社会与工程心理学研究室知识库

抑郁症是世界范围内常见的精神疾病之一,抑郁症患者往往长期伴随情绪低落,如悲伤内疚、低自尊、兴趣丧失、功能减退等,对个人、家庭及社会造成了巨大损失.抑郁症的发病原因复杂,临床诊断存在一定的困难,有必要寻找一种更加便捷、客观、高效的方式来辅助抑郁症的快速识别.语音作为一个相对客观且容易获得的变量,具有其潜在的价值.本研究旨在构建基于语音的抑郁症识别模型,探究语音与抑郁症之间的关系.收集了103名被试(45名抑郁症患者,58名健康人)的语音数据,实验组为临床确诊的抑郁症患者,年龄在23.8~44.6岁之间,控制组为健康人,年龄为20.1~41.7岁.我们采用了3(情绪状态:正性、中性、负性)×3(任务类型:语言问答、文本朗读、图片描述)的实验设计,运用机器学习的分类算法——逻辑回归(LR)来构建抑郁识别模型.实验结果表明,语音的抑郁识别精度可以达到82.9%.本文采用机器学习方法,基于语音变量建立有效的抑郁症自动识别模型,为抑郁症的辅助识别提供客观的指标和依据.

Depression is one of the common mental diseases. Patients with depression often have depressed moods such as sadness, guilty, low self-esteem, loss of interest, hypofunction and so on. They suffer from serious emotional problems, unexplained suffering, which has caused enormous losses to individuals, families and society. According to the World Health Organization, there are aproximately 322 million people suffering from depression in the whole world in 2017. While there are about 54 million depressive patients in China.

Depression can be cured effciently. However, due to the complexity of the pathogenesis of depression, clinical diagnosis is accompanied with many difficulties. Firstly, the mental disease, especially depression, are not getting enough attention and even being misinterpreted by other people. Secondly, the depression patients are less willing to ask for help. Thirdly, it is hard to select and dignose the potential depression patients precisely, as well as there are limited medical resource for depression diagnosis.

It is necessary to find a more convenient, objective and efficient way to assist the fast identification of depression. As a relatively objective and easily accessible variable, speech has its potential value. The speech of patient is easy to acquire, and also, it has been proved that the sound of depressed patients have special charcteristics such as slow speech rate, lack of cadence and so on. The purpose of this paper is to explore the relationship between speech and depression by establishing classification models of voice feature and depression prediction. In this research, 3(emotion mood: positive, neutral, negative)×3(task type: question answering, text reading, picture description) experimental design was employed, and the voice data was collected from the speech of individuals recorded during different tasks. 103 participants were inculded in this study, including 45 depression patients (age: 23.8–44.6, M=34.2, SD=10.4, males=22, females=23) and 58 healthy ones (age: 20.1–41.7, M=30.9, SD=10.8, males=27, females=31). The former were recruited in the hospital in Beijing Anding Hospital and Huilongguan Hospital, while the latter were recruited by advertisement. All of them were diagnosed by specialist with DSM-IV and MINI interview. All participants did not have substance abuse, substance dependence, personality disorders and other mental diseases, no serious physical illness or suicidal behavior. The education level of subjects are all above the elementary school. 988 Voice features were extracted from the speech data using open SMILE software. Logistic regression, a machine learning method, was used to train the predicting models. Results showed that the precision rate of predicting can reach to 82.9%. Based on machine learning methods, this paper employed voice features to establish predicting models of depression. Results show the speech of depression patients has certain predicting effect, which paves the way for the further identification of depression in a more thorough way.