Conference Agenda

Background: Underreporting is one of the main challenges when monitoring suicide attempts in electronic health records (EHR). Some information is only recorded in text fields and, thus, escape detection when using diagnostic codes. Model-based approaches based on Natural language Processing (NLP) and machine learning algorithms might help increase overall precision by utilizing multimodal and unstructured data. Our aim was to examine whether model-based approaches identified more suicide attempts than conventional diagnostic approaches, to later assess multimorbidity and polypharmacy in non-reported suicide attempts.

Method: EHRs from 2006-2016 from the Capital Region and Region Zealand of Denmark were analyzed. Data consisted of complete EHRs, including laboratory results and pharmacy prescriptions, from 2.7 million individuals. Suicide attempts were identified using i) the conventional approach based on ICD-10 codes (X60-X84) and models based on: ii) custom semi-supervised neural network designed to leverage on multimodality, and iii) language, by conducting contextual and logical inference over narratives. A validation subset consisted of half of the suicide attempts identified by the conventional approach (n=5412) and suicide attempt with no diagnostic codes, which were identified by an expert panel (n=200). Numbers of identified suicide attempts were compared across the different approaches, and later risk assessment of multimorbidity and polypharmacy will be conducted.

Results: Approximately 14 million EHRs were processed. We found a significantly higher number of suicide attempts when using the model-based approaches instead of the conventional approach. Seemingly, >65% of all suicide attempts were not identified through the conventional approach. Analyses of multimorbidity are on-going. Final results of the analyses are available at the time of the conference.

Conclusion: Model-based approaches seem to have a higher precision and identify more suicide attempts, which allows for further investigations of multimorbidity and risk assessment of polypharmacy regarding suicide attempts. Thus, these options might be better suited for EHR monitoring.

Recent advancements in digital society are rapidly transforming not only our daily lives but also professional practices. In particular, the remarkable progress of artificial intelligence-related technologies , including generative AI, has enabled widespread applications in information retrieval, work efficiency, and consultation service. In this context, the Basic Act on Suicide Countermeasures in Japan was amended in June 2025 to include the "appropriate utilization of artificial intelligence-related technologies and others" thereby providing a policy basis for the use of generative AI in suicide countermeasures. This presentation outlines the "AI Chatbot for Local Government Suicide Countermeasures Officials" (provisional name), developed by Japan Suicide Countermeasures Promotion Center (JSCP) to bolster local efforts, and reports on findings and data from its early operational phase.

In Japan, local governments are obligated to lay down local suicide countermeasure plans that consider the individual local suicide situation, and local governments play a vital role as a safety net that directly supports residents’ lives and living. However, because local officials rotate frequently, local governments have faced the challenge of sustaining and transferring knowledge and experience in suicide countermeasures. To address this challenge, JSCP developed the AI Chatbot that enables local officials—including those newly appointed with limited experience—to access necessary knowledge immediately. The AI Chatbot is trained on relevant laws, guidelines, and best practices in local government suicide countermeasures, enabling local officials to access highly specialized expertise immediately. Furthermore, by analyzing the questions (prompts) submitted by local officials to the AI Chatbot, it becomes possible to identify their immediate challenges and specific support needs.

Based on early operational insights from the AI Chatbot developed by JSCP, this presentation presents the potential applications and key considerations of generative AI for strengthening local suicide countermeasures and shares them with international stakeholders engaged in suicide countermeasures.

Introduction

The rapid adoption of conversational AI chatbots for mental health support has outpaced the development of evidence-based standards for suicide risk detection and mitigation. Recent estimates suggest substantial uptake among adolescents and adults seeking mental health advice from AI chatbots, including 1 in 8 U.S. teens and young adults and more than 1 in 3 adults in the (McBain et al, 2025; Mental Health UK, 2025). From a suicide prevention perspective, this growing use heightens the need for validated safety benchmarks.

Methods

This presentation examines key methodological challenges in defining and evaluating the adequacy of AI chatbot responses to suicide risk. We synthesize emerging usage estimates, publicly reported safety metrics from major chatbot developers, and findings from 500+ suicide-focused conversations across leading LLMs evaluated by mpathic.ai. We report on recent developments in integrating clinician led suicide safety in AI and perspectives of individuals with lived experience using chatbots while suicidal.

Results

Disclosure of suicidal ideation or planning during routine chatbot use is significant. In October 2025, OpenAI reported detecting explicit indicators of potential suicide planning or intent in ~ 0.15% of weekly users (≈1.2 million individuals), with an additional ~0.05% (≈400,000 individuals) disclosing explicit or implicit suicidal ideation. Extended conversations, ambiguous disclosures, and adversarial inputs complicate both risk detection and response evaluation estimates. Ever changing LLM models and lack of cross-platform consistency will make validating benchmarks for mitigation adequacy an ongoing effort.

Conclusions

Advancing safe and ethical use of AI chatbots in suicide-related contexts requires research frameworks that are clinically grounded, transparent, empirically testable, and current. Testing for risk detection and effective protective responses across diverse groups will be challenging, especially as technology changes rapidly. Emerging efforts to standardize suicide risk reporting and mitigation assessment (e.g., Throughline) represent early steps toward responsible clinical and public health applications.

Introduction. Suicide-related conversations with general-purpose conversational AI (chatbots), including tools such as ChatGPT, are an emerging feature of help-seeking. These interactions may precede, accompany, or substitute for contact with formal support services. Although such systems sit outside regulated therapeutic and crisis care, people with lived experience of suicidal thoughts report using them during periods of distress as a low-threshold form of support-seeking.

Current safeguarding approaches prioritise detection and signposting, yet little is known about how users understand, engage with, and evaluate these interactions. This gap risks misalignment between real-world use, platform safety design, and suicide prevention and mental health care pathways. This study aims to develop an evidence-informed framework describing patterns of use and tensions between perceived support and safety, to inform guidance for design, governance, and use.

Methods. We undertook a scoping review to map academic, policy, and regulatory literature relating to chatbots in suicidal and upstream risk contexts. Building on this review, we will run two online focus groups with people with lived experience of suicidal thoughts (6–8 participants each). Discussions will explore how AI is used during distress or crisis, focusing on perceived functions and limitations, experienced benefits and harms, and expectations of responsibility across users, services, and developers. Data will be synthesised to develop a relational risk framework to describe and link user motivations, perceived AI functionality, and safety trade-offs.

Results. Preliminary findings will integrate thematic insights from focus groups with evidence from the scoping review to examine motivations for AI use during suicidal distress, alongside perceived benefits and risks related to disclosure, autonomy, and help-seeking.

Conclusions. Suicidal conversations with AI represent an emerging and under-examined phenomenon. Rather than presuming benefit or harm, this study centres lived-experience perspectives to clarify implications for support pathways, safety, and responsibility, and to inform stakeholder-specific recommendations.

Background

There has been rapid expansion in the development of machine learning algorithms to predict suicidal behaviours. To test the accuracy of these algorithms for predicting suicide and hospital-treated self-harm, we undertook a systematic review and meta-analysis.

Methods and findings

We searched PubMed, PsycINFO, Scopus, EMBASE, IEEE, Medline, CINALH and Web of Science from database inception until 30 April 2025 to identify studies using machine learning algorithms to predict suicide, self-harm and a combined suicide/self-harm outcome.

Findings

Fifty-three studies met the inclusion criteria. The area under the receiver operating characteristic curves ranged from 0.69 to 0.93. Sensitivity was 45%–82% and specificity was 91%–95%. Positive likelihood ratios were 6.5–9.9 and negative likelihood values were 0.2–0.6. Using in-sample prevalence values, the positive predictive values ranged from 6% to 17%. Using out-of-sample prevalence values at an LR+ value of 10, the positive predictive value was 0.1% in low prevalence populations, 17% in medium prevalence populations and 66% in high prevalence populations.

Conclusions

The accuracy of machine learning algorithms for predicting suicidal behaviour is too low to be useful for screening (case finding) or for prioritising high-risk individuals for interventions (treatment allocation). For hospital-treated self-harm populations, management should instead include three components for all patients: a needs-based assessment and response, identification of modifiable risk factors with treatment intended to reduce those exposures, and implementation of demonstrated effective aftercare interventions.

Background: Most research using Google Trends to examine suicide-related search patterns relies on researcher-selected terms rather than those actually used by people at risk, limiting ecological validity. Additionally, no studies have specifically examined associations between volume of suicide-related searches and number of suicides by mental health patients, despite this population's elevated risk and prevalent suicide-related internet use (SRIU).

Methods: This study analysed suicide-related search terms provided by 196 people who had used mental health services in the UK and had engaged in SRIU. From 520 search entries, 36 terms were identified across seven categories using a mixed methods approach. Monthly Google Trends data (2011-2022) for these terms was obtained. Transfer function modelling with ARIMA/SARIMA models examined cross-correlations between monthly search volumes and the number of suicide deaths by both mental health patients and the general population.

Results: Of 448 tested lag correlations, 18 (2.0%) were statistically significant, showing small to medium effect sizes. Most significant associations involved non-specific suicide-related queries (33.3%), characteristics of suicide (22.2%), and help-seeking queries (22.2%). No pro-suicide queries showed significant associations. Concurrent positive correlations were found between search volumes for suicide ideation terms and suicide deaths by mental health patient, and between help-seeking terms and suicide deaths in the general population.

Discussion: The study reveals modest associations between suicide-related search patterns and actual suicide deaths. The associations differed between mental health patients and the general population. The absence of pro-suicide query associations may reflect effective online prevention efforts in the UK. However, the small number of significant correlations indicates limited predictive utility for population-level suicide monitoring, supporting conclusions that Google Trends data alone is insufficient for predicting trends in suicide, though valuable for generating hypotheses about suicide-related internet behaviours.