prevenIA: a Chatbot for Information and Prevention of Suicide and other Mental Health Disorders

Abstract

Suicide and mental disorders are major health and social issues worldwide; therefore, a simple access to reliable sources of information that can mitigate these problems is instrumental for people sufering them and for their relatives. The goal of this project is the development of a Whatsapp chatbot that will provide resources and advices to help in suicide prevention. In addition, the chatbot will detect sensitive users and derive them to a team of specialists. Initially, the focus of the projet is on suicide prevention and in the Whatsapp platform, but it will be extended to other mental disorders and other platforms to reach as many people as possible. In conclusion, this project aims to disseminate reliable information about suicide and mental disorders in a format that is understandable and easily accessible by the general population.

Keywords

• eol>Chatbot
• Semantic Search
• Question Answering
• Suicide
• Mental disorders

1. Introduction

vantages of these tools. Then, we explain the phases into which the development of the chatbot will be divided, as well as the team that will carry them out. Finally, we draw some conclusions about the hypothetical results that are intended to be obtained with this work.

2. Conceptual framework

user and will express concern for their well-being.

It is important to understand that chatbots also have their limitations. Human communication has nuances that are complicated for artificial intelligence to understand, and inappropriate or unclear responses must be avoided [ 10, 15 ]. In addition, the non-verbal language that is often decisive in this context is lacking in chatbots [ 8 ]. A special care aspect is to provide an adequate reaction once an emergency situation has been detected [ 12 ]. All these limitations must be treated with particular care in a subject as sensitive as suicide risk [ 15 ].

A very noteworthy aspect of the literature reviews on the use of chatbots in mental health is that most of the studies have been conducted in English-speaking populations, and there is a notable absence of works for Spanish-speakers [ 9, 11, 12, 13 ]. An exception is the work by Romero et al. [ 8 ] wherein the basis for the design of a chatbot with psychological assessment functions is presented. Research into the personalisation of chatbots in order to provide answers to diferent types of users is also highlighted as an interesting and little-studied aspect.

In particular, the complexity of the language could be adapted to the level required by the user [ 12 ]. Finally, the uses of machine learning methods that stand out include the classification and detection of people potentially at risk of suicidal behaviour, but there is no evidence of studies that involve providing information, for example to family members, about suicidal behaviour [ 13 ].

A chatbot, or conversational assistant, is a software application that simulates a conversation with a person by providing automatic responses, and from whose application it is possible to obtain some information or some kind of action [ 8 ]. Chatbots are currently being used in a wide range of fields, including health in general [ 9 ] and mental health in particular [ 10 ]. In fact, the use of chatbots in mental health is present in the very origins of these tools in the 1960s, a period in which what is considered to be the first chatbot, called ELIZA, was developed.

This chatbot made it possible to simulate a conversation with a psychologist in a psychotherapy session [ 8 ].

There are several recent literature reviews on the use of chatbots in mental health [ 9, 10, 11, 12 ] and also on the use of artificial intelligence methods in aspects related to suicide [ 13 ]. These reviews highlight aspects where chatbots can be useful in this area. On the one hand, a chatbot can give access to virtual services to certain people who would avoid using a face-to-face service, either because the latter is overburdened, because they cannot aford it, or to avoid the stigma attached to certain people with mental health problems. On the other hand, 3. Work plan the anonymity ofered by chatbots allows some people, especially the younger ones, to seek information about The development of the project is divided into the foltheir doubts or freely express their feelings and prob- lowing phases. lems; feelings that they are not comfortable to be shared to other human beings [ 10, 13, 14 ]. Furthermore, both 3.1. Phase 1. Informative chatbot people who use these chatbots [ 12 ] and mental health professionals [ 15 ] have a positive perception and opinion In this first phase, a chatbot will be developed on the of them. However, although it is emphasised that these WhatsApp platform. Through natural language processsystems can help the professional in some aspects, they ing techniques, the chatbot will be able to automatically are never intended to replace them [ 15 ]. provide information related to suicide, for example, to

In a literature review carried out in 2022 by Valizadeh prevent it or to deal with mourning situations. To achieve and Parde [ 9 ] on the application of chatbots in health, this aim, it will be necessary to build a corpus of infor70 studies were identified; and 22 of them correspond to mation sources using the materials available on websites diferent pathologies related to mental health. Among such as MenteScopia1, prevensuic2, or the prevention these pathologies are depression, anxiety, phobias or ad- plans created by the Autonomous Regions. dictions; however, none of these studies were related to Once the corpus has been built, models will be comsuicide. After that review was published, the designed of bined to perform semantic searches and to answer quesa chatbot for the detection of suicidal ideation has been tions [ 17 ]. The former will be used to extract the context proposed [ 14 ]. This detection is done through a natural that will be used by the latter to answer users’ questions. language processing model, called BERT, retrained on One of the biggest challenges in this context is that both a database obtained from a Reddit subnetwork, called types of models are mainly used in the context of the Reddit Suicide Watch [ 16 ]. If ideation is detected, the English language, so it will be necessary to use either user is asked for permission to send help; if permission is not given, the chatbot will continue chatting with the

1https://psynal.eu/mentescopia/ 2https://www.prevensuic.org/

translation models or to train new models. A small ap- or specialised staf working in community help services plication developed with Gradio is already available (see such as the “Teléfono de la Esperanza”. Instead, it is inFigure 1) which allows answers to be obtained from a tended to locate risk cases by means of surveys based series of documents identified in an initial search. on clinical interviews developed by psychologists and

Finally, in this phase, the accessibility of the chatbot psychiatrists [ 21 ]. In addition, the chatbot will make will be taken into account. Thus, easy read and plain it easier for users to create help resources such as “My language techniques will be used to allow users to find safety plan” [ 3 ]. the information they need, understand what they find and use that information. The work done in the context 3.3. Phase 3. Extension to other platforms of CLARA-MeD [ 18 ] for medical text simplification will be particularly useful for this. Moreover, the chatbot The chatbot developed in the previous two phases will be will include the necessary functionality to communicate implemented on other platforms in addition to WhatsApp with users in written or oral form. In the latter, it will in order to reach the widest possible audience. These be necessary to use models such as Whisper [ 19 ], which platforms include Facebook chat, messaging apps such will allow the chatbot to transcribe the audios sent by as Telegram, or platforms such as Discord. the user; or fairseq 2, to synthesise voice from text [ 20 ].

It is important to emphasise that at the beginning of 3.4. Phase 4. Extension to other disorders the conversation the users will be informed that they are interacting with a machine, and will always have an option to choose to speak to a person via a specialised service such as 112.

With all the experience from the previous phases, the chatbot will be enhanced to not only provide information on suicide, but also on other mental health disorders such as eating and behavioural disorders.

3.2. Phase 2. Preventive chatbot

In this phase, the chatbot developed in Phase 1 will be

extended to automatically detect possible risks. If a sensitive situation is detected, the user will be redirected to 112 or to a support team with specialised professionals in SERIS.

For the moment, it is not intended to build an open text chatbot that can replace psychologists, psychiatrists

4. The working team

The project described in this paper is carried out by a

multidisciplinary team combining experts in computer science, psychiatry and linguistics. The technical development of the project is carried out within the Computer Science Group of the University of La Rioja. In this sense, diferent members of the group collaborate in diferent tasks of the work plan.

María Soledad Campos, a doctor in psychiatry from the Espartero Mental Health Unit in La Rioja, is also collaborating with the project and will contribute her expert vision for the compilation and review of materials related to suicide as well as clinical interviews present in the literature. She will also be in charge of validating the correctness of the answers provided by the chatbot.

Finally, we are working with Ana Rosa Terroba Reinares, PhD in Linguistics and specialist in health communication at SERIS. Her contribution will be essential for the chatbot to communicate clearly with users, in addition to contributing her knowledge in the generation of the corpus that feeds the chatbot.

5. Conclusions

Mental health is a serious problem in today’s society, and the use of natural language processing and artificial intelligence tools, such as the one we intend to build with the chatbot of this project, can support patients and families. However, the aim is not to replace the human role played by agents such as psychiatric teams or organisations such as the Telefono de la Esperanza, but to complement it with user-friendly tools which provide reliable information.

Acknowledgments

This work was partially supported by Grant PID2020-115225RB-I00 funded by MCIN/AEI/10.13039/501100011033 and the research contract OTCA230110.

References

[1] Instituto Nacional de Estadística , Defunciones según la causa de muerte Año 2021 , Technical Report , 2022 . URL: https://www.ine.es/prensa/edcm_ 2021.pdf .

[2] WHO, Suicide worldwide in 2019: global health estimates , 2021 .

[3] Rioja Salud , Plan de prevención del suicidio en La Rioja , 2019 . URL: https://www.riojasalud.es/files/ content/ciudadanos/planes-estrategicos/PLAN_ PREVENCION_ CONDUCTA_SUICIDA_DEF .pdf .

[4] Servicio Canario de Salud , Programa de prevención de la conducta suicida en Canarias , 2021 . URL: https: //www3.gobiernodecanarias.org/sanidad/scs/ content/3f5ce57d-1085 - 11ec-bfb0-874800d2c074/ PPCSC.pdf .

[5] Gobierno de Navarra , Prevención y actuación ante conductas suicidas , 2014 . URL: https://www.educacion.navarra.es/ documents/27590/548485/Suicidio.pdf/ b5374981-511a - 40ed - 82c5-7c74bc23b049.

[6] T. Sufrate-Sorzano , E. Jiménez-Ramón , M. E. Garrote-Cámara , V. Gea-Caballero , A. Durante , R. Júarez-Vela , I. Santolalla-Arnedo, Health plans for suicide prevention in Spain: a descriptive analysis of the published documents , Nursing Reports 12 ( 2022 ) 77 - 89 .

[7] D. Crespo Amaro , P. Vázquez Herrera , G. Velázquez Basterra , M. S. Campos Burgui , Salud mental y covid- 19 , Zubía ( 2021 ) 131 - 140 .

[8] M. Romero , C. Casadevante , H. Montoro , Cómo construir un psicólogo-chatbot , Papeles del Psicólogo 41 ( 2020 ) 27 - 34 .

[9] M. Valizadeh , N. Parde , The AI doctor is in: A survey of task-oriented dialogue systems for healthcare applications , in: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1 : Long Papers) , 2022 , pp. 6638 - 6660 .

[10] A. N. Vaidyam , H. Wisniewski , J. D. Halamka , M. S. Kashavan , J. B. Torous , Chatbots and conversational agents in mental health: a review of the psychiatric landscape , The Canadian Journal of Psychiatry 64 ( 2019 ) 456 - 464 .

[11] T. Zhang , A. M. Schoene , S. Ji , S. Ananiadou , Natural language processing applied to mental illness detection: a narrative review , NPJ digital medicine 5 ( 2022 ) 46 .

[12] A. A. Abd-Alrazaq , M. Alajlani , N. Ali , K. Denecke , B. M. Bewick , M. Househ , Perceptions and opinions of patients about mental health chatbots: scoping review , Journal of medical Internet research 23 ( 2021 ) e17828 .

[13] S. Ji , S. Pan , X. Li , E. Cambria , G. Long , Z. Huang , Suicidal ideation detection: A review of machine learning methods and applications , IEEE Transactions on Computational Social Systems 8 ( 2020 ) 214 - 226 .

[14] J. X. Chan , S.-L. Chua , L. K. Foo , A two-stage classiifcation chatbot for suicidal ideation detection , in: International Conference on Computer, Information Technology and Intelligent Computing (CITIC 2022 ), Atlantis Press, 2022 , pp. 405 - 412 .

[15] C. Sweeney , C. Potts , E. Ennis, R. Bond , M. D. Mulvenna , S. O'neill , M. Malcolm , L. Kuosmanen , C. Kostenius , A. Vakaloudis , et al., Can chatbots help support a person's mental health? Perceptions and views from mental healthcare professionals and experts , ACM Transactions on Computing for Healthcare 2 ( 2021 ) 1 - 15 .

[16] S. Ji , C. P. Yu , S.-f. Fung, S. Pan , G. Long, Supervised learning for suicidal ideation detection in online user content , Complexity 2018 ( 2018 ).

[17] A. Esteva , A. Kale , R. Paulus , K. Hashimoto , W. Yin , D. Radev , R. Socher , Covid -19 information retrieval with deep-learning based semantic search, question answering, and abstractive summarization , NPJ digital medicine 4 ( 2021 ) 68 .

[18] L. Campillos Llanos , A. R. Terroba Reinares , S. Zakhir Puig , A. Valverde , A. Capllonch-Carrión , Building a comparable corpus and a benchmark for Spanish medical text simplification ( 2022 ).

[19] A. Radford , J. W. Kim , T. Xu, G. Brockman , C. McLeavey , I. Sutskever , Robust speech recognition via large-scale weak supervision , arXiv preprint arXiv:2212.04356 ( 2022 ).

[20] C. Wang , W.-N. Hsu , Y. Adi , A. Polyak , A. Lee , P.-J. Chen , J. Gu , J. Pino, fairseq Sˆ 2: A scalable and integrable speech synthesis toolkit , arXiv preprint arXiv:2109.06912 ( 2021 ).

[21] D. V. Sheehan , Y. Lecrubier , K. H. Sheehan , P. Amorim , J. Janavs , E. Weiller, T. Hergueta , R. Baker , G. C. Dunbar , et al., The mini-international neuropsychiatric interview (MINI): the development and validation of a structured diagnostic psychiatric interview for DSM-IV and ICD-10 , Journal of clinical psychiatry 59 ( 1998 ) 22 - 33 .