The prevalence of suicidal thoughts among young adults, those between the ages of 18 and 25 years, in the United States has increased by approximately 160% since 2013 [1,2]. While most young adults who experience suicidal thoughts do not go on to make a suicide attempt, suicidal thoughts are indicators of risk that warrant attention and intervention [3]. Among adults who experienced mental health symptoms in the last year, young adults have the lowest rates of formal mental health care treatment [2]. Structural (eg, cost, accessibility, and provider availability) and attitudinal (eg, stigma, strong preferences for self-management) barriers are particularly important for determining help-seeking behavior [4] and contribute to low rates of mental health treatment engagement [4-6].

Current approaches to suicide prevention underserve young adults facing structural or attitudinal barriers to mental health services. Most prevention efforts are delivered in the context of mental health interventions via formal outpatient, inpatient, emergency department, and other clinical care settings. Prevention is also delivered in informal settings such as crisis helplines, targeting individuals before suicidal crises arise (eg, in schools or workplaces) in an effort to bring suicide prevention approaches to those who need it in places they are most likely to access it [7]. As most young adults with suicide-related thoughts and behaviors do not engage in formal mental health treatment, where most evidence-based suicide prevention is delivered [8-15], and they experience strong preferences for self-management, young adults are less likely to encounter evidence-based suicide prevention interventions.

The high need for preventive interventions and low formal mental health treatment engagement among young adults experiencing suicidal thoughts and behaviors [1,2] signal a critical gap in suicide prevention infrastructure. Digital suicide prevention interventions are an important avenue by which access to evidence-based suicide prevention care can be increased. Digital suicide prevention interventions offer the benefit of being able to circumvent many of the attitudinal and structural barriers that prevent young adults from accessing evidence-based preventive interventions [4-6]. They can be accessed on young adults’ devices (eg, smartphones) and used privately.

The safety planning intervention (SPI) [16], a brief intervention aimed at interrupting the progression from suicidal ideation to suicidal behavior by increasing coping ability, has typically been administered collaboratively by clinicians in clinical care settings across ages [17-19] and is considered standard of care [20,21]. The SPI typically includes 6 areas of focus: warning signs, coping skills, people for distraction, people for help, professional organizations and resources, and environmental safety [16]. The product of the SPI is often a brief document with personalized lists of resources and coping activities for each of the 6 focus areas. When delivered with fidelity, clinicians ask patients targeted questions and follow-up on patients’ responses to elicit highly personalized information that can enhance the efficacy of an individual’s safety plan by making it as relevant as possible to the individual’s life circumstances [22]. However, safety planning is frequently delivered by individuals facing barriers to implementation including insufficient training and time [23,24], which poses a risk to safety plan quality and efficacy.

The SPI is one of the most common digital suicide prevention intervention elements [25]. There is an emerging literature on automated digital SPIs that guide users through the creation of a highly personalized safety plan [25]. Several have been shown to be acceptable [26-29] and even demonstrate that resultant safety plan quality can be comparable to collaboratively developed safety plans [30,31]. Extant literature on digital SPIs primarily focuses on broad adult samples and narrow adolescent populations [25]. Boudreaux et al [26] piloted an automated, computer-based SPI that enabled adults who experienced suicide-related thoughts and behaviors to complete a safety plan and offered brief guidance on how to complete the safety plan. The resultant safety plan was a 1-page PDF that users could download, access on different devices, and share with trusted others. Melvin et al [27] and Rainbow et al [29] developed a mobile app–based approach called BeyondNow that includes a customizable SPI with prepopulated suggestions for each of the 6 safety planning steps and offers the ability to edit or share the completed safety plan and a quick access button to a crisis helpline. The app is publicly accessible and has strong acceptability metrics and is associated with promising improvements in suicide-related coping. Finally, Methi et al [28] created an automated web-based e-SPI specifically designed for adolescents that includes age-appropriate instructions, examples, and illustrations that assist users in completing the online safety plan via text-fill boxes. The content adolescents add is converted to an online safety card that users can save. These efforts highlight that automated safety planning is feasible to implement and that there are opportunities to engage individuals in new ways, expanding beyond single-session safety planning and crafting safety plans via mediums that are more conversational in nature.

Among this emerging literature, several key needs and preferences have surfaced. First, having quick access to information such as one’s documented safety plan, crisis resources, or personal contact information can be extremely valuable during a crisis as crisis states make it difficult to navigate complex user interfaces [32,33]. Second, there is a favorable view of digital artifacts (eg, photos, songs) that inspire hope and increase motivation for living [27]. Third, across the literature, users of digital SPIs have ambivalence about whether or how they want to involve others in their personal networks in the creation or use of their safety plans [32,33]. One way early digital SPIs have addressed this issue is through the use of optional prescripted messages users can generate when outside of crisis states and send to trusted others to elicit social or instrumental support [32]. Fourth, extant literature highlights that strong privacy practices emerge as critical priorities for users of digital safety plans [33,34]. Fifth, the ability to download and edit created safety plans after their initial creation is critical for ensuring safety plans remain relevant and accessible over time [27,34,35]. Finally, using validating language that minimizes assumptions about a user (eg, that they have a network of others that can provide support during a crisis) and ensuring the digital SPI can provide examples or menus of items if a user has difficulty generating safety planning content independently can prove beneficial [35].

The vast majority of digital SPIs are mobile app or browser based [25,36], meaning that individuals must visit a specific app or website in order to interact with or create their safety plan. In this study, we aimed to understand the potential of an SMS text messaging–based safety plan creation approach. Texting has the technological affordance of being a conversational medium, potentially addressing some of the shortcomings of automated digital safety plans that are documentation only. By eliciting information relevant to safety planning in a conversational manner, an automated texting–based SPI may be able to increase the personalization and relevance of digital safety plans. SMS text messaging is the most commonly used communication medium and young adults are the age group sending the greatest number of SMS text messages [37,38]. It is more widely used than social media apps [39]. Therefore, an SMS text messaging–based SPI is capable of meeting young adults in a space where they are already spending a significant amount of time. Indeed, prior work has shown that texting interventions for suicide prevention have been acceptable to users [40-45].

The relatively sparse existing literature on the needs and preferences of digital SPI users is focused on the app- or web-based designs. Further, it lacks focus on the needs of the adult age group with the highest rates of suicide-related thoughts and behaviors, but the lowest rates of treatment-seeking: young adults. Moreover, there is a paucity of literature regarding automated, interactive, texting interventions to support suicide prevention, broadly, and safety planning, more specifically.

In this study, we aimed to understand the needs, preferences, constraints, and design considerations for an SMS text message–based automated safety planning tool for young adults who have lived experience of suicide-related thoughts and behaviors. Informed by prior literature, our work was guided by the following research questions: (1) What are the broad needs and preferences of young adults who experience suicide-related thoughts and behaviors for suicide prevention? (2) What are the features that can support young adults with suicide-related thoughts and behaviors to create an effective safety plan?

A total of 30 participants between the ages of 18 and 24 years were enrolled in this study (Table 1). Individuals were eligible to participate if they resided in the United States, were between the ages of 18 and 24 years, owned a smartphone, were proficient in English, and endorsed past 2-week suicidal ideation. Suicidal ideation was assessed using a modified item from the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition) Cross-Cutting Symptom Inventory (DSM-5-CCSI) [46], “During the past 2 weeks, how often have you been bothered by thoughts of actually hurting yourself [or ending your life]?” Disambiguation of suicidal thoughts from nonsuicidal self-injurious thoughts was assessed via the Depression Severity Index—Suicidality Subscale (DSI-SS) [47], a 4-item self-report measure assessing the frequency and intensity of suicidal thinking. Participants were not eligible to participate if they endorsed symptoms consistent with mania or psychosis in the last 2 weeks (assessed via the DSM-5-CCSI) or the presence of strong suicidal desire assessed via item 9 on the Beck Depression Inventory—Second Edition [48] and a subsequent score of 2 or 3 (range 0-3).

Individuals were recruited via an advertisement embedded on Mental Health America’s (MHA) website. The advertisement was shown to individuals who visited a resource page about suicidal thoughts and behaviors, endorsed suicidal or self-harm ideation in the past 2 weeks on MHA’s online screening website, or endorsed past-month suicidal ideation on MHA’s nonsuicidal self-injury survey. Individuals who clicked the ad were directed to a Northwestern University–hosted informational web page about the study. Interested individuals could elect to provide informed consent for eligibility screening via Northwestern University’s instance of REDCap [49,50] and were subsequently routed to the eligibility screening surveys.

If eligible, interested individuals could review the REDCap-based study consent documentation and the code of conduct for engaging with research staff and fellow participants on the private research forum. The code of conduct detailed guidelines such as refraining from posting identifying information or detailed descriptions of self-harm, suicide methods, or behaviors. The code of conduct also overviewed harassment and moderation policies. To be enrolled in the study, individuals needed to provide affirmative informed consent and agree to abide by the code of conduct.

Participants engaged in 1 of 2 asynchronous remote community (ARC) [51-53] groups—akin to an online focus group. We used ARC to increase the engagement of underrepresented populations by decreasing barriers related to synchronous or in-laboratory research tasks [54]. Additionally, the group format was chosen to facilitate conversation between adults with shared experiences. ARC methods are used in the field of human-computer interaction to facilitate conversation, elicit several perspectives simultaneously, and provide the environment for participants to build on each other’s creative contributions [51]. ARC groups were held between November and December 2022, or February and March 2023. Two ARC groups were held to increase the sample size and diversity of responses to several lines of questioning and protect against group effects as well as to gather data on lines of inquiry that were not covered in the first group. ARC groups involved participants registering for a private online forum using FocusGroupIt [55] in which researchers and participants could create posts visible for all in the study and with which participants could engage or respond. Each group comprised 15 participants. Past work demonstrates similar sample sizes are sufficient to elicit diverse perspectives [56]. Participants were provided with automatically generated anonymous usernames (eg, participant 1, participant 2). Each ARC group consisted of 8 total prompts, 1 every 3 days. Prompts focused on terminology young adults used for suicidal ideation, prompting events for suicide-related thoughts and behaviors, safety plan use, and technology to support both safety planning and help-seeking.

Research staff monitored forums daily for posts violating the posted code of conduct. Offending posts were deleted, and the original poster was contacted by research staff with an explanation of why the post was removed, and, if needed, to conduct a risk assessment. Only 1 such post was removed and there was no indication of suicide-risk elevation for the poster.

The corpus for this study included 3 prompts from each group containing relevant data to answering the primary research questions (see Multimedia Appendix 1). These prompts comprising the corpus covered perceptions of safety planning via SMS text messages (group 1, prompt 4; group 2, prompt 3), the role of automated messages in supporting safety planning (group 1, prompt 5; group 2, prompt 4), use of a safety plan after it is created (group 1, prompt 7), and the role of crisis services in suicide prevention care (group 2, prompt 7).

Forum transcripts were analyzed using Thematic Analysis [57,58] to extract key needs, preferences, and design considerations for a digital safety planning tool. Thematic analysis was facilitated using Dedoose software (Version 9.0.9) [59]. One coder (JM) read through the corpus and generated an initial codebook via open coding responses to each of the 3 prompts from each ARC group. Coding occurred at the sentence level within a participant’s post. Two additional coders (SAP and TL) read through the entire corpus to familiarize themselves with the data and then applied the initial codebook to overlapping transcripts. All 3 coders then met to revise the initial codebook. Discrepancies were discussed until a consensus was reached among all coders. Importantly, the analysis software, Dedoose, facilitates the attribution of code applications to the research staff who coded the data, enabling the ability to identify and resolve coding discrepancies. All coders then divided the remaining data and applied the revised codebook so that the entire corpus was coded by at least two coders. Coders met 3 more times to discuss and resolve discrepancies. The saturation of codes was determined to be reached once all discrepancies in codes were resolved and no additional codes were identified in the overlapping transcripts. Coders then met to organize codes into axial themes iteratively.

At the time of this study, JM was a clinical psychologist and faculty member with a background in both clinical and research work in suicide prevention. He held funded grants focused on increasing access to care through the design, development, and evaluation of digital mental health interventions. SAP was a research assistant who supported the administration of the ARC methods described in this paper. Both JM and SAP interacted with participants whose data comprise the study corpus through study procedures such as email and, in some cases, interviews. To maximize objectivity during coding and analysis, all participants were assigned a coded ID number. TL was a research assistant external to the project and had no prior contact with any study participants.

Before beginning any screening or study procedures, all interested individuals and participants in this study provided affirmative informed consent, both for initial eligibility screening and subsequently for participation in the study. Participants were compensated up to US $80 for responding to the study prompts and for commenting on other participants’ posts.

Throughout the study, participants’ data were, and continue to be, considered confidential and protected. Only research personnel and individuals securing appropriate credentials and data use agreements were able to access study data and complete transcripts. Study transcripts were deidentified and maintained on secure encrypted servers at Northwestern University.

All study procedures were administered at Northwestern University and all study procedures were approved by the Northwestern University Institutional Review Board (project number STU00217191).

Across participants in our ARC groups, young adults expressed their need for automated digital SPIs. Young adults discussed (1) the treatment mechanisms an automated digital SPI must address, (2) core elements of the technology that affect the perceived usefulness and acceptability of an automated digital SPI, and (3) key skill-building features they would like to see in such an intervention.

Beyond intervention and technological requirements, participants in our study shared numerous preferences and feature suggestions for an automated digital SPI that helped support skill building, not simply skill documentation—extending the role and function of a safety plan. For example, participants shared that they wanted self-tracking features so that they could identify patterns in their suicide-related thoughts and behaviors and build insights over time. G1-P4 shared one way this feature could be integrated into the larger automated digital SPI:

Others wanted suggestions for coping skills. This augments traditional safety planning by making automated digital SPIs a coping idea generation and refinement system, hosting a repository for useful coping skills that might spark inspiration for a user, especially if that user is having difficulty generating coping skills on their own. One participant shared how this might work:

Relatedly, some participants wanted reminders to use coping skills. While safety plans have traditionally been a static document, the translation to an automated digital SPI enables the system to prompt the user to assess the need to use coping skills through periodic reminders. As G1-P2 shared, “I personally feel like getting reminders to relax and check in with myself would be helpful, since suicidal thoughts can be brought on by stress.”

Taken together, participants shared that through reminders, skill suggestions, and self-tracking features, an automated digital SPI presents an opportunity to build skillfulness and insight for a potential user, rather than simply documenting effective coping skills.

This study aimed to understand the needs, preferences, and design requirements and features necessary for an automated SMS text messaging–based suicide prevention intervention. Our results suggest that young adults who experience suicide-related thoughts and behaviors are open to embedding suicide prevention interventions within SMS text messages, but these interventions must be designed to meet their unique needs and preferences.

One of the important needs arising from our ARC groups was the importance of social connection and accountability in automated digital SPI design. Young adults who experienced suicide-related thoughts and behaviors were interested in interventions that strengthen existing protective factors and social connections. This finding aligns with the Interpersonal Psychological Theory of Suicide [60,61] and the Three-Step Theory of Suicide [62,63], which each identify social disconnection as a key driver of suicidal desire and ideation. An automated digital SPI, such as the one proposed in this study, may offer psychoeducation to users on methods of leveraging existing social connections to build social accountability for symptom self-management. It may prompt users to consider available coping activities that foster a sense of connection (eg, letter writing, watching a movie with others). Additionally, an automated digital SPI may be able to support users in crafting messages to trusted others to solicit support around safety plan use and offer resources that users can send to trusted others to educate them on how to respond to someone experiencing suicidal thoughts (eg, tips to mitigate unhelpful reactions and methods of engaging in nonjudgmental listening). Importantly, our results also show that connection to trusted individuals within participants’ personal networks was particularly valued relative to simulated connection or companionship achieved exclusively through texting with an automated chatbot. Designers of automated SPIs should consider designs that encourage and support real-world interpersonal connections with people in users’ personal networks; otherwise, they could unintentionally exacerbate a sense of isolation.

Our finding that young adults believed that having social accountability may support their use of a digital safety plan, as well as their application of coping skills, also sits within the context of digital health literature that suggests, for some, framework-based human support and accountability [64,65] can facilitate longitudinal engagement in digital interventions [66]. Although human support aimed at engaging users in a digital health intervention is conceptually distinct from the social accountability young adults saw as valuable in facilitating the use of a safety plan—not necessarily a digital health intervention—there is conceptual overlap that signals the importance of others for facilitating the use of a therapeutic tool that may, at times, be difficult to engage with.

Participants in our study wanted to be highly informed about how an automated digital SPI would handle personal data. This finding has implications for the design of digital suicide prevention interventions broadly. Intervention systems must proactively communicate meta-information about how the intervention operates, as this has the potential to increase confidence and comfort using the intervention. Prior work demonstrates that individuals are more likely to make use of crisis helpline resources when they are provided alongside barrier-reduction interventions that target informational gaps such as misunderstandings of crisis helpline practices with regard to active rescues (ie, emergency responders being called to make contact with an individual) before receiving a referral to crisis helplines [67]. Moreover, U.S. Suicide and Crisis helpline (ie, 988) data suggest that several of the top barriers to contacting the Suicide and Crisis helpline are due to fears about law enforcement being called and about privacy and confidentiality [68]. These top concerns are addressable through robust transparency practices that frontload information about the decision rules and processes used within the intervention and by its administrators. For instance, having clear and simple language before a user engages in the intervention that addresses (1) how and when confidentiality is broken and emergency services are called; (2) how data inputs are used such as contact information for individuals a user can reach out to for help during a crisis; (3) if, when, and how trusted others are contacted, and (4) proactive psychoeducational information about how privacy and confidentiality are maintained within the program (ie, data security measures). Finally, it is important to offer resources that empower users to ask questions about privacy, confidentiality, and clinical decision-making in other contexts such as formal treatment settings and crisis helpline calls. Thus, while our findings suggest the value of self-management to reduce the need for crisis services, they also suggest important ways that crisis services can improve.

Themes raised across our ARC groups highlight the importance of ensuring that an automated SPI is usable and convenient. In the context of SMS text messaging–based safety planning, this means the system design must incorporate bidirectional or interactive messaging. To date, SMS text messaging interventions for suicide prevention [40-44] have primarily used unidirectional messaging (originating from the intervention system and being sent to the user), leaving individuals with limited opportunity to engage with intervention material in ways that reflect their lived experiences. To ensure an SMS text messaging–based automated SPI is convenient and usable, the system must use the conversational medium to be interactive. Eliciting safety plan entries by prompting users to consider new or different aspects of coping and prompting users to be highly specific are all unique affordances offered by texting. To make an SMS text messaging system usable and convenient, a system must also leverage a text-based navigation menu. This enables users to access settings, jump to different parts of their safety plan, and modify elements of a safety plan, if needed. To this end, participants highlighted that while texting may be useful for safety plan creation, having multiple methods of editing a safety plan may be necessary if the navigation menu becomes cumbersome. Thus, offering multiple modes of interacting with one’s safety plan (ie, via SMS text messages and via a web-based editing tool) may be required to make an automated digital SPI maximally usable.

Our results reveal an overlap between young adults’ stated needs and preferences and innovations in the interactive text–based market. Our results highlighted that young adults had several needs and preferences that may lend themselves to innovation using emerging technologies. Specifically, young adults expressed a desire for a digital SPI service to support, among other needs, (1) coping skill acquisition and use, (2) quick and simple menu navigation and easy access to desired information, (3) interpersonal connection to trusted others, and (4) clear privacy policies. The recent commercial availability of highly customizable large language models (LLMs), a class of machine learning model trained on massive linguistic data sets and designed to process and generate text-based data, has increasingly been integrated into text-based mental health interventions in the form of chatbots, content generators, and as classifier algorithms that code or categorize data so that an intervention becomes more responsive to user inputs [69]. LLMs integrated into digital chatbots can tailor responses based on context, previous interactions, and user inputs. This emerging technology is beginning to be integrated into digital suicide prevention tools, especially as it relates to the detection of suicide-related user-generated content [70], and comes with significant risks in addition to substantial promise [71]. While our study did not explicitly seek to understand the role of LLMs in an automated digital SPI, participants in our study offered important insights that are relevant to intervention development in the context of a market where commercial LLMs are available and actively being incorporated into different digital mental health technologies.

First, as it relates to supporting coping skill acquisition and use, LLM-based messages may be launched to support users in ideating coping skills when they have difficulty generating their own. Moreover, the ability of LLMs to personalize content based on individuals’ contexts and inputs creates opportunities to increase the relevance of a user’s safety plan at the moment. With a user’s existing safety plan as context, LLMs may be able to support idea generation about how to overcome barriers to coping (eg, if an individual’s safety plan contains going for a walk outside as a coping skill, but weather is prohibitive, LLM-based messages may be able to support a user identifying relevant solutions, such as going for a walk indoors, taking protective gear, using other coping skills on one’s safety plan). Second, LLMs may be able to improve the experience of navigating rigid text-based menus embedded in an SMS text messaging intervention by enabling the system to accept user responses that do not exactly match preprogrammed keywords that direct a user down different branching logic trees and instead allowing users to respond using natural language. Third, young adults were clear that they wanted a digital SPI to support interpersonal connections to other humans. While they were open to an automated digital SPI, connection to real people was crucial. LLM-based chatbots have often been used as conversational agents [70], some of which have personified the conversational agent. This may exacerbate a suicide-specific vulnerability: isolation and loneliness. Past work highlights that young adults with depression were concerned that using a robot for companionship could worsen their symptoms by making the absence of real human companions more apparent [72]. More research is needed to understand whether LLM-based suicide prevention tools should present themselves as personified agents or as self-service tools. Finally, given participants’ emphasis on the importance of privacy and transparency, if an LLM was to be integrated into an automated digital SPI to enhance messages, it would need to (1) proactively make transparent ways in which user inputs would be used to train models and (2) disclose the potential risk of confidentiality breaches through the LLM itself, as has been documented in other contexts [73].

By engaging young adults with lived experience of suicide-related thoughts and behaviors to understand their needs, preferences, constraints, and the context of existing technological solutions, we can identify design requirements that are able to be applied across digital suicide prevention tools.

This study is limited in several important ways. First, while we engaged in rigorous user-centered design practices to understand the needs and preferences of young people who experience suicide-related thoughts and behaviors, we were only able to engage a relatively small number of individuals, and our results may not be generalizable to broader young adult populations. In our sample of 30 participants, 11 (37%) were in formal mental health treatment, and 17 (57%) planned to seek formal treatment in the coming months. These treatment-engaged numbers are greater than the approximately 12% of US adults who have experienced suicidal thoughts and sought treatment in the past year [2]. Moreover, we recruited exclusively from MHA’s website which may inherently limit the perspectives of individuals included in the study to those who are open to technology-assisted interventions and services, and potentially more actively engaged in self-management of mental health symptoms. While we aimed to cover a broad range of topics in our ARC groups, our discussions were guided by researcher-posted prompts and centered on eliciting needs and preferences as they related to safety planning or texting. To this end, in this study we did not systematically examine needs and preferences for risk management in self-guided digital safety planning tools; however, future work must aim to better understand the needs and preferences across interested parties (eg, consumers, research participants, clinicians, organizations) as it relates to risk management procedures. Finally, individuals at the highest levels of suicide risk (ie, individuals who may have had elevated levels of suicidal desire and intent) were not eligible to participate in this study and thus their voices were systematically excluded as the study format lacked appropriate support for these individuals. Future work should focus on applying these design considerations and assessing their value within the context of deployable digital suicide prevention interventions.

Engaging young adults in the design process of a digital suicide prevention tool revealed critical considerations that must be addressed if the tool is to effectively expand access to evidence-based care to reach young people at risk for suicide-related thoughts and behaviors. Specifically, automated digital SPIs must support building skillfulness to cope effectively with suicidal crises, deepening interpersonal connections, system transparency, and data privacy.