格的回归预测中，五个 人格维度的相关系数最好分别达到了 0.55(外向性)、0.52(宜人性)、0.52(尽责 性)、0.56(神经质)和 0.52(开放性)。预研究的结论体现了二维步态对年龄、性 别和大五人格具有较好的预测能力。 在对抑郁的研究中，我们基于二维步态数据对抑郁量表 PHQ-9 和 CES-D 的得 分建立多种回归模型，得到最好的相关系数分别为 0.54 和 0.52，接近强相关的预测 能力;类似的，焦虑的研究使用了 GAD-7 和 T-AI 两个焦虑量表的得分建立回归模 型，步态模型最终达到的相关系数最高达到 0.52，达到了中等以上接近强相关的预 测水平。同时，我们按照性别和年龄对被试进行分组，使用 MLP 对不同组别进行单 独训练，发现其结果相对不分组有较大的改善，分组后的模型在四个量表中的最大 的 PCC 都达到 0.80 以上，并且呈现出男性组略好于女性组，大龄组(大于 22 岁) 表现好于低龄组(22 岁以下)的特点，说明了性别和年龄对步态模型表现有较大的 影响。 最后，我们使用步态和语音两种模态数据，并且在 MLP 模型上进行了特征级融 合、基于 Loss 值和 PCC 系数的决策级融合，融合效果相对单源特征的表现略有改 善，但并不明显。本文通过建立机器学习模型证明二维步态数据可以有效的识别年龄、性别和大 五人格，也证明可以有效的进行抑郁和焦虑的检测，同时揭示了性别和年龄因素对 步态模型的影响，初步探索了和语音的多源数据融合，对应用市场开发及推广此类 应用具有一定的现实意义。
语种	中文
内容类型	学位论文
源URL	[http://ir.psych.ac.cn/handle/311026/41640]
专题	心理研究所_社会与工程心理学研究室
推荐引用方式 GB/T 7714	敬春柯. 基于二维步态数据的抑郁和焦虑检测[D]. 中国科学院心理研究所. 中国科学院心理研究所. 2021.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。