Suicide Prevention Using Artificial Intelligence: Collaborative Support Approach

The World Health Organization (WHO) asserts that suicide is a significant global health problem. In 2016, the suicide rate was estimated to be 10.6 per 100,000 individuals, with 80% of these cases occurring in low-income and middle-income countries (Fazel et al., 2020). Often, individuals at risk of suicide do not seek help from their clinicians or communities due to fear of stigmatization and the possibility of forced medical treatment. Furthermore, individuals with mental illnesses—who represent a majority of suicide cases—may have limited awareness of their mental condition and may not recognize themselves as being at risk of suicide (Picardo et al., 2020). This situation is further complicated by the difficulty clinicians or psychotherapists face in accurately identifying those at risk of suicide when they do seek medical care (Walsh et al., 2018).

Epidemiological studies have also shown that young adults’ ages 15 to 21 years have the highest prevalence of mental illness leading to suicide, with a rate of 39% (Eisenberg et al., 2007). Psychiatric conditions that are often associated with suicidal ideation and behavior include depression, anxiety, substance use disorders and eating disorders (Bradyik, 2018). Warning signs of suicidal ideation may include a prior suicide attempt or intentional self-harm behaviors, such as cutting or burning oneself (Ryan & Oquendo, 2020).

In an effort to mitigate the impact of suicide, there is a growing interest in leveraging artificial intelligence (AI), data science, and other analytical techniques to enhance suicide prediction and risk identification. Broadly, these tools fall into two categories: medical suicide prediction tools and social suicide prediction tools. Medical suicide prediction tools involve clinicians and psychotherapists using AI techniques (i.e., natural language processing and machine learning) to identify patterns of information and behavior indicative of suicide risk, and to utilize data from electronic medical records and potentially other government data sources (Nugent et al., 2019).These tools are typically employed in hospital settings or general practitioner clinics to support clinicians or psychotherapists with assessing patient suicide risk. Social suicide prediction tools, on the other hand, involve AI and data tools that analyze information from social media and browsing habits to assess suicide risk. Platforms like Facebook, Google, and Apple, for example, may use data to identify users at risk of suicide and then provide appropriate interventions, such as offering free information and counseling services (Coppersmith et al., 2018; Muriello et al., 2018).

While it may not be possible to completely eradicate suicide, enhancing predictive and preventative measures through advanced analytical tools may offer the best potential for improved outcomes. However, predicting suicide risk remains challenging for traditional epidemiological studies and healthcare providers due to the complex factors involved and the difficulties with identifying a small number of at-risk individuals within a large population of individual who share similar risk factors. Lejeune et al. (2022) conducted a review to evaluate the potential of AI in identifying patients at-risk of attempting suicide. They performed a systematic review of literatures using the PubMed, EMBASE, and SCOPUS databases, with relevant keywords. Out of 296 studies identified, 17 studies published between 2014 and 2020 met the inclusion criteria and were deemed relevant. These studies focused on predicting individual suicide risk or identifying at-risk individuals within specific populations. Overall, the performance of AI was found to be good, though it varied across different algorithms and application settings. The review concluded that AI holds significant promise for identifying patients at-risk of suicide; however, the exact application of these algorithms in clinical practice, along with the ethical issues they raise, still needs further clarification and research evidence.

Approaching Suicide Prevention Using AI as a Clinician

A landmark meta-analysis by Franklin et al. (2017), which examined 365 studies over a 50-year span, revealed that suicide prediction is only slightly better than chance for all outcomes and this predictive ability has not improved over the last five decades. This review found that psychotherapists are challenged by the fact that many individuals who die by suicide never disclose their suicidal thoughts to a healthcare provider. Those experiencing suicidal ideation often fear discussing their thoughts with friends or family due to concerns about being judged, hospitalized, or medicated (Franklin et al., 2017).

Despite these challenges, a longitudinal study found that 83% of individuals who die by suicide had contact with health services within the year preceding their death, and 45% had contact within the month before (Sheehan et al., 2017). This indicates a significant opportunity for medical prediction tools to assist clinicians in assessing suicide risk when these patients seek care. Consequently, these prediction tools should move away from focusing solely on risk factors and instead utilize machine learning algorithms and data science to predict suicide risk using innovative analytical methods.

For instance, Kessler et al. (2017) applied machine learning protocols—such as Naive Bayes, random forests, and support vector regression—to predict suicide deaths among military veterans within 26-weeks of an outpatient mental health visit. The study reported an area under the curve (AUC) of 0.72 for those with a prior psychiatric hospitalization, 0.61 for those without such hospitalization, and 0.66 when both groups were combined. This implies that there was a 72% chance that the model would correctly rank a randomly chosen patient who completed suicide higher than a randomly chosen person who did not. Further, the model had only a 61% chance of correctly ranking a patient who completed suicide higher than someone who did not. However, the model was able to correctly rank suicide death cases over non-suicide cases 66% of the time. This moderate AUC reflects an overall better performance than random guessing but suggests that the model’s predictive accuracy is still limited, particularly when both groups (hospitalized and non-hospitalized) are considered together. Overall, the model performed better in predicting suicide deaths for patients with a prior psychiatric hospitalization (AUC of 0.72) than for those without (AUC of 0.61). When considering both groups together, the model’s predictive ability averaged out to an AUC of 0.66, indicating moderate performance with room for improvement. Relevant factors included suicidality, depression, bipolar disorder, and non-affective psychosis (Kessler et al., 2017). Interestingly, the AUC improved to 0.75 when predicting suicide death within five weeks of the outpatient visit.

Similarly, a study by DelPozo-Banos et al. (2018) utilized artificial neural networks, a type of machine learning technique, to analyze routinely collected data in electronic medical records (EMRs) to assess suicide risk in patients visiting health services for any reason. By analyzing EMR and hospital data from the 5years prior to a patient’s suicide, the model accurately distinguished between control patients and those who died by suicide, achieving an accuracy of over 73%. However, it is important to note that more complex models incorporating additional data points could likely yield even better results, and the researchers plan to develop such a model in the next phase of their research (DelPozo-Banos et al., 2018).

Artificial intelligence has also demonstrated high accuracy in predicting suicide attempts. By applying machine learning to electronic health records (EHRs), Walsh et al. (2017) developed machine-learning algorithms (random forest and logistic regression) that achieved AUC values of 0.80 and 0.84 when predicting the likelihood of a suicide attempt within the next 2years and within the next week, respectively. Important predictors in both long- and short-term predictions included depression with psychosis, schizophrenia, and a history of prior suicide attempts. In the study conducted by Walsh and colleagues (2017), they analyzed patient data from the records of 5,167 adults treated at Vanderbilt University Medical Center. The study reported the accuracy of their suicide prediction models in terms of area under the curve (AUC), where an AUC of 0.5 indicates “accuracy no better than chance,” and an AUC of 1.0 represents perfect accuracy (Centers for Disease Control and Prevention, 2018). Given that traditional suicide prediction methods may be only slightly more accurate than a coin flip (approximately 50% or 0.50 probability), the study’s findings are noteworthy. For patients attempting suicide for the first time, Walsh and colleagues (2017) reported AUC values ranging from 0.82 at 7 days prior to suicide attempts to 0.75 at 720 days prior to suicide attempts (VA Releases National Suicide Data Report, 2018).

Ryu et al. (2019) also employed a machine learning technique (random forest) to predict suicide attempts among individuals with suicidal ideation. Their prediction model achieved impressive results, with an AUC of 0.947 and an accuracy of 88.9%. While the clinical applicability of these tools in real-world settings has yet to be fully proven, the initial results are highly promising.

In all, an important question that arises is what actions should be taken when individuals are identified as being at-risk of suicide? For some patients, hospitalization may be the appropriate step, but for others, hospitalization could potentially cause more harm than good. Additionally, forcibly detaining patients in a hospital or other medical setting could lead to significant psychological distress and may even hasten future suicide attempts (Jackman & Kanerva, 2017).

Suicide Prevention Using Social Evidence Approach of AI

In a study, Gaur et al. (2019) analyzed Reddit posts for signs of suicidal language to assess suicide risk. They compared different clinical classification schemes against machine learning techniques, including random forest and convolutional neural networks. Convolutional neural networks outperformed the others, achieving an overall precision of 70%, which was 40% better than baseline approaches that relied solely on medical classification systems (Gaur et al., 2019).

Earlier, on April 2, 2018, Zuckerberg revealed that Facebook’s AI scans the content of users’ private messages, suggesting that both public and private user-generated content may be monitored for signs of suicidal intent (Meyer, 2014). On September 10, 2018, Facebook provided additional details about its suicide prediction algorithms. Using an AI tool called random forests, Facebook analyzes user-generated content and assigns a risk rating to specific words, word pairs, and phrases in each post. Hypothetical examples provided by the company include terms like “sadness,” “much sadness,” and “so much sadness.” Unlike medical suicide prediction, being primarily experimental, requiring approval from Institutional Review Boards (IRB), and resulting in peer-reviewed publications in academic journals, Facebook’s suicide prediction program does not undergo independent ethics reviews (Muriello et al., 2018).Additionally, Facebook’s methods and results are not published or made publicly available, raising concerns about safety, accountability, and transparency. Instead of consulting an IRB, Facebook will use an internal ethics board. Unlike mandatory IRB approval at a hospital or university, however, the review of Facebook’s projects by its ethics board occurs at the company’s discretion. This lack of transparency and accountability is troubling, given Facebook’s history of monitoring users’ emotional states and conducting experiments on users without their knowledge or consent (Muriello et al., 2018).

Coppersmith et al. (2018) utilized natural language processing, along with supervised and unsupervised machine learning methods, to analyze social media data from various platforms, including Instagram, Facebook, Twitter/X, Strava, Fitbit, Reddit, and Tumblr, with permission from the test subjects, to assess the risk of suicide attempts. The model achieved an AUC of 0.89–0.93 for time frames ranging from 1 to 6 months. According to Coppersmith et al. (2018), if a false alarm rate of 1-2% is assumed, this model could be up to 10 times more accurate in predicting suicide attempts compared to clinician averages (40–60% vs. 4–6%).

Lissak et al. (2024) employed AI methodologies to uncover hidden risk factors that contribute to or exacerbate suicidal behaviors. The primary dataset comprised 228,052 Facebook posts from 1,006 users who completed the Columbia Suicide Severity Rating Scale. The secondary dataset included responses from1, 062 participants who completed the same suicide scale, along with well-validated scales measuring depression and boredom. The results revealed that an almost fully automated, AI-guided research pipeline identified four Facebook topics that predicted suicide risk, with boredom emerging as the strongest predictor (Lissak et al., 2024). Interestingly, a comprehensive literature review using APA PsycInfo indicated that boredom is rarely considered a unique risk factor for suicide. Analysis of the secondary dataset revealed an indirect relationship between boredom and suicide, mediated by depression (Parsapoor et al., 2023). A similar mediated relationship was observed in the primary Facebook dataset, where a direct relationship between boredom and suicide risk was also found. The integration of AI methods enabled the discovery of an under-researched risk factor for suicide: boredom. This study highlights boredom as a potentially maladaptive factor that could trigger suicidal behaviors, independent of depression. Further research is recommended to draw clinicians’ attention to this burdensome and sometimes existential experience.

Future of Using Artificial Intelligence in Suicide Prevention

Study results consistently demonstrate that AI can outperform doctors in predicting both suicide completion and suicide attempts, underscoring the potential of AI-based medical suicide prediction. The research indicates a promising clinical application for AI in identifying the risk of suicide completion.

Implement Natural Language Processing (NLP) tools

It is more important for all clinicians and healthcare settings to begin utilizing NLP to analyze patient communications, such as journal entries or therapy session transcripts, for signs of suicidal ideation. This technology can identify risk factors that may not be captured in traditional screening tools like the PHQ-9 (Zaubler, 2024). For instance, studies have shown that NLP can detect suicidal thoughts in over half of patients who might otherwise go unnoticed, allowing clinicians to intervene more effectively (Gliadkovskaya, 2024). By integrating clinical data, social determinants, and behavioral indicators, these models can provide timely alerts to clinicians about patients at-risk, facilitating early intervention. This approach has been successfully implemented in youth suicide prevention, demonstrating its potential to save lives (Pediatric Health Advances, 2023).

Enhance clinical decision support systems

It is crucial to integrate AI into clinical decision support systems to assist clinicians and psychotherapists in assessing suicide risk. AI can analyze vocal biomarkers and speech patterns to provide real-time feedback during therapy sessions, helping clinicians make informed decisions about patient care (Thompson, 2023). Such systems can alert providers when a patient exhibits concerning signs, ensuring a quick response that is central to the intervention process (Gliadkovskaya, 2024).

Foster collaborative care models

Health institutions should encourage collaboration between AI systems and human clinicians to create a holistic care approach. AI can handle data analysis and risk assessment, while clinicians can focus on building therapeutic relationships and providing personalized care. This partnership can enhance the overall effectiveness of suicide prevention strategies, as clinicians are better equipped with actionable insights from AI.

Utilize AI for resource allocation

Employing AI to optimize the allocation of mental health resources, particularly in underserved communities, is very important. By analyzing data on suicide rates and mental health service availability, AI can help identify areas with the greatest need for intervention. This targeted approach ensures that resources are directed to where they have the most impact, addressing disparities in mental health care access (Gliadkovskaya, 2024).

Ensure ethical and secure data usage

Establish strict ethical guidelines and robust cybersecurity measures for using AI in suicide prevention (Onah et al., 2024). As AI relies on sensitive patient data, it’s crucial to protect this information and maintain patient confidentiality (European Parliament, 2023). Clear legal frameworks should be developed to guide the responsible use of AI in clinical settings, ensuring that ethical considerations are prioritized alongside technological advancements. Further, population-wide suicide prediction may offer an ethical and effective application for AI, helping policymakers and medical professionals allocate healthcare resources more efficiently.

While AI in suicide prevention may be much more advanced than traditional methods, current medical suicide prediction models still yield a considerable number of false positives and false negatives. Consequently, these tools are primarily used for research purposes and are not yet fully employed to guide clinical decision-making in most healthcare settings. Although advances in AI present significant opportunities for developing novel tools to predict suicide, further evidence suggests that the combined medical and social suicide prediction tools could enhance our ability to identify individuals at-risk of suicide and potentially save lives. By implementing these strategies, the integration of AI into suicide prevention efforts can enhance the ability of clinicians to identify at-risk individuals and provide timely, effective, life-saving interventions.