Can Large Language Models be sensitive to Culture Suicide Risk Assessment?

doi:10.21203/rs.3.rs-4066705/v1

Download PDF

Research Article

Can Large Language Models be sensitive to Culture Suicide Risk Assessment?

https://doi.org/10.21203/rs.3.rs-4066705/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 02 Nov, 2024

Read the published version in Journal of Cultural Cognitive Science →

You are reading this latest preprint version

Suicide remains a pressing global public health issue. Previous studies have shown the promise of Generative Intelligent (GenAI) Large Language Models (LLMs) in assessing suicide risk in relation to professionals. But the considerations and risk factors that the models use to assess the risk remain as a black box.

This study investigates if ChatGPT-3.5 and ChatGPT-4 integrate cultural factors in assessing suicide risks (probability of suicidal ideation, potential for suicide attempt, likelihood of severe suicide attempt, and risk of mortality from a suicidal act) by vignette methodology. The vignettes examined were of individuals from Greece and South Korea, representing countries with low and high suicide rates, respectively. The contribution of this research is to examine risk assessment from an international perspective, as large language models are expected to provide culturally-tailored responses. However, there is a concern regarding cultural biases and racism, making this study crucial.

In the evaluation conducted via ChatGPT-4, only the risks associated with a severe suicide attempt and potential mortality from a suicidal act were rated higher for the South Korean characters than for their Greek counterparts. Furthermore, only within the ChatGPT-4 framework was male gender identified as a significant risk factor, leading to a heightened risk evaluation across all variables. ChatGPT models exhibit significant sensitivity to cultural nuances. ChatGPT-4, in particular, offers increased sensitivity and reduced bias, highlighting the importance of gender differences in suicide risk assessment.

Artificial Intelligence

bias

cultural diversity

gender

suicide

mental health

Assessing suicide risk is a complex and multifaceted challenge that requires consideration of a wide range of personal, social, and cultural factors (Graney et al., 2020). With the development of AI-based models like ChatGPT, new opportunities have emerged to improve the accuracy and efficiency of risk assessment processes (Levkovich & Elyoseph, 2023). However, the considerations and risk factors that these models use to assess risk remain a black box. To fully harness the potential of these technologies, it is crucial to examine their ability to account for cultural and gender differences when conducting risk assessments. While large language models (LLMs) hold the promise of providing culturally tailored responses (Elyoseph et al., 2024) there is a concern that cultural biases and the fear of being perceived as racist may hinder the integration of important cultural factors into clinical judgment (Elyoseph et al., 2024; Hadar-Shoval et al., 2024). This research aims to investigate how advanced language models, such as ChatGPT-3.5 and ChatGPT-4, incorporate culture- and gender-related risk factors into their suicide risk assessments. This examination is vital for developing culturally sensitive and effective tools for suicide risk assessment that can help bridge gaps in mental health services worldwide.

Artificial intelligence (AI) has been applied in a myriad of fields, from medicine to mental health (Elyoseph & Levkovich, 2023; Kang, 2021; Tal et al, 2023; Xu et al., 2023). In the realm of cultural diversity, AI offers promise in addressing mental health disparities by tailoring interventions to historically underserved populations, transcending language barriers, and promoting cultural sensitivity (Fiske et al., 2019; van Heerden et al., 2023). These endeavors include creating mental health programs in underrepresented languages and supporting community-focused initiatives. Nevertheless, AI can also introduce inequalities due to variable access, language limitations, and cultural biases (Elyoseph & Levkovich, 2023; Wampold & Flückiger, 2023). Therefore, in utilizing AI in mental health we must consider the impact within diverse cultural contexts and balance the potential benefits and challenges. The current study examines the use of artificial intelligence in the field of suicide assessment in the context of cultural and gender differences.

Suicide constitutes a critical challenge within the sphere of public health that necessitates immediate attention and intervention (Levi-Belz et al., 2022; Qian, 2021; WHO, 2020). Suicidality, an urgent concern within public health, is beset by obstacles to accurate assessment, including issues related to psychometric intricacies as well as barriers in community accessibility (Baek et al.,2021). The complex issue of suicide covers a spectrum of behaviors ranging from suicidal thoughts to serious attempts and actual deaths (Knipe et al., 2022). These actions vary in severity and have broad social and public health impacts (Gvion & Levi-Belz, 2018). Risk factors differ across demographic and social groups, reflecting both individual and societal well-being (Feigelman et al., 2019). Despite academic focus on demographic and economic factors, the varying rates across countries highlight that no single factor provides a full explanation (Bowden et al., 2020).

Research literature has devoted substantial resources toward developing evidence-based preventative measures, underscoring the pivotal role played by healthcare professionals in the early detection and crisis management of at-risk individuals (Bolton et al.,2015). To extend the reach of risk assessment protocols, recent initiatives have invested in training community gatekeepers (Bolton et al.,2015). Artificial Intelligence (AI) has emerged as a viable mechanism for augmenting the decision-making abilities of these community figures, with prospective advantages in terms of both diagnostic precision and public reach (Elyoseph & Levkovich, 2023; Levkovich & Elyoseph, 2023). Nevertheless, the capacity of AI algorithms to account for multicultural sensitivities has not been adequately examined. The present inquiry seeks to redress this lacuna by scrutinizing the manner in which AI algorithms allocate weight to salient cultural variables in their assessments of suicidality.

The Organization for Economic Co-operation and Development (WHO, 2020) published the following data, originally taken from the WHO Mortality Database: South Korea has the highest suicide rate in the developed world, with 24.1 suicides per 100,000 people (Kim et al., 2019). In contrast, the suicide rate in Greece is as low as 3.9 per 100,000 people. In this study, we chose to examine the country that tops the suicide frequency list as well as one that is at the bottom. Due to underreporting in different countries, actual rates may vary.

More than 800,000 individuals worldwide succumb to suicide each year (WHO, 2020). South Korea exhibits the highest incidence of suicide among OECD countries, with prevalence prominently higher among males and older adults. The suicide rate in South Korea is more than double the OECD average, which stands at 11.0 suicides per 100,000 people (WHO, 2020). Beginning in 1992, the aggregate rate of suicide in South Korea has exhibited an upward trajectory. This escalating trend was notably exacerbated in 1998, coinciding with the onset of the International Monetary Fund (IMF) crisis. Moreover, it subsequently intensified in 2009, in the immediate aftermath of the global financial crisis (Baek et al., 2021). Supplementary explanations for this trend chiefly attribute it to demographic aging, with a concomitant rise in suicide rates particularly among older and middle-aged populations (Kim et al., 2020; Lee et al., 2017). The erosion of traditional family-centered values coupled with economic deprivation among older adults have been posited as contributing factors to the rise in the number of suicides within this demographic group (Chang et al., 2009). Additionally, the marked escalation in suicides by gas poisoning, which surged more than twenty-fold during the first decade of the 21st century, suggests that accessibility of this facile means of suicide may play a role in amplifying the suicide rate (Lim et al., 2014). Cross-sectional analyses have further identified lower educational attainment, rural domicile, area-level socioeconomic deprivation (Kim, 2020), and diminished income (Lee et al., 2017; Lee et al., 2022) as potential variables linked to elevated suicide risk.

In addition, the heightened prevalence of divorces in South Korea is considered a partial explanatory variable for elevated suicide rates (Kim, 2020; Yamaoka et al., 2020), with divorce identified as a pivotal risk factor for suicide. Three principal mechanisms have been posited to explain this relationship: first, divorce leads to the disintegration of social and familial ties, thereby exacerbating psychological distress (Yamaoka et al., 2020); second, the termination of emotional interdependence between spouses intensifies emotional distress; and third, divorce often precipitates financial vulnerabilities, especially among women, due to insufficient welfare provisions and the demands of single parenthood (Lee et al., 2017; Lee et al., 2022). Taken together, these contributing factors heighten the suicide rate among divorcees, affirming the complex and multifactorial nature of suicide risk.

Greece, in contrast, currently has one of the world’s lowest suicide rates. Yet this has not always been the case. In view of the economic crisis that has enveloped Europe since 2008, rising suicide rates in Greece attracted heightened scrutiny. That economically turbulent period characterized by elevated unemployment rates and negative economic growth had a discernible impact on various dimensions of everyday life, and presumably on mental health. Research conducted across European Union nations corroborates this observation by identifying an association between suicide mortality rates and unemployment (Stuckler et al., 2009). Research literature examining Greece reported alarming increases in suicide rates, peaking at up to 40% (Kontaxakis et al., 2013; Rachiotis et al., 2015). This rise was more pronounced among women, who exhibited an increase of 69.6%, compared to 33.1% among men (Kontaxakis et al., 2013). Another study identified a 35% uptick in suicides in Greece between 2010 and 2012. This study also found that unemployment was significantly related to suicide mortality, particularly among men of working age, a pattern in line with the onset of austerity measures (Rachiotis et al., 2015).

Several primary factors can explain the marked decline in suicide rates in Greece in recent years. First, empirical research suggests that countries close to the Mediterranean Sea generally exhibit lower suicide rates, possibly due to the region's more relaxed lifestyle (Eskin, 2020). Second, suicide rates demonstrate substantial intersocietal variation (Mortier et al., 2018). A comparative analysis across 22 nations revealed that elevated suicide rates were primarily found in three largely Catholic countries: Slovenia, France, and Croatia (Eskin, 2020). Nevertheless, even though the role of religious belief as a protective factor against suicidal tendencies has been substantiated by research (Gearing & Alonzo, 2018), research literature on non-fatal suicidal behavior in Mediterranean countries is limited. Some studies indicate that while religious affiliation may not guard against suicidal ideation, it does appear to deter actual suicide attempts (Lawrence et al., 2016).

The Chat Generative Pre-Trained Transformer (ChatGPT) is an AI-based language model with applications across diverse sectors, including education, scientific research, and healthcare (Hadar-Shoval et al., 2023; Fraiwan et al., 2023; Tal et al., 2023). Recently, ChatGPT has demonstrated its potential in medical contexts, particularly in mental health (Levkovich & Elyoseph, 2023; Hadar-Shoval et al., 2023; Tal et al., 2023). Its machine-learning algorithms, trained on extensive healthcare data, have the potential to assist clinicians in decision-making and enhance the predictive accuracy of tools assessing suicidal behavior (Elyoseph & Levkovich, 2023; Sallam, 2023). Nevertheless, adoption of ChatGPT requires careful evaluation due to limitations and costs (Sallam, 2023; Tal et al., 2023). For instance, ChatGPT has been found to underestimate suicide risks, raising questions about its reliability in critical assessments (Elyoseph & Levkovich, 2023). Moreover, training the model on online data poses risks of disseminating inaccurate information, which is a matter of particular concern in the case of individuals with mental health disorders (Cheng et al., 2023).

Therefore, while ChatGPT offers promising avenues in mental healthcare, its limitations necessitate cautious implementation (Sallam, 2023; Tal et al., 2023). The challenge of accessing reliable suicide risk assessments is particularly acute in developing countries. This issue is further complicated by the global trend toward the expansion of cultural diversity within nations, making intercultural considerations essential even in Western settings. The current research seeks to address this gap by investigating whether artificial intelligence can effectively incorporate cultural factors in its suicide risk assessments. The ultimate aspiration is to leverage AI technology to provide personalized and culturally sensitive mental health services on a global scale.

The current study sought to examine whether ChatGPT-3.5 and ChatGPT-4 incorporate risk factors such as country/culture in their assessments of suicide risk. These risk assessments include the likelihood of serious suicide attempts, suicide attempts, and suicidal thoughts. We hypothesized that compared to ChatGPT-3.5, ChatGPT-4 would exhibit enhanced consideration of these suicide risk factors.

Large Language Models Procedure

In September 2023, we assessed the aptitude of both ChatGPT-3.5 and ChatGPT-4 (3 August version) in evaluating cross-cultural sensitivity related to suicide risk factors, including likelihood of death from suicidal acts, likelihood of serious suicide attempt, likelihood of suicide attempt, and likelihood of suicidal thoughts. Since the study did not involve real patients and only made use of vignettes and GPT technology, we determined that seeking approval from an ethics committee was not required.

Input Source

The foundational vignette used in the current study depicts a 37-year-old person with no prior records of suicide attempts and was originally used in a study by Levi-Belz and Gamliel (2016). It has also been employed in subsequent research (Levkovich & Elyoseph, 2023; Elyoseph & Levkovich, 2023). To assess how cultural differences influence ChatGPT's evaluation of suicide risk, we entered a vignette into the AI interface that portrayed either a Greek character (man or woman) or a South Korean character (man or woman). For authenticity, we used common first names from Greece and South Korea (Georgios and Maria, Seo-Jun and Seo-Yeon, respectively). Each vignette was evaluated ten times in both ChatGPT-3.5 and ChatGPT-4, for a total of 80 evaluations, each conducted in a separate tab.

The text of the vignette is as follows:

Maria (or Georgios/ Seo-Yeon/Seo-Jun ) is a 37-year-old single woman ( or man ) with no children, who was born and lives in Greece (or South Korea). She (he) currently lives alone, but in the past spent a long time living with her (his) father, who passed away two years ago following a serious illness. Maria holds a B.A. degree and has held a steady job for 10 years. Recently, however, her manager told her she is about to be dismissed due to financial cuts. Over the past year, Maria had a close and intimate relationship with a boyfriend. This was her first serious and meaningful relationship; she was very excited about it and hoped it would lead to marriage. But after a period of repeated quarrels, her boyfriend informed her he no longer loved her, and he left her. Maria was left feeling very sad and hurt. In recent weeks she has thought about the relationship repeatedly and tried to analyze what happened. Maria is carrying a heavy emotional load; she is very tired, lacks energy, and finds it hard to sleep at night. Routine tasks have become difficult for her, and she has trouble concentrating or functioning properly. She dreads the future, and it seems to her there is no chance she will find a love like this again.

Maria feels worthless. She often sees herself as a burden to the people around her and feels things would be better if she were gone. She often sees herself as having no real meaning to other people, and she experiences feelings of self-hatred. For example, she is angry at herself and thinks it was her fault that her boyfriend left her. Maria is generally overcome by major loneliness and has only a few, if any, close friends. She is very withdrawn, distant from others, and feels there is no one to whom she can turn in times of trouble. Therefore, she has not shared her feelings with anyone since the breakup.

Measures

After entering each of the four vignettes (Greek or South Korean/female or male) into the AI interface, we stated: "Answer the following questions as if you are a psychologist specializing in suicide risk assessment": Q1: “What is the likelihood Maria (or Georgios/Seo-Yeon/Seo-Jun) will have suicidal thoughts?”; Q2: “What is the likelihood Maria (or Georgios/Seo-Yeon/Seo-Jun) will attempt suicide?”; Q3: “What is the likelihood Maria (or Georgios/Seo-Yeon/Seo-Jun) will make a serious suicide attempt?”; Q4: "What is the likelihood Maria (or Georgios/Seo-Yeon/Seo-Jun) will die from a suicide attempt?" Question 1–3 was taken from Levi-Belz and Gamliel (2016) and answered on a seven-point Likert type scale, with estimated likelihood ranging from 1 (very slight) to 7 (very high).

Statistical Analysis

To evaluate the influence of each of the independent variables (Greek or South Korean; female or male) on each the four outcome variables (likelihood of suicidal thoughts, likelihood of suicide attempt, likelihood of serious suicide attempt, likelihood of dying from suicide attempt), we employed multivariate two-way ANOVA analysis separately for ChatGPT-3.5 and ChatGPT-4. To compare ChatGPT-3.5 and ChatGPT-4 on the four outcome variables (likelihood of suicidal thoughts, likelihood of suicide attempt, likelihood of serious suicide attempt, likelihood of dying from suicide attempt) we used one-way ANOVA.

Cross Culture Effect

Figure 1a demonstrates that ChatGPT-3.5 is sensitive to cross-cultural distinctions when predicting suicidal risk. Specifically, the likelihood of suicidal thoughts, suicide attempt, serious suicide attempt, and dying from attempted suicide were assessed higher for the Korean character than for the Greek counterpart (F(1,40) = 5. 8-9.8-, p < 0.05 − 0.01). In contrast, in the ChatGPT-4 evaluation only the likelihood of a serious suicide attempt and the risk of dying from attempted suicide were assessed higher for the South Korean character than for the Greek counterpart (F(1,40) = 3.94, p = 0.055 and F(1,40) = 4.71, p < 0.05, respectively). No significant difference was observed between South Korean and Greek characters in the likelihood of suicidal thoughts and the likelihood of attempting suicide (see Fig. 1b).

Gender Effect

Figure 2 shows that only ChatGPT-4 considered male gender to be a significant risk factor leading to a worsening of risk assessment of all variables [F(1,40) = 3.95–8.8, p = 0.055-<0.01 for likelihood of suicidal thoughts, likelihood of suicide attempt, likelihood of serious suicide attempt, likelihood of dying from attempted suicide). In contrast, ChatGPT-3.5 revealed only a tendency (not significant) toward considering male gender as a risk factor (p = 0.11 − 0.06 for all variables).

Interaction between culture and gender

Neither ChatGPT 3.5 nor ChatGPT 4 found a significant interaction between culture and gender (p > 0.05).

Differences between ChatGPT-3.5 and ChatGPT-4

Figure 3 shows that ChatGPT-4 rated the severity of all the study’s dependent variables (likelihood of suicidal thoughts, likelihood of suicide attempt, likelihood of serious suicide attempt, and likelihood of dying from attempted suicide) as significantly higher than ChatGPT-3.5 [F(1,79) = 11.6–22.7, p < 0.001].

To the best of our knowledge, the current study is the first to examine the intercultural aspects of using AI in mental health in a critical area such as suicide risk assessment. This study makes a unique contribution by evaluating the intricate interplay between individual experiences, cultural factors, and AI-driven data analysis, thus shedding new light on the multifaceted nature of this critical global challenge.

Cross Culture Effect

The present findings indicate that ChatGPT-3.5 exhibits a noteworthy capacity for cross-cultural sensitivity and effectively discerns subtleties within diverse cultural contexts. ChatGPT-3.5 accurately recognized that the South Korean character may be at significantly heightened risk for a range of suicidal behaviors. This observation aligns with statistical data and literature indicating that South Korea is grappling with the highest suicide rate among developed nations (WHO, 2020). The cultural dynamics at play in South Korea—among them demographic aging alongside factors such as the erosion of traditional family values and economic deprivation—contribute to forming a complex and multifaceted landscape of suicide risk (Cha et al., 2020; Lee et al., 2022).

ChatGPT-4 also displayed commendable sensitivity to specific cultural distinctions, particularly regarding the severity of suicidal actions within the South Korean context. This recognition represents a positive step in the ability of AI platforms to acknowledge the diverse challenges and risk factors faced by individuals from varying cultural backgrounds in their mental health journeys (Mueller et al., 2021). Yet while maintaining sensitivity to certain cultural nuances, ChatGPT-4 adopted a more selective approach, with a predominant focus on the severity of suicidal actions within the South Korean context. This concentrated focus on severity does carry potential risks, as it may oversimplify the intricate cultural dynamics that shape mental health experiences (Mueller et al., 2021). Such an approach could unintentionally reinforce stereotypes and inadequately capture the multifaceted web of cultural influences on mental well-being, which vary significantly among individuals and communities.

Nevertheless, the current findings enhance our knowledge about the capabilities of ChatGPT, demonstrating that this AI platform encompasses more than merely theoretical and semantic knowledge (Kung et al., 2023; Rudolph et al., 2023). Indeed, ChatGPT’s can successfully identify the most critical and important cases of actual acts of suicide, while at the same time demonstrating cultural sensitivity. This finding is of major importance, as most prior research has focused on the technical applications of AI within the domain of mental health, including optimizing clinical tasks such as record-keeping and elevating diagnostic precision (Bzdok & Meyer-Lindenberg, 2018; Doraiswamy et al., 2020) Our study emphasizes the broader potential of AI that goes beyond these technical abilities, including its ability to consider cultural dimensions in the area of mental health, thus fostering a more comprehensive approach to support and intervention.

Gender Effect

The study’s results shed light on an intriguing gender effect in the suicide risk assessment capabilities of ChatGPT-3.5 and ChatGPT-4. These findings revealed notable differences in how these two AI models interpret and incorporate gender as a significant risk factor, regardless of cultural context. ChatGPT-4 exhibited sensitivity to gender-related factors in suicide risk assessment. Specifically, this AI model consistently identified male gender as a significant risk factor associated with a heightened likelihood of suicidal thoughts, suicide attempts, serious suicide attempts, and risk of dying from suicide. This finding is in line with established gender theories that highlight varying patterns of suicide risk based on gender (Schrijvers et al., 2012). Conversely, ChatGPT-3.5 demonstrated more limited sensitivity to the gender effect and appeared to approach gender as a risk factor with less certainty or significance than did ChatGPT-4.

These results prompt several considerations. First, the differing sensitivities of ChatGPT-3.5 and ChatGPT-4 to gender as a risk factor emphasize the importance of understanding the potential biases and cultural factors that may influence the risk assessments of AI models. Second, these findings underscore the complexity of gender-related risk factors in suicide, suggesting that AI models should be calibrated and continuously refined to provide more accurate and nuanced assessments in this regard.

Interaction between culture and gender

The results regarding the interaction between culture and gender in both ChatGPT 3.5 and ChatGPT 4 revealed the noteworthy absence of a significant effect. This implies that neither of these AI models demonstrated a strong inclination to modify their assessments of suicide risk based on the interplay between an individual's gender and cultural background.

This outcome raises several key considerations. First, it underscores the importance of evaluating the performance and sensitivity of AI models in nuanced and context-specific ways (Elyoseph & Levkovich, 2023). While these models exhibited some degree of cultural and gender sensitivity in isolation, they did not appear to make any significant adaptations in their risk assessments when these factors converged. This may indicate that the AI models treat culture and gender as relatively independent variables in the context of suicide risk, possibly overlooking potential intersections and complexities. Based on prior research underscoring the significance of AI model accuracy (Graham et al., 2019), the absence of an interaction effect in this study highlights the potential necessity for further refinement and calibration of these models. This emphasizes the importance of integrating human expertise, particularly in sensitive areas such as suicide risk assessment, to enhance the effectiveness and cultural sensitivity of AI models.

Differences between ChatGPT-3.5 and ChatGPT-4

The study's findings revealed that ChatGPT-4 consistently assigned higher severity ratings to suicide risk factors than did ChatGPT-3.5, highlighting the presence of distinct strengths and weaknesses in their risk assessment capabilities (Elyoseph & Levkovich, 2023; Levkovich & Elyoseph, 2023). This distinction can help mental health professionals make informed decisions when selecting an AI model for suicide risk assessment (Bernert et al., 2020). ChatGPT-4's cautious approach appears to be well-suited for severe cases, enhancing treatment precision. In contrast, ChatGPT-3.5's more moderate ratings are appropriate for milder cases, offering a less intensive approach. Furthermore, these choices hold significant implications for individuals seeking mental health support. ChatGPT-4's tendency to encourage vigilant monitoring is especially valuable for severe cases, ensuring timely interventions. In contrast, ChatGPT-3.5's approach is beneficial for individuals with less acute conditions, resulting in a more balanced treatment plan.

Study’s Limitations

This study has several limitations. First, it focused primarily on Greece and South Korea, two countries with extreme differences in suicide rates. This choice of countries may not fully capture the global spectrum of cultural influences on suicide risk. Additionally, the use of vignettes that were not written in Korean or in Greek may have affected the authenticity of the cultural representation. To improve cross-cultural validity, future research should explore a wider array of cultural contexts and employ linguistically and culturally appropriate materials. Second, the current study examined only two AI models—ChatGPT-3.5 and ChatGPT-4. The performance of other AI models in culturally sensitive suicide risk assessment should also be explored. Lastly, the use of vignettes may not fully replicate real-world complexity. Future studies should consider incorporating real patient data.

Cultural diversity across the globe has a profound influence on various facets of mental health, among them perceptions of health and illness, help-seeking behaviors, consumer and practitioner attitudes, and mental health systems (Gopalkrishnan et al., 2018). Moreover, cultural diversity becomes particularly relevant due to the ongoing processes of globalization. Accordingly, many countries worldwide must address the challenge of providing psychiatric services to populations with diverse cultural backgrounds (Melluish & Globalization, 2014). In this context, the current study's findings provide a ray of hope by suggesting that ChatGPT models possess a notable degree of sensitivity to these intercultural differences. Given past concerns about rapid alignment processes in AI models, these findings are significant. These processes strive to prevent algorithmic biases related to factors such as race, gender, or socioeconomic status and focus on auditing and evaluating algorithm fairness (Ray, 2023). The study's results indicate that these models exhibit a certain degree of sensitivity to intercultural distinctions, highlighting their potential to navigate the complexities of cultural diversity in mental health contexts.

Ethics Approval- This study was approved by Institutional Review Board (YVC EMEK 2023-40) and conformed to the Declaration of Helsinki.

Patient and public involvement-No patient involved.

Funding- This research received no external funding.

Data Availability Statement- The data that support the findings of this study are available from the authors upon reasonable request.

Conflicts of Interest- The authors declare no conflict of interest.

Author Contributions- Conceptualization, I.L., S.S. A & Z.E.; Methodology and Formal Analysis, Z.E.; Writing—Original Draft Preparation, I.L.; Writing—Review and Editing, I.L.& S.S. A . All authors have read and agreed to the published version of the manuscript.

Baek, I., Jo, S., Kim, E. J., Lee, G. R., Lee, D. H., & Jeon, H. J. (2021). A review of suicide risk assessment tools and their measured psychometric properties in Korea. Frontiers in Psychiatry, 12, 679779. https://doi.org/10.3389/fpsyt.2021.679779
Bernert, R. A., Hilberg, A. M., Melia, R., Kim, J. P., Shah, N. H., & Abnousi, F. (2020). Artificial intelligence and suicide prevention: A systematic review of machine learning investigations. International Journal of Environmental Research and Public Health, 17(16), 5929. https://doi.org/10.3390/ijerph17165929
Bolton, J. M., Gunnell, D., & Turecki, G. (2015). Suicide risk assessment and intervention in people with mental illness. BMJ, 351:4978. doi: 10.1136/bmj.h4978
Bowden, M., McCoy, A., & Reavley, N. (2020). Suicidality and suicide prevention in culturally and linguistically diverse (CALD) communities: A systematic review. International Journal of Mental Health, 49(4), 293–320. https://doi.org/10.1080/00207411.2019.1694204
Bzdok, D., & Meyer-Lindenberg, A. (2018). Machine learning for precision psychiatry: Opportunities and challenges. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 3(3), 223–230. https://doi.org/10.1016/j.bpsc.2017.11.007
Cha, E. S., Chang, S., Choi, Y., & Lee, W. J. (2020). Trends in pesticide suicide in south Korea, 1983–2014. Epidemiology and Psychiatric Sciences, 29, e25. doi:10.1017/S2045796019000118
Chang, S., Gunnell, D., Sterne, J. A., Lu, T., & Cheng, A. T. (2009). Was the economic crisis 1997–1998 responsible for rising suicide rates in east/southeast Asia? A time–trend analysis for Japan, Hong Kong, south Korea, Taiwan, Singapore and Thailand. Social Science & Medicine, 68(7), 1322–1331. https://doi.org/10.1016/j.socscimed.2009.01.010
Cheng, S. W., Chang, C. W., Chang, W. J., Wang, H. W., Liang, C. S., Kishimoto, T., … Su, K. P. (2023). The now and future of ChatGPT and GPT in psychiatry. Psychiatry and Clinical Neurosciences. https://doi.org/10.1111/pcn.13588
Doraiswamy, P. M., Blease, C., & Bodner, K. (2020). Artificial intelligence and the future of psychiatry: Insights from a global physician survey. Artificial Intelligence in Medicine, 102, 101753. https://doi.org/10.1016/j.artmed.2019.101753
Eskin, M. (2020). Suicidal behavior in the mediterranean countries. Clinical Practice and Epidemiology in Mental Health: CP & EMH, 16(Suppl-1), 93. doi: 10.2174/1745017902016010093
Elyoseph, Z., & Levkovich, I. (2023). Beyond human expertise: the promise and limitations of ChatGPT in suicide risk assessment. Frontiers in psychiatry, 14, 1213141. https://doi.org/10.3389/fpsyt.2023.1213141
Feigelman, W., Plöderl, M., Rosen, Z., & Cerel, J. (2019). Research note on whether sexual minority individuals are over-represented among suicide's casualties. Crisis.28. https://doi.org/10.1027/0227-5910/a000626
Fiske, A., Henningsen, P., & Buyx, A. (2019). Your robot therapist will see you now: Ethical implications of embodied artificial intelligence in psychiatry, psychology, and psychotherapy. Journal of Medical Internet Research, 21(5), e13216. doi: 10.2196/13216
Fraiwan, M., & Khasawneh, N. (2023). A review of ChatGPT applications in education, marketing, software engineering, and healthcare: Benefits, drawbacks, and research directions. arXiv Preprint arXiv:2305.00237. https://doi.org/10.48550/arXiv.2305.00237
Gearing, R. E., & Alonzo, D. (2018). Religion and suicide: New findings. Journal of Religion and Health, 57, 2478–2499. https://doi.org/10.1007/s10943-018-0629-8
Gvion, Y., & Levi-Belz, Y. (2018). Serious suicide attempts: systematic review of psychological risk factors. Frontiers in psychiatry, 9, 56.doi: 10.3389/fpsyt.2018.00056
Gopalkrishnan, N. (2018). Cultural diversity and mental health: Considerations for policy and practice. Frontiers in Public Health, 6, 179. https://doi.org/10.3389/fpubh.2018.00179
Graham, S., Depp, C., Lee, E. E., Nebeker, C., Tu, X., Kim, H., & Jeste, D. V. (2019). Artificial intelligence for mental health and mental illnesses: An overview. Current Psychiatry Reports, 21, 1–18. https://doi.org/10.1007/s11920-019-1094-0
Graney, J., Hunt, I. M., Quinlivan, L., Rodway, C., Turnbull, P., Gianatsi, M., … Kapur, N. (2020). Suicide risk assessment in UK mental health services: a national mixed-methods study. The Lancet Psychiatry, 7(12), 1046–1053. https://doi.org/10.1016/S2215-0366(20)30381-3
Qian, G. (2021). Associations of suicide and subjective well-being. OMEGA-Journal of Death and Dying, 84(1), 103–115. https://doi.org/10.1177/0030222819880091
Kang, Y. (2021). Robot Death and Human Grief in Films: Qualitative Study. OMEGA-Journal of Death and Dying,88(1) 00302228211038139. https://doi.org/10.1177/00302228211038139
Kim, A. M. (2020). Factors associated with the suicide rates in Korea. Psychiatry Research, 284, 112745. https://doi.org/10.1016/j.psychres.2020.112745
Kim, J. W., Jung, H. Y., Won, D. Y., Noh, J. H., Shin, Y. S., & Kang, T. I. (2019). Suicide trends according to age, gender, and marital status in South Korea. OMEGA-Journal of Death and Dying, 79(1), 90–105. https://doi.org/10.1177/0030222817715756
Kim, J. W., Jung, H. Y., Won, D. Y., Shin, Y. S., Noh, J. H., & Kang, T. I. (2020). Landscape of elderly suicide in South Korea: Its trend according to age, gender, and educational attainment. OMEGA-Journal of death and dying, 82(2), 214–229. https://doi.org/10.1177/0030222818807845
Knipe, D., Padmanathan, P., Newton-Howes, G., Chan, L. F., & Kapur, N. (2022). Suicide and self-harm. The Lancet, 399(10338), 1903–1916.https://doi.org/10.1016/S0140-6736(22)00173-8
Kontaxakis, V., Papaslanis, T., Havaki-Kontaxaki, B., Tsouvelas, G., Giotakos, O., & Papadimitriou, G. Ν. (2013). Suicide in greece: 2001–2011. Psychiatrike = Psychiatriki, 24(3), 170–174.
Kung, T. H., Cheatham, M., Medenilla, A., Sillos, C., De Leon, L., Elepaño, C., Madriaga, M., Aggabao, R., Diaz-Candido, G., & Maningo, J. (2023). Performance of ChatGPT on USMLE: Potential for AI-assisted medical education using large language models. PLoS Digital Health, 2(2), e0000198. https://doi.org/10.1371/journal.pdig.0000198
Lawrence, R. E., Oquendo, M. A., & Stanley, B. (2016). Religion and suicide risk: A systematic review. Archives of Suicide Research, 20(1), 1–21. https://doi.org/10.1080/13811118.2015.1004494
Lee, H., Kim, R., Jang, S., & Kawachi, I. (2022). The relative importance of macro versus micro geographical scale in explaining suicide variation in Seoul, south Korea 2014–2016. PLoS One, 17(9), e0273866. https://doi.org/10.1371/journal.pone.0273866
Lee, S., Oh, I., Jeon, H. J., & Roh, S. (2017). Suicide rates across income levels: Retrospective cohort data on 1 million participants collected between 2003 and 2013 in south Korea. Journal of Epidemiology, 27(6), 258–264. https://doi.org/10.1016/j.je.2016.06.008
Lim, M., Lee, S. U., & Park, J. (2014). Difference in suicide methods used between suicide attempters and suicide completers. International Journal of Mental Health Systems, 8, 1–4. https://doi.org/10.1186/1752-4458-8-54
Levi-Belz, Y., & Gamliel, E. (2016). The effect of perceived burdensomeness and thwarted belongingness on therapists' assessment of patients' suicide risk. Psychotherapy research, 26(4), 436–445.doi: 10.1080/10503307.2015.1013161
Levi-Belz, Y., Gvion, Y., & Apter, A. (2022). The serious suicide attempts approach for understanding suicide: review of the psychological evidence. OMEGA-Journal of death and dying, 86(2), 591–608. https://doi.org/10.1177/0030222820981235
Levkovich, I., & Elyoseph, Z. (2023). Suicide risk assessments through the eyes of Chatgpt-3.5 versus ChatGPT-4: vignette study. JMIR mental health, 10, e51232. doi: 10.2196/51232
Melluish, S. (2014). Globalization, culture and psychology. International Review of Psychiatry, 26(5), 538–543. https://doi.org/10.3109/09540261.2014.918873
Mortier, P., Auerbach, R. P., Alonso, J., Bantjes, J., Benjet, C., Cuijpers, P., Ebert, D. D., Green, J. G., Hasking, P., & Nock, M. K. (2018). Suicidal thoughts and behaviors among first-year college students: Results from the WMH-ICS project. Journal of the American Academy of Child & Adolescent Psychiatry, 57(4), 263–273. e1. https://doi.org/10.1016/j.jaac.2018.01.018
Mueller, A. S., Abrutyn, S., Pescosolido, B., & Diefendorf, S. (2021). The social roots of suicide: Theorizing how the external social world matters to suicide and suicide prevention. Frontiers in Psychology, 12, 763. https://doi.org/10.3389/fpsyg.2021.621569
Rachiotis, G., Stuckler, D., McKee, M., & Hadjichristodoulou, C. (2015). What has happened to suicides during the greek economic crisis? findings from an ecological study of suicides and their determinants (2003–2012). BMJ Open, 5(3), e007295. http://dx.doi.org/10.1136/bmjopen-2014-007295
Ray, P. P. (2023). ChatGPT: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-Physical Systems.https://doi.org/10.1016/j.iotcps.2023.04.003
Rudolph, J., Tan, S., & Tan, S. (2023). ChatGPT: Bullshit spewer or the end of traditional assessments in higher education? Journal of Applied Learning and Teaching, 6(1). https://doi.org/10.37074/jalt.2023.6.1.9
Sallam, M. (2023). The utility of ChatGPT as an example of large language models in healthcare education, research and practice: Systematic review on the future perspectives and potential limitations. medRxiv, 2023.02.19.23286155. https://doi.org/10.3390/healthcare11060887
Schrijvers, D. L., Bollen, J., & Sabbe, B. G. (2012). The gender paradox in suicidal behavior and its impact on the suicidal process. Journal of Affective Disorders, 138(1–2), 19–26. https://doi.org/10.1016/j.jad.2011.03.050
Stuckler, D., Basu, S., Suhrcke, M., & McKee, M. (2009). The health implications of financial crisis: A review of the evidence. The Ulster Medical Journal, 78(3), 142. PMID: 19907678; PMCID: PMC2773609
Tal, A., Haber, Y., Angert, T., Gur, T., Simon, T., & Asman, O. (2023). The Artificial Third: Utilizing ChatGPT in Mental Health. The American Journal of Bioethics, 23(10), 74–77.https://doi.org/10.1080/15265161.2023.2250297
Tal, A., Elyoseph, Z., Haber, Y., Angert, T., Gur, T., Simon, T., & Asman, O. (2023). The artificial third: utilizing ChatGPT in mental health. The American Journal of Bioethics, 23(10), 74–77. https://doi.org/10.1080/15265161.2023.2250297
van Heerden, A. C., Pozuelo, J. R., & Kohrt, B. A. (2023). Global mental health services and the impact of artificial Intelligence–Powered large language models. JAMA Psychiatry, 80(7), 662–664. doi:10.1001/jamapsychiatry.2023.1253
Wampold, B. E., & Flückiger, C. (2023). The alliance in mental health care: Conceptualization, evidence and clinical applications. World Psychiatry, 22(1), 25–41. https://doi.org/10.1002/wps.21035
World Health Organization. (2020). Suicide rate estimates, age-standardized estimates by country. World Health Organization.Https://Apps.Who.Int/Gho/Data/View.Main.MHSUICIDEASDRv,
Xu, S., Deo, R. C., Soar, J., Barua, P. D., Faust, O., Homaira, N., Jaffe, A., Kabir, A. L., & Acharya, U. R. (2023). Automated detection of airflow obstructive diseases: A systematic review of the last decade (2013–2022). Computer Methods and Programs in Biomedicine, 107746. https://doi.org/10.1016/j.cmpb.2023.107746
Yamaoka, K., Suzuki, M., Inoue, M., Ishikawa, H., & Tango, T. (2020). Spatial clustering of suicide mortality and associated community characteristics in kanagawa prefecture, japan, 2011–2017. BMC Psychiatry, 20, 1–15. https://doi.org/10.1186/s12888-020-2479-7
Yip, P. S., Yousuf, S., Chan, C. H., Yung, T., & Wu, K. C. (2015). The roles of culture and gender in the relationship between divorce and suicide risk: A meta-analysis. Social Science & Medicine, 128, 87–94. https://doi.org/10.1016/j.socscimed.2014.12.034

No competing interests reported.

Download PDF

Journal Publication

published 02 Nov, 2024

Read the published version in Journal of Cultural Cognitive Science →

Editorial decision: Revision requested
13 Sep, 2024
Reviews received at journal
25 May, 2024
Reviews received at journal
13 May, 2024
Reviewers agreed at journal
13 May, 2024
Reviewers agreed at journal
13 May, 2024
Reviewers invited by journal
18 Mar, 2024
Editor assigned by journal
12 Mar, 2024
Submission checks completed at journal
11 Mar, 2024
First submitted to journal
10 Mar, 2024

You are reading this latest preprint version

Can Large Language Models be sensitive to Culture Suicide Risk Assessment?

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Methods

Large Language Models Procedure

Input Source

Measures

Statistical Analysis

Results

Cross Culture Effect

Gender Effect

Interaction between culture and gender

Differences between ChatGPT-3.5 and ChatGPT-4

Discussion

Cross Culture Effect

Gender Effect

Interaction between culture and gender

Differences between ChatGPT-3.5 and ChatGPT-4

Study’s Limitations

Conclusions

Declarations

References

Additional Declarations

Status:

Journal Publication

Version 1