社会与工程心理学研究室知识库

抑郁症是一种以抑郁情绪为核心并伴随多种症状的严重心理疾病。找到能够有效反应抑郁症疾病状态的客观指标有利于辅助抑郁症的诊断。作为一种常见易得的行为线索，已有研究表明抑郁被试的语音和抑郁症状存在显著相关，并且语音能够用于区分抑郁症和健康人。但是，通过语音自动检测抑郁症的方法当前还有两个重要问题有待探究：一是语音识别抑郁症的诊断效力问题；二是语音的区分效果是否具备跨情境的稳定性问题。诊断效力问题将通过逐层加深区分难度的方式，利用语音分别区分抑郁症与不同症状相似程度的人群，考察最终可以利用语音区分抑郁和哪种类型的人群。研究中利用机器学习中的分类算法来考察语音特征识别抑郁症的效果，通过分类算法中的如F值等指标评价模型预测结果的好坏。跨情境稳定性问题将通过比较不同实验情境、任务和情绪下的语音的预测结果来探讨语音的区分效果是否具备跨情境的稳定性。
研究一考察语音纳入疾病之能力，旨在利用语音区分抑郁症和非心理疾患，包括3个实验。实验1为重复测量设计，比较了39名健康人、26名双相障碍和46名抑郁患者在不同情绪诱导下的情绪主观体验之间的差异。该实验的结果说明通过自陈提供的情绪信息无法区分抑郁症、双相障碍和健康人。实验2为重复测量设计，收集了58名健康人和45名抑郁患者在不同任务和情绪下的语音用于分类预测。结果显示利用语音区分抑郁症患者和健康人的F值能达到78.2%的中等预测水平，且在不同任务和情绪下的预测结果都达到了中等预测水平。实验3利用语音特征区分759名抑郁患者和996名躯体疾病患者，结果显示语音区分抑郁症和躯体疾患的F值能达到75.6%。研究一的结果表明，语音特征能够有效的用于区分抑郁症患者和健康人、抑郁症患者和躯体疾病患者，且区分效果具备跨情境的稳定性。
研究二考察语音排除疾病之能力，旨在利用语音区分抑郁症和其它心理疾患，包括2个实验。实验4利用语音特征区分抑郁症患者内部的不同子型，结果显示语音区分抑郁症中有无恐惧症的F值能达到75%。实验5为重复测量设计，利用语音特征区分45名抑郁症和17名双相障碍、58名健康人，结果显示：（1）语音区分抑郁症和双相障碍的F值能达到80.9%；（2）语音区分抑郁症、双相障碍和健康人的F值能达60.8%；（2）在不同任务和情绪下，语音的预测结果都达到了中等预测水平。研究二的结果表明，语音特征能够有效的用于区分抑郁症和与其症状相近的其它心理疾病，且区分效果具备跨情境的稳定性。
本文采用栅栏式的研究方式，逐步探索语音区分不同人群的效果，最终发现通过语音能够从抑郁症和双相障碍的混合人群中有效地识别出抑郁症且区分效果具备跨情境的稳定性。未来需要尝试在更大规模的样本上利用语音区分不同的抑郁程度和抑郁的不同子型。

Major depression disorder (MDD) is a kind of mental illness which is accompany with a core symptom of depressive mood and various other symptoms. To improve the diagnostic effect, we are eager to find an objective indicator that could effectively reflect the real status of depression. Speech is one kind of easily available behavioral clue. Many studies indicated that there are significant correlations between acoustic features and depressive symptoms, and it is possible to identify depression via acoustic features. Nevertheless, there are still two questions need to be figured out: how about the diagnostic power of speech when used it to identify depression? Are the predictive effects of speech cross-situational consistency?
To figure out the question about diagnostic power of speech, we designed a series of experiments to identify depression for finding out the utmost of identification in our case. In our studies, we applied classification algorithms of machine learning to analyze the predictive effects of acoustic features, and used metrics like F-measure to estimate the predictive models. The question about cross-situational consistency of speech’ predictive effects was investigated by comparing predictive effects under different experimental situations, tasks and emotions.
The main aim of study 1 is discriminate depressed people from non-mental patients. Study 1 includes three experiments. Experiment 1 compared the subjective emotional experiences among depressed, bipolar and healthy people, the results suggested that it is hard to distinguish these three groups while used self-report emotions only. Experiment 2 found that acoustic features can be used to differentiate depression and healthy people, the F-measure was reach 78.2%. Besides, the predictive effects can reach moderate levels under different situations. Experiment 3 indicated that acoustic features can be applied to distinguish depressed people and physical patients, the best F-measure was 75.6%.
The main aim of study 2 is distinguish depressed people from other kinds of mental patients. Study 2 includes two experiments. Experiment 4 shown that acoustic features can be applied to differentiate depressed people with one kind of comorbidity and depressed people without this comorbidity, the F-measure reached 75%. The goal of experiment 5 is to distinguish depression from bipolar disorder. The results reported:
(1) acoustic features can be used to differentiate depression and bipolar disorder, the best F-measure was 80.9%; (2) acoustic identification was able to discriminate healthy, bipolar and depressed people, the best F-measure was 60.8%; (3) the predictive effects of acoustic features can reach moderate levels under different situations.
This paper explores the diagnostic power of speech, and find that depression could be effectively identified from the mixture crowd of depressed and bipolar people. And this diagnostic power has cross-situational consistency. In the future studies, researchers should try to identify different degrees of MDD or subtypes within MDD in a larger sample.