2023 IEEE International Conference on Bioinformatics and Biomedicine

Classifying Anxiety and Depression through LLMs

Virtual Interactions: A Case Study with ChatGPT
2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) | 979-8-3503-3748-8/23/$31.00 ©2023 IEEE | DOI: 10.1109/BIBM58861.2023.10385305

1st Yongfeng Tao 2nd Minqiang Yang* 3rd Hao Shen

School of Information Science and School of Information Science and School of Information Science and
Engineering, Lanzhou University Engineering, Lanzhou University Engineering, Lanzhou University
Lanzhou, China Lanzhou, China Lanzhou, China
[email protected] [email protected] [email protected]

4th Zhichao Yang 5th Ziru Weng 6th Bin Hu*

School of Information Science and School of Information Science and School of Information Science and
Engineering, Lanzhou University Engineering, Lanzhou University Engineering, Lanzhou University
Lanzhou, China Lanzhou, China Lanzhou, China
[email protected] [email protected] [email protected]

Abstract—Mental health has long been studied, and many of mental disorders include both clinical diagnosis by pro-
AI-based approaches have been proposed for diagnosis and fessionals and treatment after diagnosis. Clinical diagnosis
adjunctive therapy. The emergence of Pre-trained Large Lan- has the disadvantages of subjectivity of experts and uncoop-
guage Models (LLMs) has had a profound impact on various
fields, but the potential of using ChatGPT for cognitive behavior erativeness of participants. To alleviate this problem, some
therapy is largely unexplored. Therefore, there is an urgent researchers in recent years have used physiological [3], [4]
need to build and design a virtual interactive framework for or behavioural [5], [6] data to model objective detection
assisted diagnosis/treatment. In this paper, we present a virtual methods. Treatment of mental disorders consists mainly of
interaction framework based on LLMs that allows participants medication and psychotherapy. However, studies have shown
to engage in a dialogue with a virtual character, analyse mental
health issues through augmented LLMs, and make suggestions that psychotherapy has comparable effects to medication in
during the dialogue to alleviate the psychological problems they the short term, but may be more effective in the long term [7].
are currently facing. Based on this framework, we develop a use Current psychotherapies include cognitive behavioural therapy
case for the application of ChatGPT in the field of emotional (CBT), problem-solving therapy, etc., with CBT being the
disorders. Specifically, we use data from question-and-answer most widespread and common method [7]. CBT was also
dialogues in real-life scenarios to populate the current exploration
of ChatGPT’s potential for depression and anxiety detection. slowly being introduced into the remote treatment of abnormal
The case study shows the great potential of ChatGPT in the mental states, with relatively good results. Psychotherapy
analysis of depression and anxiety tests. The feasibility of a virtual based on asynchronous messaging has shown its viability as an
interaction framework based on LLMs has been preliminarily adjunct treatment in combination with traditional counselling
demonstrated. services [8]. A virtual voice coach, Lumen, has been proposed
Index Terms—affective disorders, depression detection, anxiety
detection, virtual character
for the emotional treatment of subjects and has shown good
results in the cognitive control of depression and anxiety
symptoms [9]. With the development of avatar technology
I. I NTRODUCTION and the Pre-trained Large Language Models (LLMs), it has
The assessment of mental disorders in the field of affective become possible to build a model with interactive capabilities
computing has been one of the research hotspots in recent for the adjunctive treatment of mental disorders. Therefore, we
years. Mental disorders, as a common mental health problem, propose a virtual interaction framework based on enhanced
have affected people’s daily activities and most of them require LLMs, and the avatars in this framework is able to provide
treatment for improvement and recovery. There has also been participants with some advice and methods to alleviate the
a gradual increase in the prevalence of depression in recent mental health problems they are facing, based on the content
years, and there is already a related social burden. Previous of the dialogue.
studies have shown that up to 85% of depressive disorders For LLMs, such as ChatGPT [10], have been widely used in
are associated with anxiety [1], but [2] have shown that various domains since their release. ChatGPT has been used in
there are also phenotypic differences between depression and text classification, text generation, education, medicine, etc.,
anxiety, and that improving these phenotypic characteristics and has a wide range of applications in data visualisation,
could improve clinical assessment. Prevention and treatment information extraction, data enhancement, quality assessment
and multimodal data processing [11]. Overall, it can be divided
*Corresponding author. into two main categories: fine-tuning the application of mod-

2259
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.
979-8-3503-3748-8/23/$31.00 ©2023 IEEE
els [12], and testing the potential performance of models in used to assess and support the treatment of abnormal mental
different areas. Because ChatGPT can process text flexibly and states in a non-intrusive way. Our proposed virtual interaction
has good contextual understanding and information inference, framework is based on enhanced LLMs is shown in Fig. 1,
there are many studies testing its potential performance in and mainly consists of user-side and cloud-based augmented
sentiment analysis, personality recognition and affective disor- large model composition.
ders. In [13], the results of the zero-shot experiment suggest a
potential application of the ChatGPT in mental health classifi- TABLE I
T HE APPLICATION OF LLMS TO THE EXPLORATION OF POTENTIAL ON
cation tasks. In [14], the potential for text classification in three AFFECTIVE DISORDERS
affective computing problems was explored using ChatGPT.
In order to better use ChatGPT for analysing relevant tasks, Database Condition Use Post Num Platform Language
studies have explored the ability to zero-point classify text by CLPsych15 [18] DP / HC 300 [15] Twitter English
232,895/
introducing prompts. In [15], a variety of emotion-related cues TMDD [19] DP / HC
879,025
[16] Twitter English
were used to further explore the interpretability of ChatGPT 39,595/
WU3D [20] DP / HC [16] Weibo Chinese
on LLM for the 5-task emotional and mental health dataset 80,167
using ChatGPT. For affective disorders, the study [16] used DR [17] DP / HC 406 [15] Reddit English
the Diagnostic and Statistical Manual of Mental Disorders CAMS [21] DP / SP 626 [15] Reddit English
[22] DP / SP - [14] Reddit English
(DSM-IV) depression indicators as prompts and fed some of
Note: DP, HC, and SP denote depressed patients, healthy controls and
the social media users’ posts into the LLM model to determine suicidal patients, respectively.
whether they were depressed or not.
To explore the performance of LLM in addition to ab- Facial behaviour includes facial expression data and eye
normal sentiment analysis, researchers have used LLM to movement data, and are the most commonly used forms of
classify depressed and non-depressed, depressed and suicidal non-verbal communication. Studies of eye movement data
on different datasets, as shown in Table I. These studies in depressed patients have shown a negative attentional bias
were mainly based on datasets of posts on social media plat- [23]. The collection of facial behaviour data does not require
forms (Reddit, Twitter, Weibo, etc.), e.g. Depression Reddit much cooperation from the participants, and it can be used
(DR) [17], CLPsych15 [18], the Twitter multimodal depression as an input data to the virtual interaction framework is
dataset (TMDD) [19], the Weibo user depression detection based on enhanced LLMs to correct for effective changes in
dataset (WU3D) [20], the depression/suicide cause detection the virtual character’s facial emotions during the interaction
dataset (CAMS) [21], etc., and most of these datasets were with the participants. Significantly different body movements,
in English language. Based on the proposed framework, a such as slower arm swinging and walking, can be observed
case study of the potential of ChatGPT in affective disorders in depressed people [24], which is associated with fewer
is investigated. To the best of our knowledge, no Chinese- brain neurotransmitters in their negative mood. Similarly,
language Question and Answer (Q&A) data from real-life the acquisition of body movement data can be captured by
scenarios have been used to validate the performance of LLM a high-resolution camera in a distraction-free environment.
in anomalous sentiment detection. The main contributions of Participants with abnormal emotions showed different speech
this paper are as follows: characteristics compared to normal individuals [25]. Depressed
• We have constructed a virtual interaction framework patients mainly showed longer speech pauses, lack of rhythm,
based on enhanced LLMs. We also discuss the outstand- slower speech rate, etc. The main task of the user terminal
ing issues in the framework and provide ideas for future is to collect the participant’s behavioural data, pre-process it,
research. and provide real-time feedback on some suggestions and nec-
• We evaluated for the first time the potential of the Chat- essary intervention methods of the virtual character during the
GPT model to detect depression and anxiety in real-life interaction, such that these methods are used to alleviate the
scenarios using a database of question-answer dialogues. current participant’s abnormal mental state. In our framework,
• The descriptive information prompted by speech en- we have designed an assisted treatment scheme specifically for
hances ChatGPT’s ability to classify depression and anx- patients with mental disorders, which mainly uses LLMs and
iety, and this introduces an easy way to prompt input for avatar generation technology by taking the behavioural data
multiple modalities. captured by the camera and microphone and pre-processing
The rest of the paper is structured as follows. Section II them before inputting them as prompts to the corresponding
describes the virtual interaction framework based on enhanced modules in the framework.
LLMs, Section III describes the affective disorder case study, The framework can provide real-time feedback on the user’s
and Section IV discusses the challenges and open issues of psychological state and gradually guide the user to solve the
the framework. Section V concludes this paper. psychological problems faced in response to bad emotions.
The extracted effective emotional features are processed by
II. V IRTUAL I NTERACTIVE F RAMEWORK BASED ON LLM S processing multimodal data such as video, audio and text.
Physiological data, such as EEG, require specialised equip- These features not only serve as a reference for the virtual
ment to collect, so the use of visual and auditory data can be character’s emotional adjustment, but also help to prompt the

2260
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. The virtual interaction framework is based on enhanced LLMs. The main multimodal data collected by camera, microphone, etc. are processed and
transmitted to the virtual doctor in real time. The virtual doctor takes into account the textual content of the speech and the emotion conveyed by the patient
in order to provide timely feedback.

LLMs to accurately assess the participants’ psychological state patients with anxiety and 64 patients with depression were
in real time and provide assessment results, psychological recruited from the Peking University Sixth Hospital in China,
counselling methods, and so on. The feedback information and demographic information is shown in Table II. Diagnosis
is fed back to the user through the virtual character’s voice, recommended by at least one clinical psychiatrist. The local
enabling him or her to effectively reduce stress and improve ethics committee approved consent forms and studies designed
the quality of life and work. At the end of the interaction with for biomedical research at the Peking University Sixth Hospital
the user, the framework summarises the problems faced by the in China, following the Code of Ethics of the World Medical
user and assigns follow-up tasks that need to be attended to Asso-ciation (Declaration of Helsinki), and written informed
and completed in life. consent forms were signed by all participants before the
With the development of LLMs and virtual character tech- experiment.
nology, the virtual interaction framework based on enhanced The collected speech data consisted of nine Q&A tasks, and
LLMs is technically feasible. This framework allows patients six of the Q&A dialogue tasks were related to the participants’
to be treated regardless of time and place, and better protects daily activities and did not involve personal privacy, and were
their privacy without fear of being discovered or commented therefore chosen for this experiment. To input the speech data
on by others, allowing them to be more relaxed and focused into ChatGPT, the text data in the speech is extracted using a
on the therapeutic process. the virtual interaction framework combination of the Whisper [26] model and manual methods.
is designed to simulate conversations in real-life scenarios The text data contains some private information about places,
and uses CBT to achieve the purpose of assisting in the people, time and so on, we have done the privacy protection
treatment of psychiatric disorders. By interacting with the processing and used “xxx” to replace the original text.
virtual character, patients can provide feedback on their needs
and preferences to create a more personalised treatment plan TABLE II
D EMOGRAPHIC VARIABLES FOR PARTICIPANTS WITH DEPRESSIVE
and improve outcomes. This framework can alleviate the DISORDER (DP) AND ANXIETY PATIENT (AP).
problem of low doctor-patient ratios to some extent and save
time and effort for the therapist, thereby reducing the cost of Variables DP AP
treatment for the patient. N (M/F) 64 (19/45) 75 (33/42)
Age (m+SD) 32.09 ± 10.55 37.83 ± 11.31
III. C ASE S TUDY ON A FFECTIVE D ISORDERS
2) Feature Extraction: Previous studies have shown sig-
A. Materials and Methods nificant differences in speech characteristics between people
1) Participants and Preprocessing: To investigate the po- with different affective disorders, e.g. depressed people speak
tential performance of ChatGPT for the detection of anxiety more slowly, have longer pauses, etc [27]. We found texts
and depression in real-life dialogue scenarios. Seventy-five that appeared to have relatively normal descriptions, but the

2261
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.
speech sounded as if it was spoken in a choked, sighing 1) Compare Single Question: Before exploring multiple
voice. Therefore, we believe that there is a limit to the questions together, it is important to explore the potential
amount of emotional information that can be conveyed by of ChatGPT to discriminate between depression and anxiety
text alone. To characterise speech, we extracted speech rate using each individual question. We explored individual ques-
(words/sec), average pitch (HZ), speech rhythm (BPM), and tions using the prompt information question a) in Question
pauses (/sec), which are entered as prompts in ChatGPT. Formulation. Classification accuracy is best for question Q4
Features are extracted from speech using open source Python and second best for question Q2, as shown in Table III. And
packages, e.g. librosa, etc. In addition, as many current LLMs the values of Rec, F1, and Pre under questions Q2 and Q4
can only process text, the following prompt description is are relatively stable compared to the evaluation metrics under
given for speech speed features (speed): other questions, for example the Rec evaluation metrics of Q1
If speed > 200, make Dspeed = “faster speech speed, and Q6 are as high as 96%, suggesting that ChatGPT is more
{speed} words per minute”. If speed < 120, make Dspeed inclined to categorise the population into one category through
= “slower speaking speed, {speed} words per minute”. Oth- these questions. We found that Q2 and Q4 are questions about
erwise, Dspeed = “moderate speaking speed, {speed} words the participant’s recent mood and ‘what to do if the participant
per minute”. The description of pitch of speech is similar to is unhappy’, and we chose these two question dialogue data for
the above. the subsequent experiments. In order to explore the contextual
3) Question Formulation: The format of the question information between question and answer dialogues, we use
whether to add a voice message description is shown in the the prompt information from question a) in Question Formu-
following snippets. The question in the question and answer lation to determine the input data for the union of the two
dialogue is replaced by the section [question], the participant’s questions. As shown in the last row of Table III, the evaluation
answer is replaced by [text] and the voice message description metrics of Acc, Rec, F1, and Pre are 67.62%, 72.00%, 70.58%,
is replaced by D. The formula for calculating the prompt and 69.23%, respectively, which are significantly higher than
information for these two questions is as follows: the categorical metrics for single questions. This indicates that
a) For the question without adding voice description infor- the contextual information in the question-answer dialogue is
mation, we pose such a question: beneficial for improving the classification effect.
Analyze the dialogue to determine whether the respondent’s
emotional state is depression or anxiety. Question: [question], TABLE III
Answer: [text],..., Question: [question], Answer: [text], tell me ACCURACY, RECALL , F1- SCORE , AND PRECISION UNDER SINGLE
PROBLEMS .
the respondent’s emotion in the following format: “anxiety” or
“depression”. Just give me the final word, no further analysis. Question Acc(%) Rec(%) F1(%) Pre(%)
b) For questions with voice descriptions added, we ask the Q1 56.83 96.00 70.59 55.81
following question: Q2 57.55 76.00 65.89 58.16
Q3 52.52 88.00 66.66 53.66
Analyse the conversation to determine whether the respon- Q4 60.43 80.00 68.57 60.00
dent’s emotional state is depression or anxiety. Question: Q5 53.23 96.00 68.89 53.73
[question], Answer: [text], where [D],..., Question: [question], Q6 53.96 92.00 68.31 54.33
Answer: [text], where [D]. Tell me the mood of the respondent Q2/Q4 67.62 72.00 70.58 69.23
in the following format “anxiety” or “depression”. Just give
me the final word, no further analysis. 2) Compare Speech Descriptions: Several studies have
ChatGPT’s response answers may appear in arbitrary for- extracted audio features in speech for use in mental disorders,
mats, although we have fixed the output answers. Sometimes and speech features can be used as objective biomarkers for
the given answers have results that are inconsistent with the assessing the severity of depression, suggesting that speech
given answer, such as extra punctuation, ”my guess is ...”, and features may reflect abnormal mental states in people [27].
we adjust the output according to the principle of semantic Compared to healthy controls, depressed patients had slower
similarity. speech, longer pauses and longer reaction times, indicating the
presence of a psychomotor disorder. To explore the effective-
B. Results ness of speech features and speech descriptions in prompting
In this section, we explored the potential of ChatGPT ChatGPT to categorise depression and anxiety, we set up
for anxiety and depression classification, mainly following the input data according to the prompting message question
different prompts. First, as the experimental data used in b) in Question Formulation. As shown in Table IV, the
this paper were six task questions about participants’ daily classification results for the prompted input with added speech
activities, we need to explore the single question and answer description information are significantly improved in every
dialogues for depression and anxiety classification effects. Sec- evaluation index compared to the input without added speech
ondly, we chose two of the questions with high accuracy for description prompts. This shows that ChatGPT is better able
anxiety and depression classification to subsequently explore to distinguish between the two categories for text data with
the effect on classification between different descriptions of speech description information. Better classification results
speech information. are achieved with the addition of rhythm and speech rate,

2262
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.
which can reach 79.14% and 76.97% accuracy respectively.
The confusion matrix for pause and speech rhythm is shown
in Fig. 2. From the figure, it is evident that compared to
adding pause prompts, incorporating speech rate prompts is
more beneficial for the classification of anxiety. It is worth
noting that of these four speech feature descriptions, pitch has
the lowest classification accuracy.
Analysing negative words in speech texts through word (a) (b)
cloud (Fig. 3) and word frequency (Fig. 4) visualisation, we
observed differences in negative words between people with Fig. 4. Subfigure a and subfigure b represent the top ten negative words in
anxiety and depression. Words such as ‘anxious’ and ‘angry’ terms of word frequency for negative words in the speech texts of anxious
and depressed patients, respectively.
appear frequently in the speech texts of people with anxiety
disorders, whereas people with depression use the term ‘bad
in the framework to collaborate on virtual interactions, we
mood’ more often. Fig. 3 shows that, overall, anxiety has a
believe that the following aspects are worth exploring further:
more negative vocabulary, which is related to differences in the
verbalisation of anxiety and depression. As shown in Fig. 4, A. Generation of Virtual Characters
the highest frequency of the negative word ‘insomnia’ is found The behavioral capabilities of virtual characters include
in the anxiety and depression speech texts, which indicates that facial expressions, body movements, and language. Facial
both groups have some degree of sleep disorders. expressions are the core of constructing virtual characters
as they enhance the sense of immersion for patients. The
challenge in creating a virtual doctor system lies in achieving
realistic, smooth, and emotionally rich facial expressions and
movements. To reduce the cognitive load on participants,
the language, facial expressions, and body movements of
virtual characters must be natural and fluid during interactions.
Additionally, it is necessary to design language and behavioral
(a) (b) guidelines for medical assistance, such as determining which
behaviors should be prohibited. To accomplish these objec-
Fig. 2. Subfigure a and subfigure b represent confusion matrices of pauses
(/sec) and speech rhythm (BPM), respectively.
tives, extensive experimentation and consultation with industry
experts are required to develop various interaction strategies
and control mechanisms.
B. Privacy and Ethics
In the proposed framework, the collection of large amounts
of multimodal data from participants poses ethical challenges
to participant privacy. How to introduce federated machine
learning into the framework is a concern. As CBT therapy is a
long-term treatment, the framework must fully respect the pa-
(a) (b)
tient’s wishes without coercing or inducing the patient to make
Fig. 3. Subfigure a and subfigure b represent the word cloud distribution unwanted or inappropriate decisions. Following professional
of negative words in the speech texts of anxiety and depression patients, knowledge and ethical guidelines during the treatment process
respectively. to ensure the reliability, controllability and trustworthiness of
IV. R ESEARCH C HALLENGES AND O PEN I SSUES the virtual treatment is an important challenge.
We conducted a case study of chatgpt’s potential in depres- C. Robust Model
sion and anxiety using the framework, which proved to have To provide a robust CBT process, a multimodal fusion
significant potential. Combined with the various modules used model is required that combines different types of data such as
speech, facial behaviour, etc. to provide more accurate feed-
TABLE IV
back. The challenge is how to effectively represent the fusion
ACCURACY, RECALL , F1- SCORE , AND PRECISION OF DIFFERENT SPEECH of different modal data, exploiting their complementarities and
FEATURES AS PROMPTS UNDER THE UNION OF TWO PROBLEMS . redundancies to improve the generalisation and robustness of
Acoustic Acc(%) Rec(%) F1(%) Pre(%)
the model, and how to accurately align the different modal
Pause 75.74 80.00 77.92 75.94 data to improve the understanding and inference of the model.
Pitch 74.10 80.00 76.92 74.07 The patient’s mental state also affects speech habits such as
Rhythm 79.14 82.66 81.05 79.49 speech rate, pauses and accents, which may lead to incorrect
Speed 76.97 78.66 78.66 78.67 or missing recognition results from existing speech models,

2263
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.
affecting dialogue comprehension and generation. How to use [9] T. Kannampallil, O. A. Ajilore, N. Lv, J. M. Smyth, N. E. Wittels, C. R.
large language models in relation to the condition is also a Ronneberg, V. Kumar, L. Xiao, S. Dosala, A. Barve et al., “Effects
of a virtual voice-based coach delivering problem-solving treatment on
problem that needs to be overcome. emotional distress and brain function: a pilot rct in depression and
anxiety,” Translational Psychiatry, vol. 13, no. 1, p. 166, 2023.
V. C ONCLUSION [10] T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal,
This paper introduces a virtual interaction framework based A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language mod-
els are few-shot learners,” Advances in neural information processing
on LLMs to help alleviate negative psychological states in systems, vol. 33, pp. 1877–1901, 2020.
participants. By analysing the data obtained from the Q&A [11] Y. Liu, T. Han, S. Ma, J. Zhang, Y. Yang, J. Tian, H. He, A. Li,
dialogues, the findings have revealed the significant potential M. He, Z. Liu et al., “Summary of chatgpt/gpt-4 research and per-
spective towards the future of large language models,” arXiv preprint
of ChatGPT in classifying depression and anxiety. To improve arXiv:2304.01852, 2023.
the classification performance, four language features were [12] M. Maaz, H. Rasheed, S. Khan, and F. S. Khan, “Video-chatgpt: Towards
extracted as additional clues, with prosodic and speech rate detailed video understanding via large vision and language models,”
arXiv preprint arXiv:2306.05424, 2023.
features showing positive effects on classification. However, [13] B. Lamichhane, “Evaluation of chatgpt for nlp-based mental health
the current research cases still have some limitations, including applications,” arXiv preprint arXiv:2303.15727, 2023.
a small number of tasks and recruited participants. Future [14] M. M. Amin, E. Cambria, and B. W. Schuller, “Will affective computing
emerge from foundation models and general ai? a first evaluation on
studies will continue to recruit more participants and increase chatgpt,” IEEE Intelligent Systems, vol. 38, p. 2, 2023.
the number of tasks to further optimize this framework. [15] K. Yang, S. Ji, T. Zhang, Q. Xie, and S. Ananiadou, “On the evaluations
Despite the challenges in fine-grained classification, the virtual of chatgpt and emotion-enhanced prompting for mental health analysis,”
arXiv preprint arXiv:2304.03347, 2023.
interaction framework based on LLM is worth exploring and [16] W. Qin, Z. Chen, L. Wang, Y. Lan, W. Ren, and R. Hong, “Read,
implementing based on current technological feasibility and diagnose and chat: Towards explainable and interactive llms-augmented
potential demand. depression detection in social media,” arXiv preprint arXiv:2305.05138,
2023.
ACKNOWLEDGMENT [17] I. Pirina and Ç. Çöltekin, “Identifying depression on reddit: The effect of
training data,” in Proceedings of the 2018 EMNLP workshop SMM4H:
This work was supported in part by the National Key the 3rd social media mining for health applications workshop & shared
Research and Development Program of China (Grant No. task, 2018, pp. 9–12.
[18] G. Coppersmith, M. Dredze, C. Harman, K. Hollingshead, and
2019YFA0706200), in part by the National Natural Science M. Mitchell, “Clpsych 2015 shared task: Depression and ptsd on twitter,”
Foundation of China (Grant No.62227807), in part by the in Proceedings of the 2nd workshop on computational linguistics and
Natural Science Foundation of Gansu Province, China(Grant clinical psychology: from linguistic signal to clinical reality, 2015, pp.
31–39.
No. 22JR5RA488), in part by the Fundamental Research [19] T. Gui, L. Zhu, Q. Zhang, M. Peng, X. Zhou, K. Ding, and Z. Chen,
Funds for the Central Universities (Grant No. lzujbky-2023- “Cooperative multimodal approach to depression detection in twitter,”
16). The authors would like to acknowledge Yanfeng Zhang at in Proceedings of the AAAI conference on artificial intelligence, vol. 33,
no. 01, 2019, pp. 110–117.
Georgetown University for providing the relevant computing [20] Y. Wang, Z. Wang, C. Li, Y. Zhang, and H. Wang, “A multitask deep
resources. learning approach for user depression detection on sina weibo,” arXiv
preprint arXiv:2008.11708, 2020.
R EFERENCES [21] M. Garg, C. Saxena, S. Saha, V. Krishnan, R. Joshi, and V. Mago,
[1] J. M. Gorman, “Comorbid depression and anxiety spectrum disorders,” “Cams: An annotated corpus for causal analysis of mental health issues
Depression and anxiety, vol. 4, no. 4, pp. 160–168, 1996. in social media posts,” in Proceedings of the Thirteenth Language
[2] Y. N. Gao, B. Coombes, E. Ryu, V. Pazdernik, G. Jenkins, R. Pendegraft, Resources and Evaluation Conference, 2022, pp. 6387–6396.
J. Biernacka, and M. Olfson, “Phenotypic distinctions in depression [22] V. Desu, N. Komati, S. Lingamaneni, and F. Shaik, “Suicide and
and anxiety: a comparative analysis of comorbid and isolated cases,” depression detection in social media forums,” in Smart Intelligent Com-
Psychological Medicine, pp. 1–9, 2023. puting and Applications, Volume 2: Proceedings of Fifth International
[3] D. Wang, C. Lei, X. Zhang, H. Wu, S. Zheng, J. Chao, and H. Peng, Conference on Smart Computing and Informatics (SCI 2021). Springer,
“Identification of depression with a semi-supervised gcn based on eeg 2022, pp. 263–270.
data,” in 2021 IEEE International Conference on Bioinformatics and [23] M. Yang, C. Cai, and B. Hu, “Clustering based on eye tracking
Biomedicine (BIBM). IEEE, 2021, pp. 2338–2345. data for depression recognition,” IEEE Transactions on Cognitive and
[4] X. Zhang, W. Liang, T. Ding, J. Pan, J. Shen, X. Huang, and J. Gao, Developmental Systems, pp. 1–1, 2022.
“Individual similarity guided transfer modeling for eeg-based emotion [24] H. Lu, S. Xu, X. Hu, E. Ngai, Y. Guo, W. Wang, and B. Hu, “Postgrad-
recognition,” in 2019 IEEE International Conference on Bioinformatics uate student depression assessment by multimedia gait analysis,” IEEE
and Biomedicine (BIBM). IEEE, 2019, pp. 1156–1161. MultiMedia, vol. 29, no. 2, pp. 56–65, 2022.
[5] Y. Tao, M. Yang, Y. Wu, K. Lee, A. Kline, and B. Hu, “Depressive [25] Z. Liu, H. Yu, G. Li, Q. Chen, Z. Ding, L. Feng, Z. Yao, and B. Hu,
semantic awareness from vlog facial and vocal streams via spatio- “Ensemble learning with speaker embeddings in multiple speech task
temporal transformer,” Digital Communications and Networks, 2023. stimuli for depression detection,” Frontiers in Neuroscience, vol. 17, p.
[6] B. Hu, Y. Tao, and M. Yang, “Detecting depression based on facial 1141621, 2023.
cues elicited by emotional stimuli in video,” Computers in Biology and [26] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and
Medicine, vol. 165, p. 107457, 2023. I. Sutskever, “Robust speech recognition via large-scale weak super-
[7] P. Cuijpers, C. Miguel, M. Harrer, C. Y. Plessen, M. Ciharova, D. Ebert, vision,” in International Conference on Machine Learning. PMLR,
and E. Karyotaki, “Cognitive behavior therapy vs. control conditions, 2023, pp. 28 492–28 518.
other psychotherapies, pharmacotherapies and combined treatment for [27] Y. Yang, C. Fairbairn, and J. F. Cohn, “Detecting depression severity
depression: a comprehensive meta-analysis including 409 trials with from vocal prosody,” IEEE transactions on affective computing, vol. 4,
52,702 patients,” World Psychiatry, vol. 22, no. 1, pp. 105–115, 2023. no. 2, pp. 142–150, 2012.
[8] J. Song, B. Litvin, R. Allred, S. Chen, T. D. Hull, and P. A. Areán,
“Comparing message-based psychotherapy to once-weekly, video-based
psychotherapy for moderate depression: Randomized controlled trial,”
Journal of Medical Internet Research, vol. 25, p. e46052, 2023.

2264
Authorized licensed use limited to: Amrita School of Engineering. Downloaded on October 09,2024 at 02:58:59 UTC from IEEE Xplore. Restrictions apply.