Mental health professionals’ perspectives on dynamic learning of individual-level trajectories in youth mental health care: A qualitative study

Mental health professionals’ perspectives on dynamic learning of individual-level trajectories in youth mental health care: A qualitative study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Mental health professionals’ perspectives on dynamic learning of individual-level trajectories in youth mental health care: A qualitative study Adam Poulsen, Ian B Hickie, Min K Chong, Haley M LaMonica, Ashlee Turner, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7062632/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Background: Assessment of suicidal ideation in youth mental health remains largely static, subjective, and inaccurate. Digital tools present a possible solution. This research examines a novel digital mental health tool, the Dynamic Learning Tool, which applies a machine learning model for individual-level continuous-time predictive trajectories of clients’ suicidal ideation to inform clinical care and was developed for an existing platform used in Australian and Canadian clinical practice. This work aims to explore professionals’ attitudes towards machine learning of predictive trajectories and artificial intelligence and thereafter determine the Dynamic Learning Tool’s design and content requirements. Methods: Following a co-design methodology, semi-structured interviews and usability testing were conducted with 21 mental health professionals in Australia, Canada, and the United States. Data analysis employed inductive reflexive thematic analysis. Results: Findings indicate that professionals are open to using predictive trajectories strictly to enhance their decision-making but not replace clinical judgement. Furthermore, they emphasised the need for greater machine learning model transparency, incorporation of real-world contextual data, clear guidelines for use, and overt adherence to privacy and accountability standards. Conclusions: The findings presented here have broader implications for the ethics, design, development, implementation, governance, and evaluation of machine learning of predictive trajectories and artificial intelligence in youth mental health. Artificial intelligence machine learning suicide prediction practitioner attitudes interviews qualitative research Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Background Globally, youth mental health is deteriorating [1], and youth suicide rates remain alarming [2-4]. Mental ill health is a multifaceted problem impacted by environmental, social, economic, political, and technological change stressors and persistent stigma and discrimination about mental health [1, 5, 6]. Young people are particularly vulnerable as many mental disorders emerge during developmental stages [7], further compounded by contemporary challenges such as climate change, economic instability, intergenerational inequity, and unregulated social media exposure [1, 8, 9]. The high demand for mental health care necessitates more effective strategies and systems for indicated prevention and early intervention [10]. Current systems struggle to identify and respond to risk in a timely manner, relying heavily on static, one-off sampling of clinical state and risk markers (e.g., based on single baseline assessments) to predict clients’ mental health trajectories, yielding only modest accuracy and replicability [11]. Responding to these multidimensional problems requires a highly personalised and real-time measurement-based approach that incorporates dynamic change into predictive modelling to better understand individual illness trajectories in young people [11, 12]. The continued digitisation of mental health, especially with artificial intelligence (AI) applications such as machine learning, represents a crucial advancement [13]. Digital mental health tools (DMHTs) have been shown to improve mental health outcomes [14] and are positioned to advance the accessibility, efficiency, and scalability of interventions to complement traditional care and services [15-17]. Youth suicide remains a significant challenge requiring urgent attention. Current suicidal ideation assessment, conducted primarily through needs-based clinical interviews, has limited temporal validity and frequently raises false positives and negatives [18]. Moreover, the dynamic nature of suicidal ideation makes accurate prediction of trajectories difficult [19]. Clinicians therefore need more objective and precise tools to supplement their clinical judgement and predict individual suicidal ideation trajectories [20, 21]. Big data and machine learning models, applied through DMHTs, offer promising opportunities for dynamic trajectory learning and prediction of suicidal ideation in mental health services. These technologies can help to overcome the highly subjective nature and inaccuracy of human-driven prediction [22, 23] and advance towards greater personalisation through, for instance, gathering user-generated and passive sensor data relevant to suicidal ideation trajectories (e.g., sleep behaviour data collected via smartwatches) [18, 24]. To date, the accuracy of algorithmically modelling suicidal ideation trajectories demonstrates promising results [18, 22, 23, 25, 26]. This data-driven, individual-level, ongoing approach represents an imperative shift from conventional risk stratification-based prediction towards highly personalised and measurement-based care [12, 27-29]. This digital, personalised approach would be a transformative change in traditional mental health care. Therefore, understanding mental health professionals’ attitudes and requirements concerning DMHTs is a crucial step before implementation [30], particularly for identifying possible barriers [31] and determining factors affecting usefulness, acceptance, and effectiveness [32]. Research on DMHTs reports a diversity of perspectives among mental health professionals [33-37]. In the United Kingdom, a 2019 online survey of child and adolescent mental health services clinical staff found that mental health professionals considered DMHTs helpful in their clinical work, advancing accessibility, convenience, and appeal to young clients, though they expressed little knowledge about the safety, security, privacy, and reliability of DMHTs used [37]. Similarly, a 2023 Portuguese study revealed that while 87.2% of professionals supported prescribing DMHTs to clients, only 43.6% planned to implement them, citing barriers such as a lack of information about DMHTs, required training effort, and the need to adjust clinical processes and records [33]. Professionals also reported key determinants of DMHT adoption that were not adequately met, including information about alignment with health objectives, evidence of DMHT validity, and expert-based recommendations for specific DMHTs [33]. A 2022 study in China by Zhang, Lewis (34) reported that clinicians held generally positive attitudes towards DMHTs, but they expressed concerns about online safety, social and cultural barriers, and limited digital competence among professionals and clients. Despite the broader evidence on DMHTs, key underexplored attitudes and design requirements among mental health professionals about machine learning of predictive trajectories include clinical utility, safety, ethics, effectiveness, permissibility, user interface (UI) design, implementation, and governance [22-24]. Going forward, the development and evaluation of implementable, scalable, and culturally sensitive machine learning models and tools for predicting suicidal ideation trajectories requires a deeper understanding of these factors, particularly regarding acceptability, feasibility, cost-effectiveness, real-world implementation, ethical considerations, and potential impact to ultimately realise the full potential of these technologies [38, 39]. Thus, this work aims to explore mental health professionals’ attitudes towards AI (e.g., machine learning models) in youth mental health and determine the design and content requirements for a novel DMHT, the Dynamic Learning Tool (DLT). Principally, the DLT provides predictive trajectories of individual clients’ suicidal ideation utilising machine learning to inform clinical decision-making by youth mental health professionals and is integrated into an existing platform used for clinical practice in both Australia and Canada. As outlined in an earlier technical paper [40], the DLT operates by: (1) using longitudinal suicidal ideation data collected in a clinical setting, captured with the suicidal ideation attribute scale (SIDAS) [41], (2) generating predicted trajectories of an individual’s suicidal ideation level and variability to guide clinical decisions and determine treatment progress (e.g., changes in the frequency that individuals may need to be observed), (3) incorporating a continuous-time modelling framework to handle irregularly time-spaced observations common in clinical data and update predictions as new observations are collected [40]. For the purposes of this study, the tool was incorporated into an online platform, Innowell (see Figure 1), which supports self-reporting, management, and monitoring of mental ill health and maintenance of well-being [42, 43]. As a precursor to the DLT, Innowell also incorporates static categorical suicide risk assessment indicators (low, moderate, or high risk) (see Figures 1a and 1c), based on self-reported Columbia-Suicide Severity Rating Scale (C-SSRS) [44] data. Figure 2 visualises the DLT as integrated into Innowell. Per best practices in developing digital health solutions [45], this study employs a co-design methodology using semi-structured interviews and usability testing to understand mental health professionals’ attitudes and co-create design and content requirements for the DLT. Methods Study setting This study was conducted primarily online by a researcher in Sydney, Australia, from June 20 th to August 23 rd , 2024, via the Zoom videoconferencing software, enabling participants across Australia, Canada, and the United States to participate. Ethics approval This research received ethical approval from the University of Sydney’s Human Research and Ethics Committee (Project number: 2021/HE000680). Participants provided informed consent prior to completing the study. Sampling and data collection Purposive sampling was utilised to recruit participants for this study. Eligibility criteria included mental health clinicians who assess and treat young people or adults (including mental health service managers and administrators in this context) and required English language proficiency. The researchers identified potential participants (e.g., known professionals and local professionals advertising services online) and contacted them via email. Additional recruitment was facilitated through mental health organisations, professional associations, and university departments that agreed to advertise the study details. Potential participants joined the study by completing an online form via Qualtrics, which provided them with a participant information sheet and allowed them to consent electronically. Qualtrics automatically notified the researchers when a participant joined the study, sending the participant contact details to schedule the interview. Participants were free to select a convenient interview time and mode of participation (i.e., in-person and remote videoconferencing participation were offered). Two participants elected to participate in person at the Brain and Mind Centre. The average interview length was approximately 47 minutes (shortest=28, longest=64). Interview audio and visuals were recorded and later transcribed verbatim and de-identified. Researcher notes were taken throughout the study to collect rich contextual information and benefit transcription [46]. Data collection stopped when the sample size was within a pre-determined range (15-25 participants for single case studies [47]) following the information power principle [48], which takes into consideration the study aim, sample specificity, theoretical background, quality of dialogue, and strategy for analysis. Following this principle, a study needs the least amount of participants when the study primarily addresses a narrow aim (i.e., attitudes about AI in mental health), the participants represent similar experiences and knowledge (i.e., mental health clinicians), the study is grounded in existing theory (i.e., personalised and measurement-based care and co-design), an experienced interviewer conducts it, and it addresses a single case analysis. Following a co-design methodology, the interviews involved questioning as well as a think-aloud activity as a usability test [49, 50] of the DLT, adhering to the following procedure. Participants were informed about the study aim, interview procedure, DLT, and Innowell. Consent to record the audio and visuals of the interview was reaffirmed. The researcher shared their computer screen to demonstrate how to access the DLT online and answer any related participant questions. Participants freely tested the DLT, completing a think-aloud activity that involved them verbalising what they see, do, think, and feel while testing the tool [49, 50], with researcher support (e.g., answering participant questions about the tool) provided throughout. A semi-structured interview guide of open-ended questions and possible follow-up questions was developed for this study and followed. The questions were adapted to ensure a natural conversational flow and account for numerous other factors, such as a participant’s digital literacy. Unguided usability testing without extensive instruction and training simulated real-world first-time use of the DLT, eliciting data relating to tool implementation and governance. Box 1 presents the interview agenda topics that guided the interview (see Additional file 1 for the semi-structured interview guide). Box 1. Interview agenda topics. Acceptability Initial thoughts Appropriateness Intended users Clinical utility Fit for the intended purpose Potential uses Function, utility, and outputs Impact on clinical decisions, practice, services, and clients Design Usability Engagement Aesthetics Clinical workflow and implementation Integration into practice Integration into mental health services Ethics and safety Ethical and safety considerations Trust Transparency Privacy Professional codes of conduct DMHTs Existing tools used Personalised and measurement-based care Next steps Features wanted Feedback on proposed future features In total, 21 mental health professionals were interviewed between June–August 2024 in Australia (n=18), the United States (n=2), and Canada (n=1). Participants comprised clinical psychologists (n=12), clinical neuropsychologists (n=3), and a clinical nurse consultant, mental health nurse, occupational therapist, social worker, service manager, and registered psychologist. A total of 19 participants (mean age=38 years; range=25-63; 58% female; 42% male) provided further demographic data. Summarised demographics are presented in Table 1. Notably, most participants worked in metropolitan areas (n=16), and a few reported working in a regional area (n=2) or rural area (n=1). Furthermore, most participants working in multiple sectors also worked in private practice (5/6). Table 1. Participants’ demographic data. Variable Number (n) Percentage (%) (n=21) Country Australia 18 85.7 United States 2 9.5 Canada 1 4.8 Occupation Psychologist 12 57.1 Clinical neuropsychologist 3 14.3 Social worker 1 4.8 Service manager 1 4.8 Clinical nurse consultant 1 4.8 Mental health nurse 1 4.8 Occupational therapist 1 4.8 Registered psychologist 1 4.8 (n=19) Age 25-29 5 26.3 30-34 4 21.1 35-39 3 15.8 40-44 3 15.8 45-49 2 10.5 50-54 1 5.3 55-59 0 0 60-64 1 5.3 Gender Female 11 57.9 Male 8 42.1 Highest educational attainment Doctorate 12 63.2 Master’s degree 4 21.1 Bachelor’s degree 3 15.8 Work sector Academia 3 15.8 Community organisation/NGO 3 15.8 Public 3 15.8 Hospital 2 10.5 Private 2 10.5 Multiple Private and academia 2 10.5 Private, community organisation/NGO, and academia 2 10.5 Private and community organisation/NGO 1 5.3 Public and academia 1 5.3 Years in profession 0-5 3 15.8 6-10 8 42.1 11-20 6 31.6 21+ 2 10.5 Geographical area Metropolitan 16 84.2 Regional 2 10.5 Rural 1 5.3 Analysis Inductive reflexive thematic analysis was applied in this research to develop and report themes based on patterns of shared meaning (or meaning-based interpretative stories [51]) in the interview data without a pre-determined set of themes [52, 53]. This study adhered to conventional thematic techniques, following six iterative analytic phases: data familiarisation, generating initial codes, generating candidate themes, reviewing candidate themes and establishing final themes, defining and naming final themes, and producing the report [54]. NVivo 14 was utilised to organise and code the interview data. Per best practices for reflexive thematic analysis [55], the remainder of this section reports theoretical assumptions that underpin the analysis and details how thematic analysis was implemented in this study. Given that reflexive analysis never neatly fits into either an inductive or deductive approach and it instead combines both to some extent as researchers must make some assumptions (e.g., what pieces of information are worth coding), one key theoretical assumption underpins the analysis. That is, clinical utility, acceptance, and design are key determinants of DMHT use, and thus these factors were considered foci during the analysis process. Researcher reflexivity was crucial throughout the analysis to confront and interrogate assumptions [52]. Here, thematic analysis was implemented as follows. Partway through data collection, two coauthors (AP and HML) independently read and engaged in familiarisation of three interview transcripts and generated initial codes. The initial codes were reviewed and discussed, resulting in numerous codes being merged, discarded, and revised and new codes being created. Following this review, noting an adaption of the reflexive thematic analysis method in this study [55], a mutable coding glossary was developed by AP and HML to create a shared language for the remaining analysis processes. When data collection stopped, three coauthors (AP, MKC, and AT) each independently reviewed and coded a portion of the remaining transcripts with reference to the flexible coding glossary. Then, AP, MKC, and AT met to iteratively generate, review, and label candidate themes and subthemes in consultation with the senior author (FI). After that, AP independently examined the candidate themes and subthemes, leading to the themes being renamed and several subthemes being merged. Finally, AP and FI established the set of final themes and subthemes and their labels. Results Three themes and seven subthemes were developed through reflexive thematic analysis of the interview data. Following best practices for reporting thematic analysis [54], Figure 3 presents the final thematic map. Figure 4 illustrates the combined initial and final thematic maps, indicating four additional candidate subthemes that were merged into the final set of themes. Table 2 reports the themes, subthemes, associated descriptions, and illustrative participant quotes. Table 2. Themes, subthemes, and illustrative participant quotes. Theme Subtheme Illustrative participant quote Value-add of predictive trajectories in mental health care Clinical utility in the ‘care as usual’ milieu ‘This is the problem with services in acute care. It's all about symptom reduction, not resolution. That's where I come in, and I, trust me, I had to headbutt the brick wall a lot with other mental health services that are supposed to jump in and help us, but they don't. So, if I've got this, I can then actually provide evidence to a service we can cut and paste this into a report per se, and say, “This is what we've seen. These were the events”, and it provides some historical data to the service to say, “This has been your [a client’s] input. This is where they've come from, this is where they're going to, and we don't know where this person's going to go”’ (P2 Mental health nurse) Accuracy: Clinical judgement versus machine learning ‘And seeing some of these predictions, the predictions aren't accurate enough. […] We're not very good at saying what level of confidence we have for this individual yet. I do think it's an area of great interest and importance for clinicians. But, it still feels like there's a lot to do in terms of the data itself, the accuracy and the validity of that data and the predictions because clinicians really need to be convinced if they're going to take an action that is against what they would have taken without that data. […] If you've got data and you've got ways to show people that they can increase their faith in these predictions, they're now likely to change their clinical decision. If not, they might consider it in amongst 100 other things they're considering about what to do with this person’ (P6 Clinical psychologist) Illuminating and magnifying the trajectory landscape More transparency is needed ‘I'd want to know a bit more about this algorithm. And, beyond the black box of how it is actually making this prediction, how seriously I need to take it and what the margin for error is. I can see the margin error, I guess, I'm inferring that the bigger the [trajectory] range in the display over time infers that confidence. […] In terms of AI clinical decision tools, there's a lot written around the transparency for a clinician about how it is calculating this and how it is arriving at this conclusion. So, taking that information out of the black box a bit; otherwise, you just won't have trust in the model. I don't know why that's telling me that there's a high risk of this person getting worse or a low risk of this person getting worse’ (P6 Clinical psychologist) Incorporating real-world considerations ‘I'd love to know if Innowell was able to take into account things like age, marital status […] social connection, or if there's any extra bits because I think that there might be people who would say, "No, I haven't had any [suicidal] thoughts", or they wouldn't want to talk about it. But, they might still be someone who […] statistically is at a greater risk, or is not. They're coming in, and they're really starting to neglect their personal hygiene, and you notice those things, or just those other factors that I would be taking into account with the risk assessment on top of these [SIDAS and C-SSRS] questions. […] I think demographic stuff, and that could be built into a tool like this. And I know there are subsets of the population who are at increased risk, and there are determinants that we know of or that we know make people at increased risk. So that would be cool if that was built in as well’ (P9 Registered psychologist) Moving towards implementation Machine learning amplifies ethical challenges ‘I remember being trained and specifically being told, ‘Cover your arse.’ That is your job. Don't expose yourself to risk. […] If this person goes into hospital tomorrow, have you done all the steps? Have you done all the right things? Did you miss something that you should have not missed? And so, for a tool like this [the DLT] for me, it is about providing the guidelines to ensure that I have done those steps. I have checked in. I have delivered a safety plan. I have cross-validated that right now, I understand that the guidelines tell me that this is probably the best recommended care. This is the evidence-based care. Have I done that, and have I box ticked that? Yes, I have. Did I miss anything? No, I have not’ (P7 Clinical psychologist) Clear guidance on use is necessary ‘Clinicians will need to be trained on how to read it for sure. I think once you've explained that to me, now I can kind of see what you're talking about. It's not evident very clearly at the beginning what the dots and the lines mean, and the grey parts [DLT trajectory output] mean? […] When does this come in handy? […] Is it to help support us in risk assessment? Or is it to help us support in what we could do to bump it up higher towards the more healthy kind of aspect?’ (P17 Occupational therapist) Engaging the right users 'I think there's a potential that people see their computer [DLT trajectory output] and say, “I'm feeling suicidal and I'm not going to feel less suicidal, and so why bother. It's never going to get better.” I think that's definitely a risk. That potentially could be interpreted very negatively by clients. […] That could be, that could be pretty counter therapeutic' (P16 Clinical psychologist) Value-add of predictive trajectories in mental health care Clinical utility in the ‘care as usual’ milieu . During the interviews, participants described various possible uses of the DLT in clinical practice. Primarily, the discussion centred on the DLT playing an added role in coordinating care and data-driven trajectory learning and prediction. Coordinating care with the DLT may involve using it as a collaborative tool between clinicians and clients, documenting clinical support roles, evidencing referrals, and supporting triage processes. On data-driven trajectory learning and prediction, most participants indicated that the DLT adequately indicates risk, timeliness of data, trends and change, and volatility, positioning it to support risk mitigation, decision-making, care planning, streamlining work, and a shift towards more targeted and personalised interventions. Some participants also suggested that the DLT could be used as a training or supervisory tool for junior clinicians or peer workers. Discussions often situated the DLT in regular mental health care settings, with participants reflecting on typical processes, failings, existing tools, and the fit of the DLT among these components. For the most part, participants noted that assessing suicidal ideation is very difficult, and they reported varying, standardised and non-standardised ways of assessing suicidal ideation, including measures (e.g., SIDAS), interviewing, and observations. Furthermore, each participant prioritised different interventions, depending on the population and mental health disorders they typically encounter in clinical practice. Despite a diversity of approaches, all participants emphasised the importance of collaborative care and sought to understand client goals and values in developing care decisions and plans, including those relating to suicide. When discussing where the DLT could fit within existing services, participants remarked on current issues within services to highlight the difficulty in introducing new digital technologies. These include high demand and low resources, long waitlists, a perceived focus on symptom reduction rather than resolution, slow adoption of digital technology, non-interoperable software across services, limited time available to integrate and use new tools, and bureaucratic processes and administrative tasks burdening clinicians. Most participants used digital tools in their practice. However, this was mostly limited to electronic health records for mandatory data collection procedures, and a few participants reported using AI tools to support report writing and notetaking, but none were for making predictions. Framing the DLT within real care settings created a focused space in which participants suggested utilising the DLT as a tool to enhance and accommodate existing processes, diverse approaches, and technology use, namely centralising data collected via measures and supplementing non-standardised assessments with quantitative data. Furthermore, reflecting on current issues, participants highlighted the need to ensure that the DLT is non-disruptive and meaningful. Accuracy: Clinical judgement versus machine learning . Accuracy was a salient topic. During user testing, many participants flagged concerns about the accuracy of the DLT trajectory output, which is chiefly displayed in a graph that visualises historical SIDAS scores and a trajectory range for the next SIDAS score indicated by a grey-coloured area on the graph (see Figure 5). Specifically, in instances when the trajectory range of the client’s predicted future score was wide (indicating uncertainty, informed by either volatility in clients’ historical scores [see Figure 5a] or a sudden recent change after a long period of stability [see Figure 5b]), the accuracy and clinical utility was in question. Furthermore, in instances when the trajectory range was narrow (indicating high confidence, informed by a long period of consistent stability [see Figure 5c]), participants expressed that the DLT trajectory output is not evidently any more accurate than a clinician who could determine this by looking at the same data trend, thus limiting the added value of the DLT. In sum, it was not readily apparent to participants how to accurately interpret the trajectory ranges as indicating, in part, uncertainty (wide range) and confidence (narrow range). Views on clinicians’ perceived ability to change their approach to suicidal ideation assessment, poor adoption of digital interventions, and the superiority of clinical judgement compared to algorithmic prediction in mental health play a role in the perceived accuracy of the DLT. Specifically, as reported in the interview data, inspiring a change in approach to suicidal ideation assessment among clinicians is difficult due to the challenging nature of predicting suicide, reliance on favoured approaches for assessment, and the perceived burden and incongruence between digital interventions and trained human-driven approaches. Participants expressed that suicide prediction requires, in some part, the human touch and that they would more likely rely on a combination of their clinical judgement and the DLT compared to the DLT alone in decision-making, framing the DLT as an instrument to either validate or invalidate a clinician’s assessment. Yet, only a minority of participants expressed a willingness to use the DLT as a tool to enhance existing practice, and none were willing to completely delegate suicidal ideation assessment to AI, due to lingering uncertainties about accuracy and the perceived advantages of clinical judgement to draw on additional contextual data in collaboration with a client in situ. Illuminating and magnifying the trajectory landscape More transparency is needed . The DLT allows users to browse each client’s responses to each SIDAS question to provide transparency. During user testing, most participants appreciated this feature as it allowed them to investigate individual assessment items closely to better understand clients’ overall scores. Nevertheless, most participants sought greater transparency about the DLT model to improve their understanding of how it works, improve their interpretation of trends across clients’ scores and trajectories, and, ultimately, advance trust. Specifically, they wanted to know what data is incorporated into the DLT model, including the data utilised during model training and operation, what constituents of SIDAS are incorporated in the predictive trajectories, and whether static factors (e.g., demographics) and population data are integrated. Furthermore, participants wanted more details about DLT trajectory outputs' calculations and confidence intervals. To that end, participants made various redesign recommendations to ensure that key information about the DLT model was plainly communicated on the UI. These included a confidence interval indicator for the DLT trajectory outputs; overtly flagging key driving predictors for each client’s trajectory; highlighting and elaborating on different psychiatric constructs underlying SIDAS that inform the DLT trajectory output; creating links between DLT trajectory outputs and the research evidence base and proof of the DLT model efficacy (e.g., providing external links to academic papers explaining the efficacy and importance of measures used, mental health dimensions considered, and the DLT model itself); and identifying the data used to generate outputs. Incorporating real-world considerations . The various existing approaches to assessing suicidal ideation demand individual-level, mutable, and diverse contextual data. This includes, for instance, client demographics, social dimensions (e.g., family or romantic relationships), major life stressors (e.g., financial or legal challenges), and mental health history (e.g., past suicide plans or attempts). Some of this data also relies on being face-to-face with clients, such as observations of body hygiene and posture. Thus, participant feedback on the data used by the DLT to model suicidal ideation trajectories (i.e., SIDAS questionnaire responses) was modest and mixed. In short, whereas some noted that SIDAS is the gold standard for assessing suicidal ideation, others reported that the measure lacked nuance and missed crucial context. The same sentiment was expressed about C-SSRS (although it is not integrated into the DLT trajectory outputs). During user testing, participants recommended incorporating real-world considerations that they currently rely on or would like to have access to for assessing suicidal ideation into the DLT in two key ways. First, feature key client demographics next to the DLT trajectory output to help clinicians recall key client details in situ. One participant highlighted that this feature might prove essential when assessing risk, especially when dealing with high-risk populations, noting that what appears to be a high-risk score for the general population might represent an average score among specific high-risk populations, such as Indigenous Australian brotherboys or sistergirls. Second, the variables considered by the DLT model should be expanded. Vital data to incorporate into suicidal ideation modelling included client demographics (e.g., age, sex, relationship status) and measures of impulsivity, substance use (alcohol, tobacco, cannabis), mental health (depression, anxiety, post-traumatic stress, psychosis-like and mania-like experiences), social factors (connectedness, occupational functioning, quality of life), and behavioural risks (self-harm, homicidal ideation, oppositionality, rule-breaking, sleep disturbances, distress, and violence). One participant recommended determining average SIDAS scores based on population data (e.g., sex, age, educational attainment level, and risk factors) and showing a comparison between a client’s score and their population group average. Another participant suggested integrating key events in a client’s life or mental health treatment (e.g., the start dates of the school year, new therapy intervention or medication, and a suicide attempt) to both inform the DLT trajectory output and provide greater context and perhaps help to interpret a client’s historical data and trajectory. To that end, participants discussed sourcing additional data to advance the contextualisation of the DLT model from client notes, health records, police records, coronial reports, and client session transcripts, leveraging data mining techniques and large language models to parse data. Beyond suicidal ideation, participants also recommended modelling and generating predictive trajectories for numerous mental health dimensions. These included ‘more classic, major aspects of symptomatology’ (P10 Clinical neuropsychologist) such as psychological distress, alcohol, tobacco, and cannabis use, sleep, anxiety, psychosis-like experiences, and depression. However, emphasising the importance of context, participants noted that the utility of these mental health dimensions trajectories (including suicidal ideation) depends on individual client presentation and a service’s typical population. Moving towards implementation Machine learning amplifies ethical challenges . Two key ethical considerations about machine learning in mental health and the DLT were discussed: privacy and accountability. Participants were very concerned about privacy relating to DMHTs. Many noted the importance of DMHT platforms ensuring that sensitive personal information was protected, remained confidential, and managed according to national digital health standards governing privacy, especially given the sensitive nature of data regarding suicide. They raised questions about data ownership, security, governance, and storage in digital mental health and conveyed that the integration of AI obfuscates these issues further. When discussing the use of AI scribes in mental health, i.e., speech recognition technology that automatically converts voice-to-text for clinical documentation, and the possibility of integrating the same functionality with the DLT, participants flagged that the data should be deidentified before being utilised and that appropriate procedures to obtain client consent to use their data is necessary. Some participants said explicit adherence to Australian privacy and cybersecurity laws and standards is essential. Accountability was a significant concern. Many reflected on existing accountability for acting upon assessments that indicate a high risk of suicide and wondered if the DLT added additional complexities to practice. For instance, one participant noted that if the DLT trajectory output indicates a negative future trend and the clinician does not expect the same trend, the accountability for action based on conflicting assessments is unclear. Furthermore, the accountability for any subsequent negative consequences to the client is confused. Multiple participants made explicit requests for guidelines on using and actioning the DLT. Clear guidance on use is necessary . During user testing, most participants found it challenging to discern the DLT trajectory output (see Figure 5) intuitively. This presented an initial barrier to clinical use, and most participants needed comprehensive instructions to interpret the output. Participants reported that the output was hard to understand for several reasons. These were (1) data plotted, and the trajectory itself was inadequately labelled; (2) the y-axis directions were confusing and did not meet expectations (i.e., whereas the top of the y-axis represents ‘healthy’ scores, participants expected the opposite); and (3) a tooltip (i.e., a text box that appears when a user hovers over a specific UI element to describe the element) integrated into the graph was considered frustrating as it sometimes covered elements of the graph and did not provide helpful, applicable, and tailored information. Moreso than resolving the aforementioned UI problems that created confusion, the participants indicated that the DLT should be implemented with adequate guidelines and associated onboarding training. This includes formalised comprehensive guidelines, training procedures, and helpful tips featured on the UI to provide simple, ad hoc guidance. Specifically, guidelines and training are needed to advise how to interpret the outputs and subsequently develop an appropriate care plan based on the outputs. Participants indicated that guidance is crucial to ensure correct use among junior clinicians who may experience more difficulties interpreting and generating relevant treatment options than a more experienced clinician. Engaging the right users . Identifying target user groups for the DLT was a salient, divisive point of discussion. Participants agreed that mental health professionals were the appropriate primary target end users. Those working in private psychology practices or smaller service centres were more interested in using the DLT than those working in public and large service centres (e.g., headspace). Participants representing the latter group reported disinterest due to the existing burden of large clinical caseloads and administrative obligations in public and large psychology service centres. There were strong attitudes about whether young people (i.e., clients in youth mental health services) should have access to the DLT. Some participants conveyed that young people should have access to information about their care, especially to track and observe trends and create opportunities to discuss data with clinicians. Thus, they should have access to the DLT. Other participants expressed the opposite attitude, noting some risks of young clients seeing the DLT trajectory outputs, including misinterpreting the outputs, creating a sense of hopelessness, and generating negative expectation-induced behaviours (i.e., if a client observes that the DLT indicates a likely negative trend, then they may be disempowered or withdraw from treatment participation). As pointed out by the participants, depending on the client and their current mental health experiences and trajectory, accessing the DLT risks empowerment or disempowerment. During the interviews, to further explore the possibility of young clients accessing the DLT, a feasible design feature was presented to gather participant feedback: clinician direct control over which aspects of the DLT each client can access. Most participants rejected this potential feature, suggesting it negatively impacts young people’s autonomy and may create a sense of infantilisation. Ultimately, it was determined that for the sake of safety and to avoid these negative impacts, it is best to avoid clients accessing the DLT altogether. As the next best alternative that aligns with how clinicians typically share data with clients, participants favoured an approach that enabled clinicians and clients to collaboratively access the DLT during a client’s session at the clinician’s discretion. Participants also noted that it is difficult to encourage young people to engage with DMHTs for self-reporting purposes, and thus, enabling clients to access the DLT may be ineffective and require substantial integration into service delivery. Discussion This qualitative study employed semi-structured, in-depth interviews, usability testing, and reflexive thematic analysis to explore youth mental health professionals’ attitudes towards AI and determine the design and content requirements for the DLT. Most participants were open to utilising AI and the dynamic learning of illness trajectories across mental health dimensions in clinical practice as a tool to enhance and support their decision-making. Furthermore, most found that the DLT indicates risk, volatility, trends and change, and timeliness of data. Yet, they expressed scepticism about predictive trajectories of clients’ suicidal ideation, compared to other mental health dimensions, citing fears of perceived low accuracy and AI tools replacing clinical judgement. Participants emphasised that suicidal ideation assessment largely remains a human-centric task, leveraging existing quantitative measures, interviewing, observations, understanding client goals and values, and collaborative care. They viewed AI and the DLT as tools to enhance, but not replace, care coordination and data-driven trajectory learning and prediction in Australian youth mental health, emphasising that implementation should support existing processes and technology use without adding to existing issues, including administrative burden and slow technology adoption. Following the co-design methodology, participants served a vital role in informing the redesign of the DLT model and its implementation. Participants called for greater model transparency to inspire trust among users. Users should be able to examine what data is used to develop DLT trajectory outputs and train the model, the confidence intervals of the outputs, key driving predictors of each trajectory, and supporting evidence validating the model. Moreover, participants wanted the DLT model to be expanded to incorporate individual-level, mutable, and diverse contextual data that they find useful in assessing suicidal ideation, including mental health history, client demographics, and major life stressors. Beyond suicidal ideation, participants supported modelling trajectories for other mental health dimensions, such as psychological distress, anxiety, and depression. Although suicidal ideation was the focus of this study, the DLT could be adapted to learn trajectories of other mental health dimensions with appropriate evidence-based measurements. Table 3 summarises the key, in-scope recommendations for the DLT co-designed during interviews, including new and redesigned features. Table 3. Additional requested features for the DLT. Source of frustration (Re)design recommendations Related theme (subtheme) Lack of content personalisation Content associated with the static risk indicator should include more personal insights into a client’s risk factors, with reference to significant results from questionnaires Any instance referring to the client should use their proper name Value-add of predictive trajectories in mental health care (Clinical utility in the ‘care as usual’ milieu) No link between trajectory and recommended treatment options/care plan Generate recommended potential treatment options based on predictions for clinician users to consider and action if deemed appropriate. For example, if the client’s trajectory range is beyond an uncertainty threshold, inform the clinician user and recommend a higher frequency of client follow-up appointments and review of their treatment plan No flags of significant risk factors Instances that could be flagged and conveyed to the mental health professional include: Significant areas of risk related to client questionnaire responses The probability of reaching a high-risk category, as indicated by an integrated high-ideation probability (IHIP) score greater than 20 (of 50, i.e., 40+%) [40] A new data point that is higher or lower than the last expected trajectory range No discrete indicator of client questionnaire response volatility to differentiate volatility from risk. At present, the DLT trajectory output indicates three factors: volatility, uncertainty, and timeliness of the data. Establish an appropriate numerical boundary that flags instances of volatility. Add a volatility indicator that does one of the following when volatility is flagged: Deactivate and restrict access to the DLT trajectory output (i.e., the graph) and display the message: ‘Volatility flagged in clients’ questionnaire responses, speak to the client’. Once volatility is no longer detected, return access rights Add a discrete volatility indicator to distinguish the three factors, consigning the DLT trajectory output to indicate only the uncertainty and timeliness of the data To not inadvertently convey uncertainty with the DLT trajectory output at baseline and risk confusing people that it is conveying volatility at baseline, do not display trajectories at baseline Value-add of predictive trajectories in mental health care (Accuracy: Clinical judgement versus machine learning) Timeliness of the data is not explicitly communicated Add a data timeliness indicator. For example, divide the DLT trajectory output into 1, 2, 3, and 4-week timepoints segments, indicating that as the time increases and the output widens, the timeliness of the data degrades Graph is hard to read (i.e., data points overlap) Add a zoom function on the graph Scepticism about the DLT replacing clinical judgement Feature a reminder on the UI: ‘This tool is not intended to replace clinical judgement. It should be used to support clinical judgement’ DLT trajectory output does not sufficiently convey uncertainty and confidence; some DLT trajectory output ranges are too wide or too narrow, affecting perceived accuracy A wide output range primarily indicates uncertainty based on either volatility in clients’ historical scores or a sudden recent change after a long period of stability. Revise to determine an uncertainty threshold, label the output with this information for clarity, and advise a potential treatment option based on predictions for clinician users to consider and action if deemed appropriate (in this instance, recommend more regular client follow-up appointments and treatment plan review) A narrow output range primarily indicates confidence based on a long period of consistent stability, which participants thought could be interpreted intuitively from the data points. Revise to add multiple confidence levels to the output to evidence the trajectory and possibilities for clarity and advise a subsequent potential treatment option (in this instance, recommend normal frequency follow-up with the client and typical review of their treatment plan) Need more details on the DLT model and data incorporated Flag key driving predictors for DLT trajectory outputs Provide external links to the research evidence base (e.g., academic papers explaining the efficacy and importance of each measure used, mental health dimensions considered, and the model) Illuminating and magnifying the trajectory landscape (More transparency is needed) Need more details on the DLT trajectory output calculations and confidence intervals Highlight and explicate psychiatric constructs underlying SIDAS that inform outputs Add a confidence interval indicator for the outputs. Include a feature to enable users to switch between different confidence intervals (e.g., 68% and 95%) Add a tooltip connected to data points to label significant links between the client’s historical data and trajectory and individual items answered on the SIDAS questionnaire Need to show contextual data typically used for assessing suicidal ideation Feature key client demographics on the UI to help clinicians recall key details about clients in situ and interpret the output (e.g., population group, age, and relationship status). Consider collating this data by specific decisions to be made by the clinician (e.g., data relating to developing a treatment plan, introducing an intervention, or choosing an appropriate assessment scale or measure) Illuminating and magnifying the trajectory landscape (Incorporating real-world considerations) Need to expand the variables incorporated in the DLT model for assessing suicidal ideation Consider incorporating the following data from in-scope measures already captured in Innowell that may be incorporated into the DLT model: For baseline predictive trajectories: Client demographics (e.g., age, sex, and relationship status), Innowell onboarding questionnaires (e.g., Quick Inventory of Depressive Symptomatology), and relationships between parameters For continuous-time predictive trajectories: social connectedness and social and occupational functioning. Incorporating this data may prove useful for data imputation to resolve instances when clients did not self-report suicidal ideation data but provided data on other mental health dimensions to support prediction. In the future, consider incorporating data not yet captured in Innowell, including: Relevant life events in a client’s life or mental health treatment (e.g., the start dates of the school year, new therapy intervention or medication, and a suicide attempt) Measures of impulsivity, quality of life, substance use (alcohol, tobacco, cannabis), behavioural risks (self-harm, homicidal ideation, oppositionality, rule-breaking, sleep disturbances, distress, and violence), and other mental health dimensions (depression, anxiety, post-traumatic stress, psychosis-like and mania-like experiences) Need comparison scores against clients’ population group To contextualise individual-level trajectories in a population to benefit interpretation, add a population group average line on the graph evidencing average SIDAS scores based on population data relevant to the client (e.g., sex, age, educational attainment level, and risk factors). Expand data sources Source additional contextual data using data mining techniques and large language models to search, for instance, client notes, health records, police records, coronial reports, and client session transcripts. Adhere to Australian standards and laws on health data usage, retention, consent, and privacy Concerns about data privacy and protection Overtly communicate adherence to Australian standards and laws concerning data privacy, protection, and cybersecurity on the UI Moving towards implementation (Machine learning amplifies ethical challenges) Data points and the DLT trajectory output were inadequately labelled Add relevant labels to data point labels (e.g., date data gathered, score, and whether the score has increased or decreased since the last data point) Revise DLT trajectory outputs with the abovementioned labels Moving towards implementation (Clear guidance on use is necessary) Graph y-axis is confusing Reframe and revise y-axis directions so that the top represents ‘unhealthy’ or undesirable scores (i.e., suicidal ideation is increasing), and the bottom represents ‘healthy’ or desirable scores (but do not label the bottom to avoid additional confusion) The existing tooltip connected to data points is frustrating and blocks key information Remove and revise as outlined above There is risk associated with showing outputs to clients Do not develop a client view of the DLT Moving towards implementation (Engaging the right users) Prior to implementing the DLT, participants recommended resolving ethical issues, providing clear guidance on how to use the DLT, and identifying and accommodating the appropriate target user groups. Privacy and accountability were key ethical considerations intensified by the application of machine learning in the study context. Personal data privacy and data security, storage, governance, and ownership and accountability for DLT-supported decisions and subsequent client outcomes need to be addressed and made explicit to the user in the DLT UI. Guidance, training, and dynamic tips featured on the UI to support the interpretation of the DLT trajectory outputs are crucial to avoid misinterpretation and associated risks. Participants agreed that the primary users of the DLT should be mental health clinicians. In response to mixed attitudes on whether young people should have access to the DLT and to alleviate risks associated with clients misinterpreting the DLT, client access to the DLT should be at the discretion of their clinician, and a tailored client portal of the DLT should be avoided. As the next steps for this work, co-design of the DLT with key stakeholders in youth mental health services is crucial to ensure value-add, increase system transparency and data incorporated into algorithmic decision-making, and progress appropriate implementation with the right safeguards in place, with the right users in the right environments, and with the right know-how to utilise the DLT in a way that minimises risks and maximises benefits. These findings have broader implications for the ethics, design, development, implementation, governance, and evaluation of machine learning of predictive trajectories and AI in youth mental health. While evidencing prediction accuracy to health practitioners is an enduring problem generally [56], digital health systems face additional concerns about trust and reliability as health data and decision-making are increasingly delegated to digital solutions [57]. Research suggests several methods to alleviate the problem, including conveying system accuracy information with simple, non-technical language [56] and visual representations [57]. Although participants understood the DLT’s language and were presented with a line graph visualisation, difficulties among participants interpreting the graph indicate a need for visualisation redesign. Schneider, Wayrauther (57) suggest that visualising deviations or algorithmic errors between historical health data and predictive trajectories improves understanding by providing reference to real past data, and so redesigning the DLT in this way may be appropriate. While health research advocates for communicating system accuracy by presenting natural frequencies rather than probability [58] (e.g., 1 out of 10 people are positive versus the probability that one person is positive is 10%), human-computer interaction research shows that such analogy-based explanations like natural frequencies do not increase perceived accuracy among expert users (e.g., mental health clinicians) unless adjusted to their needs [59]. Greater stakeholder needs assessment and subsequent personalisation of user experience are necessary to ensure that the DLT accuracy is communicated to the target users in the way they need it to be communicated. To that end, further co-design is needed. Digital health literature highlights the importance of incorporating highly personalised data and communicating to the user what personal data is integrated into algorithmic decision-making to dispel fears about perceived system inaccuracies [60]. This affirms the study finding that incorporating real-world considerations into the DLT and transparently and simplistically communicating this process and data to the users is essential in future iterations of the DLT and other state-of-the-art AI in youth mental health broadly. The findings presented here support existing research that emphasises the need to frame and design AI in youth mental health as a tool to enhance but not replace clinical judgement to ensure value-add going forward [61, 62]. This framing is particularly important to alleviate concerns about potential conflicts between clinical judgement and algorithmic prediction. Furthermore, this framing is key to curbing the evident dichotomy between the realistic amount of data that a mental health professional can reasonably consider in their day-to-day decision-making and their high expectations on the volume of data and data sources to be incorporated into machine learning of predictive trajectories. In this way, data-driven trajectory learning and prediction (such as the DLT’s predictive trajectories) are positioned as another information source to enhance a clinician’s decision, rather than a replacement and improve personalised care services. For the DLT, this could include generating recommendations for follow-up appointment frequency (e.g., advising whether to aim to see a client next week or next month) or indicating treatment effectiveness based on a client’s volatility or stability (see Table 3 for more details), addressing lasting unanswered questions in mental health regarding determining appointment frequency [63] and treatment effectiveness [64]. More broadly, emphasising system simplicity by design to effectively convey system accuracy, clearly highlighting key information incorporated into data-driven decisions and relevant information to inform a clinician’s decisions, and ensuring system ubiquity to work alongside existing systems, workflows, and different therapy approaches will play a key role in establishing the dynamic learning of individual-level trajectories, and AI more broadly, in youth mental health as enhancing rather than replacing human decision-making. Calls for greater transparency of AI in mental health are well-documented in the literature [65]. Yet, limited studies report practical design recommendations relevant to AI systems intended for dynamic trajectory learning and prediction to best communicate explainability to users and, ultimately, improve user experience, trust, comprehension of underlying machine learning models, and interpretation of system outputs. To best address this gap, this study found that AI generally, and data-driven trajectory learning and prediction tools specifically, operating in mental health should directly communicate and evidence, at least, the following: Visually indicate confidence intervals for the trajectory outputs (e.g., explicitly label each output displayed with the appropriate confidence interval). Overtly flag key driving predictors for each output (e.g., list driving predictors alongside associated outputs). Highlight and elucidate different psychiatric constructs underlying clinical assessment measured that are incorporated into outputs (e.g., list constructs alongside associated outputs). Create and show links between outputs and the research knowledge base and proof of the underlying machine learning model efficacy (e.g., feature external links to academic papers evidencing the efficacy and importance of measures used, mental health dimensions considered, and the model used). Identify all data used to generate outputs (e.g., list appropriate data alongside associated outputs, ranked by their corresponding weight). Other studies about consumer-facing AI in health generally have identified and warned of “inappropriate explainability,” i.e., excessive or improper levels of transparency of explanations [66]. Existing research remarks that inappropriate explainability may lead to information overload among users, negative user experiences, user harm caused by ignoring system prompts, accidental disclosure and privacy violation of sensitive consumer data, and counteracting the interests of AI providers [66]. Findings from the present study identify additional dimensions of potential user harm in the youth mental health space. Based on participant feedback, some interactions with AI in mental health and their outputs should not be made available to young clients in mental health services despite potential opportunities for empowerment due to risks associated with misinterpreting outputs, developing negative expectation-induced behaviours, and creating a sense of hopelessness, all of which threaten disempowerment, care service disengagement, low digital tool uptake among clients. This empowerment-disempowerment dichotomy must inevitably be dealt with at the design, development, implementation, and governance stages with an approach that integrates a highly personalised approach to care and AI in mental health design to best tailor outputs to different user groups based on user capacity, needs, values, preferences, expectations, and role in the caring relation [67]. Some limitations of this research should be noted. First, the majority of participants were located in Australian metropolitan areas, and thus, the findings are not generalisable to mental health care delivery practices and services in other countries or regional and rural areas. Second, most participants were psychologists, and therefore, the study lacks a salient representation of other mental health professionals that should be accounted for in the research, design, implementation, use, and governance of DMHTs. Conclusions Identifying risk and predicting future trajectories in youth mental health using conventional measures produces only modest accuracy and replicability, failing to confront the worsening global youth mental health crisis. Although the use of AI applications, such as machine learning, demonstrates a promising alternative, there are key underexplored attitudes and design requirements among mental health professionals in this space. This work addresses this gap by developing three themes, Value-add of predictive trajectories in mental health care , Illuminating and magnifying the trajectory landscape , and Moving towards implementation , reporting a rich understanding of mental health professionals’ attitudes towards AI in youth mental health generally and their design and content requirements for a novel prediction tool, the DLT, specifically. Professionals are willing to use AI and predictive trajectories, but only to enhance their decision-making and clinical judgement. Moreover, they require considerable machine learning model transparency, evident privacy and accountability standards adherence, use guidance, and contextualisation. To advance the design, development, implementation, ethics, governance, and evaluation of AI and prediction tools in youth mental health more broadly, this study also provides stakeholder-informed practical design recommendations to best communicate system explainability to users and thus benefit trust, comprehension, and user experience. Abbreviations AI Artificial intelligence C-SSRS Columbia-Suicide Severity Rating Scale DLT Dynamic Learning Tool DMHTs Digital mental health tools SIDAS Suicidal ideation attribute scale UI User interface Declarations Ethics approval and consent to participate This research received ethical approval from the University of Sydney’s Human Research and Ethics Committee (Project number: 2021/HE000680). Clinical trial number: not applicable. Participants provided informed consent prior to completing the study. Participant informed consent, participant privacy, confidentiality, and deidentification, and secure data transmission, storage, and handling procedures were carried out in accordance with the Declaration of Helsinki. Consent for publication Not applicable. Availability of data and materials Raw data are not available to the public due to data protection and privacy reasons. Anonymised metadata (e.g., coding glossary and code frequency) that support the findings are available from the corresponding author upon reasonable request. Competing interests IBH is the Co-Director, Health and Policy at the Brain and Mind Centre (BMC) University of Sydney. The BMC operates an early-intervention youth services at Camperdown under contract to headspace. He is the Chief Scientific Advisor to, and a 3.2% equity shareholder in, InnoWell Pty Ltd which aims to transform mental health services through the use of innovative technologies. AP, MKC, HML, AT, MV, and FI have nothing to disclose. Funding IBH is supported by a NHMRC L3 Investigator Grant (GNT2016346). FI is supported by an NHMRC EL1 Investigator Grant (GNT2018157). MV was supported by philanthropic funding from The Johnston Fellowship and from other donor(s) who are families affected by mental illness who wish to remain anonymous. This work was supported by the Medical Research Future Fund Grant (MRFCRI000279, Using AI to personalise treatment decisions in youth mental health services). The funder played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript. Authors' contributions Author contributions: AP: Conceptualization, Methodology, Formal analysis, Investigation, Data Curation, Writing–Original Draft, Writing–Review & Editing, Visualization. IBH: Resources, Writing–Review & Editing, Supervision, Project administration, Funding acquisition. MKC: Formal analysis, Writing–Original Draft, Writing–Review & Editing. HML: Formal analysis, Writing–Review & Editing. AT: Formal analysis, Writing–Review & Editing. MV: Software, Writing–Review & Editing. FI: Conceptualization, Methodology, Formal analysis, Resources, Writing–Review & Editing. References McGorry PD, Mei C, Dalal N, Alvarez-Jimenez M, Blakemore S-J, Browne V, et al. The Lancet Psychiatry Commission on youth mental health. Lancet Psychiatry. 2024;11(9):731-74. doi: 10.1016/S2215-0366(24)00163-9. Bertuccio P, Amerio A, Grande E, La Vecchia C, Costanza A, Aguglia A, et al. Global trends in youth suicide from 1990 to 2020: an analysis of data from the WHO mortality database. eClinicalMedicine. 2024;70:1-12. doi: 10.1016/j.eclinm.2024.102506. Lovero KL, Dos Santos PF, Come AX, Wainberg ML, Oquendo MA. Suicide in Global Mental Health. Current Psychiatry Reports. 2023;25(6):255-62. doi: 10.1007/s11920-023-01423-x. Uddin R, Burton NW, Maple M, Khan SR, Khan A. Suicidal ideation, suicide planning, and suicide attempts among adolescents in 59 low-income and middle-income countries: a population-based study. The Lancet Child & Adolescent Health. 2019;3(4):223-33. doi: 10.1016/S2352-4642(18)30403-6. Ahad AA, Sanchez-Gonzalez M, Junquera P. Understanding and Addressing Mental Health Stigma Across Cultures for Improving Psychiatric Care: A Narrative Review. Cureus. 2023;15(5):1-8. doi: 10.7759/cureus.39549. Pizzol D, Trott M, Butler L, Barnett Y, Ford T, Neufeld SAS, et al. Relationship between severe mental illness and physical multimorbidity: a meta-analysis and call for action. BMJ Mental Health. 2023;26:1-5. doi: 10.1136/bmjment-2023-300870. Patel V, Saxena S, Lund C, Kohrt B, Kieling C, Sunkel C, et al. Transforming mental health systems globally: principles and policy recommendations. The Lancet. 2023;402(10402):656-66. doi: 10.1016/S0140-6736(23)00918-2. Iorfino F, Carpenter JS, Cross SPM, Crouse J, Davenport TA, Hermens DF, et al. Social and occupational outcomes for young people who attend early intervention mental health services: a longitudinal study. Medical Journal of Australia. 2022;216(2):87-93. doi: 10.5694/mja2.51308. Hrabok M, Delorme A, Agyapong VIO. Threats to Mental Health and Well-Being Associated with Climate Change. Journal of Anxiety Disorders. 2020;76:1-5. doi: 10.1016/j.janxdis.2020.102295. McGorry P, Gunasiri H, Mei C, Rice S, Gao CX. The youth mental health crisis: analysis and solutions. Frontiers in Psychiatry. 2025;15:1-20. doi: 10.3389/fpsyt.2024.1517533. Nelson B, McGorry PD, Wichers M, Wigman JTW, Hartmann JA. Moving From Static to Dynamic Models of the Onset of Mental Disorder: A Review. JAMA Psychiatry. 2017;74(5):528-34. doi: 10.1001/jamapsychiatry.2017.0001. Hickie IB, Scott EM, Cross SP, Iorfino F, Davenport TA, Guastella AJ, et al. Right care, first time: A highly personalised and measurement-based care model to manage youth mental health. Medical Journal of Australia. 2019;211(S9):S3-S46. doi: 10.5694/mja2.50383. Hornstein S, Forman-Hoffman V, Nazander A, Ranta K, Hilbert K. Predicting therapy outcome in a digital mental health intervention for depression and anxiety: A machine learning approach. DIGITAL HEALTH. 2021;7:1-11. doi: 10.1177/20552076211060659. Wies B, Landers C, Ienca M. Digital Mental Health for Young People: A Scoping Review of Ethical Promises and Challenges. Front Digit Health. 2021;3:1-11. doi: 10.3389/fdgth.2021.697072. Torous J, Bucci S, Bell IH, Kessing LV, Faurholt-Jepsen M, Whelan P, et al. The growing field of digital psychiatry: current evidence and the future of apps, social media, chatbots, and virtual reality. World Psychiatry. 2021;20(3):318-35. doi: 10.1002/wps.20883. Uhlhaas P, Torous J. Digital tools for youth mental health. npj Digital Medicine. 2019;2:1-3. doi: 10.1038/s41746-019-0181-2. McGorry PD, Mei C, Chanen A, Hodges C, Alvarez-Jimenez M, Killackey E. Designing and scaling up integrated youth mental health care. World Psychiatry. 2022;21(1):61-76. doi: 10.1002/wps.20938. Barrigon ML, Romero-Medrano L, Moreno-Muñoz P, Porras-Segovia A, Lopez-Castroman J, Courtet P, et al. One-Week Suicide Risk Prediction Using Real-Time Smartphone Monitoring: Prospective Cohort Study. J Med Internet Res. 2023;25:1-11. doi: 10.2196/43719. Hughes JL, Horowitz LM, Ackerman JP, Adrian MC, Campo JV, Bridge JA. Suicide in young people: screening, risk assessment, and intervention. BMJ. 2023;381:1-16. doi: 10.1136/bmj-2022-070630. Østergård Ole K, Grønnebæk L, Nilsson Kristine K. Do Therapists Know When Their Clients Deteriorate? An Investigation of Therapists' Ability to Estimate and Predict Client Change During and After Psychotherapy. Clinical Psychology & Psychotherapy. 2024;31(6):1-10. doi: 10.1002/cpp.70015. Roza TH, Salgado TA, Machado CS, Watts D, Bebber J, Freitas T, et al. Prediction of Suicide Risk Using Machine Learning and Big Data. In: Passos IC, Rabelo-da-Ponte FD, Kapczinski F, editors. Digital Mental Health: A Practitioner's Guide. Cham: Springer; 2023. p. 173-88. Fonseka TM, Bhat V, Kennedy SH. The utility of artificial intelligence in suicide risk prediction and the management of suicidal behaviors. Australian & New Zealand Journal of Psychiatry. 2019;53(10):954-64. doi: 10.1177/0004867419864428. Daniel DH, Erwin L. AI enabled suicide prediction tools: a qualitative narrative review. BMJ Health Care Inform. 2020;27(3):1-10. doi: 10.1136/bmjhci-2020-100175. Haines-Delmont A, Chahal G, Bruen AJ, Wall A, Khan CT, Sadashiv R, et al. Testing Suicide Risk Prediction Algorithms Using Phone Measurements With Patients in Acute Mental Health Settings: Feasibility Study. JMIR Mhealth Uhealth. 2020;8(6):1-15. doi: 10.2196/15901. McHugh CM, Ho N, Iorfino F, Crouse JJ, Nichles A, Zmicerevska N, et al. Predictive modelling of deliberate self-harm and suicide attempts in young people accessing primary care: a machine learning analysis of a longitudinal study. Social Psychiatry and Psychiatric Epidemiology. 2023;58(6):893-905. doi: 10.1007/s00127-022-02415-7. Iorfino F, Ho N, Carpenter JS, Cross SP, Davenport TA, Hermens DF, et al. Predicting self-harm within six months after initial presentation to youth mental health services: A machine learning study. PloS one. 2021;15(12):1-16. doi: 10.1371/journal.pone.0243467. Franklin JC, Ribeiro JD, Fox KR, Bentley KH, Kleiman EM, Huang X, et al. Risk factors for suicidal thoughts and behaviors: A meta-analysis of 50 years of research. Psychol Bull. 2017;143(2):187-232. doi: 10.1037/bul0000084. Wang SB, Van Genugten RDI, Yacoby Y, Pan W, Bentley KH, Bird SA, et al. Building personalized machine learning models using real-time monitoring data to predict idiographic suicidal thoughts. Nature Mental Health. 2024;2(11):1382-91. doi: 10.1038/s44220-024-00335-w. Lutz W, Schwartz B, Delgadillo J. Measurement-Based and Data-Informed Psychological Therapy. Annu Rev Clin Psychol. 2022;18:71-98. doi: 10.1146/annurev-clinpsy-071720-014821. Sinclair C, Holloway K, Riley G, Auret K. Online Mental Health Resources in Rural Australia: Clinician Perceptions of Acceptability. J Med Internet Res. 2013;15(9):1-10. doi: 10.2196/jmir.2772. Ganapathy A, Clough BA, Casey LM. Organizational and Policy Barriers to the Use of Digital Mental Health by Mental Health Professionals. Telemedicine and e-Health. 2021;27(12):1332-43. doi: 10.1089/tmj.2020.0455. Lamela D, Cabral J, Coelho S, Jongenelen I. Personal stigma, determinants of intention to use technology, and acceptance of internet-based psychological interventions for depression. International Journal of Medical Informatics. 2020;136:1-6. doi: 10.1016/j.ijmedinf.2020.104076. Nogueira-Leite D, Diniz JM, Cruz-Correia R. Mental Health Professionals’ Attitudes Toward Digital Mental Health Apps and Implications for Adoption in Portugal: Mixed Methods Study. JMIR Hum Factors. 2023;10:1-19. doi: 10.2196/45949. Zhang X, Lewis S, Chen X, Berry N, Bucci S. Mental health professionals views and the impact of COVID-19 pandemic on implementing digital mental health in China: A nationwide survey study. Internet Interv. 2022;30:1-11. doi: 10.1016/j.invent.2022.100576. Rifkin-Zybutz R, Turner N, Derges J, Bould H, Sedgewick F, Gooberman-Hill R, et al. Digital Technology Use and Mental Health Consultations: Survey of the Views and Experiences of Clinicians and Young People. JMIR Ment Health. 2023;10:1-14. doi: 10.2196/44064. Bucci S, Berry N, Morris R, Berry K, Haddock G, Lewis S, et al. “They Are Not Hard-to-Reach Clients. We Have Just Got Hard-to-Reach Services.” Staff Views of Digital Health Tools in Specialist Mental Health Services. Frontiers in Psychiatry. 2019;10:1-14. doi: 10.3389/fpsyt.2019.00344. Cliffe B, Croker A, Denne M, Stallard P. Clinicians' use of and attitudes towards technology to provide and support interventions in child and adolescent mental health services. Child and Adolescent Mental Health. 2020;25(2):95-101. doi: 10.1111/camh.12362. Atmakuru A, Shahini A, Chakraborty S, Seoni S, Salvi M, Hafeez-Baig A, et al. Artificial intelligence-based suicide prevention and prediction: A systematic review (2019–2023). Information Fusion. 2025;114:1-29. doi: 10.1016/j.inffus.2024.102673. D’Hotman D, Loh E, Savulescu J. AI-enabled suicide prediction tools: ethical considerations for medical leaders. BMJ Leader. 2021;5(2):102-7. doi: 10.1136/leader-2020-000275. Varidel M, Hickie IB, Prodan A, Skinner A, Marchant R, Cripps S, et al. Dynamic learning of individual-level suicidal ideation trajectories to enhance mental health care. npj Mental Health Research. 2024;3(1):26. doi: 10.1038/s44184-024-00071-0. van Spijker BAJ, Batterham PJ, Calear AL, Farrer L, Christensen H, Reynolds J, et al. The Suicidal Ideation Attributes Scale (SIDAS): Community-Based Validation Study of a New Scale for the Measurement of Suicidal Ideation. Suicide and Life-Threatening Behavior. 2014;44(4):408-19. doi: 10.1111/sltb.12084. Iorfino F, Cross SP, Davenport T, Carpenter JS, Scott E, Shiran S, et al. A Digital Platform Designed for Youth Mental Health Services to Deliver Personalized and Measurement-Based Care. Front Psychiatry. 2019;10:1-9. doi: 10.3389/fpsyt.2019.00595. Capon W, Hickie IB, McKenna S, Varidel M, Richards M, LaMonica HM, et al. Characterising variability in youth mental health service populations: A detailed and scalable approach using digital technology. Australasian Psychiatry. 2023;31(3):295-301. doi: 10.1177/10398562231167681. Posner K, Brown GK, Stanley B, Brent DA, Yershova KV, Oquendo MA, et al. The Columbia–Suicide Severity Rating Scale: Initial Validity and Internal Consistency Findings From Three Multisite Studies With Adolescents and Adults. American Journal of Psychiatry. 2011;168(12):1266-77. doi: 10.1176/appi.ajp.2011.10111704. Papoutsi C, Wherton J, Shaw S, Morrison C, Greenhalgh T. Putting the social back into sociotechnical: Case studies of co-design in digital health. Journal of the American Medical Informatics Association. 2021;28(2):284-93. doi: 10.1093/jamia/ocaa197. Phillippi J, Lauderdale J. A guide to field notes for qualitative research: Context and conversation. Qualitative Health Research. 2018;28(3):381-8. doi: 10.1177/1049732317697102. Marshall B, Cardon P, Poddar A, Fontenot R. Does Sample Size Matter in Qualitative Research?: A Review of Qualitative Interviews in is Research. Journal of Computer Information Systems. 2013;54(1):11-22. doi: 10.1080/08874417.2013.11645667. Malterud K, Siersma VD, Guassora AD. Sample Size in Qualitative Interview Studies: Guided by Information Power. Qualitative Health Research. 2015;26(13):1753-60. doi: 10.1177/1049732315617444. Christie GI, Shepherd M, Merry SN, Hopkins S, Knightly S, Stasiak K. Gamifying CBT to deliver emotional health treatment to young people on smartphones. Internet Interv. 2019;18:1-9. doi: 10.1016/j.invent.2019.100286. Bevan Jones R, Stallard P, Agha SS, Rice S, Werner-Seidler A, Stasiak K, et al. Practitioner review: Co-design of digital mental health technologies with children and young people. Journal of Child Psychology and Psychiatry. 2020;61(8):928-40. doi: 10.1111/jcpp.13258. Braun V, Clarke V. Toward good practice in thematic analysis: Avoiding common problems and be(com)ing a knowing researcher. International Journal of Transgender Health. 2023;24(1):1-6. doi: 10.1080/26895269.2022.2129597. Braun V, Clarke V. Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health. 2019;11(4):589-97. doi: 10.1080/2159676X.2019.1628806. Braun V, Clarke V. One size fits all? What counts as quality practice in (reflexive) thematic analysis? Qualitative Research in Psychology. 2021;18(3):328-52. doi: 10.1080/14780887.2020.1769238. Braun V, Clarke V. Thematic analysis. In: Cooper H, Coutanche MN, McMullen LM, Panter AT, Rindskopf D, Sher KJ, editors. APA handbook of research methods in psychology: Research designs: Quantitative, qualitative, neuropsychological, and biological. 2nd ed. Washington, D.C.: American Psychological Association; 2023. p. 65-81. Braun V, Clarke V. A critical review of the reporting of reflexive thematic analysis in Health Promotion International. Health Promot Int. 2024;39(3):1-12. doi: 10.1093/heapro/daae049. Steurer J, Fischer JE, Bachmann LM, Koller M, ter Riet G. Communicating accuracy of tests to general practitioners: a controlled study. BMJ. 2002;324:824-6. doi: 10.1136/bmj.324.7341.824. Schneider H, Wayrauther J, Hassib M, Butz A. Communicating Uncertainty in Fertility Prognosis. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. Glasgow, Scotland UK: ACM; 2019. p. 1-11. Whiting PF, Davenport C, Jameson C, Burke M, Sterne JAC, Hyde C, et al. How well do health professionals interpret diagnostic information? A systematic review. BMJ open. 2015;5(7):1-8. doi: 10.1136/bmjopen-2015-008155. He G, Buijsman S, Gadiraju U. How Stated Accuracy of an AI System and Analogies to Explain Accuracy Affect Human Reliance on the System. Proc ACM Hum-Comput Interact. 2023;7(CSCW2):1-29. doi: 10.1145/3610067. He X, Hong Y, Zheng X, Zhang Y. What Are the Users’ Needs? Design of a User-Centered Explainable Artificial Intelligence Diagnostic System. International Journal of Human–Computer Interaction. 2023;39(7):1519-42. doi: 10.1080/10447318.2022.2095093. Xian X, Chang A, Xiang Y-T, Liu MT. Debate and Dilemmas Regarding Generative AI in Mental Health Care: Scoping Review. Interact J Med Res. 2024;13:1-18. doi: 10.2196/53672. Ramadan OME, Alruwaili MM, Alruwaili AN, Elsehrawy MG, Alanazi S. Facilitators and barriers to AI adoption in nursing practice: a qualitative study of registered nurses' perspectives. BMC Nursing. 2024;23:1-16. doi: 10.1186/s12912-024-02571-y. Cousineau M, Verter V, Turecki G. Impact of Psychiatric Follow-Up Frequency on Outcomes and Waiting Times. Am J Manag Care. 2024;30(2):e52-e8. doi: 10.37765/ajmc.2024.89501. Hawthorne SCC, Williams-Wengerd A. Is treatment helping? How providers gauge effectiveness in treating serious mental illness. SSM - Mental Health. 2022;2:1-8. doi: 10.1016/j.ssmmh.2022.100110. Joyce DW, Kormilitzin A, Smith KA, Cipriani A. Explainable artificial intelligence for mental health through transparency and interpretability for understandability. npj Digital Medicine. 2023;6:1-7. doi: 10.1038/s41746-023-00751-9. He X, Zheng X, Ding H. Existing Barriers Faced by and Future Design Recommendations for Direct-to-Consumer Health Care Artificial Intelligence Apps: Scoping Review. J Med Internet Res. 2023;25:1-22. doi: 10.2196/50342. Moggia D, Lutz W, Brakemeier E-L, Bickman L. Treatment Personalization and Precision Mental Health Care: Where are we and where do we want to go? Administration and Policy in Mental Health and Mental Health Services Research. 2024;51:1-6. doi: 10.1007/s10488-024-01407-w. Additional Declarations Competing interest reported. IBH is the Co-Director, Health and Policy at the Brain and Mind Centre (BMC) University of Sydney. The BMC operates an early-intervention youth services at Camperdown under contract to headspace. He is the Chief Scientific Advisor to, and a 3.2% equity shareholder in, InnoWell Pty Ltd which aims to transform mental health services through the use of innovative technologies. AP, MKC, HML, AT, MV, and FI have nothing to disclose. Supplementary Files Additionalfile1.pdf Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 29 Nov, 2025 Reviewers agreed at journal 13 Nov, 2025 Reviews received at journal 02 Sep, 2025 Reviewers agreed at journal 28 Aug, 2025 Reviews received at journal 26 Aug, 2025 Reviewers agreed at journal 26 Aug, 2025 Reviewers invited by journal 09 Jul, 2025 Editor assigned by journal 09 Jul, 2025 Editor invited by journal 09 Jul, 2025 Submission checks completed at journal 09 Jul, 2025 First submitted to journal 09 Jul, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7062632","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":483201482,"identity":"ddfa838f-2926-4b80-8a3e-d7c57cc13979","order_by":0,"name":"Adam Poulsen","email":"data:image/png;base64,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","orcid":"","institution":"The University of Sydney","correspondingAuthor":true,"prefix":"","firstName":"Adam","middleName":"","lastName":"Poulsen","suffix":""},{"id":483201483,"identity":"6a96a664-b4b0-4533-bfea-6a275c2d86e1","order_by":1,"name":"Ian B Hickie","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Ian","middleName":"B","lastName":"Hickie","suffix":""},{"id":483201484,"identity":"392d981b-df84-4b24-8c46-ab1073ecf576","order_by":2,"name":"Min K Chong","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"K","lastName":"Chong","suffix":""},{"id":483201485,"identity":"632e07c7-a243-4aa1-9ef3-f9d783aecac8","order_by":3,"name":"Haley M LaMonica","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Haley","middleName":"M","lastName":"LaMonica","suffix":""},{"id":483201486,"identity":"b555b213-dff7-4a22-a9ff-0c90017e7355","order_by":4,"name":"Ashlee Turner","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Ashlee","middleName":"","lastName":"Turner","suffix":""},{"id":483201487,"identity":"2047d426-f6e7-4f07-a8f7-3a476d7c8085","order_by":5,"name":"Mathew Varidel","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Mathew","middleName":"","lastName":"Varidel","suffix":""},{"id":483201488,"identity":"dc36c15d-49ca-4866-bb8c-4bfe61e023fe","order_by":6,"name":"Frank Iorfino","email":"","orcid":"","institution":"The University of Sydney","correspondingAuthor":false,"prefix":"","firstName":"Frank","middleName":"","lastName":"Iorfino","suffix":""}],"badges":[],"createdAt":"2025-07-07 07:53:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7062632/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7062632/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":86697657,"identity":"6c5a1171-c743-4d63-ac5d-0a709c1e7c8e","added_by":"auto","created_at":"2025-07-14 15:42:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":296613,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eInnowell platform (clinician’s and client’s view).\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/a100c15b9b1ab8d4aabf0b10.png"},{"id":86697885,"identity":"b2c21559-7327-4817-8735-00c3771dcfb9","added_by":"auto","created_at":"2025-07-14 15:50:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":204355,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe Dynamic Learning Tool (DLT) (clinician-only view).\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/006faea340468386e9d9f2a9.png"},{"id":86697660,"identity":"19638c07-388b-4807-bba0-d9a7c298465c","added_by":"auto","created_at":"2025-07-14 15:42:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":230713,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFinal thematic map.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/c679242597447f9c234f890d.png"},{"id":86697661,"identity":"55fea004-f9d4-4d16-936a-8c970f2aad53","added_by":"auto","created_at":"2025-07-14 15:42:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":179480,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eCombined initial and final thematic maps.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/6888b33e784492bdb1bd19dc.png"},{"id":86697663,"identity":"8bba0a2a-4dfc-485e-82be-39670887bb5a","added_by":"auto","created_at":"2025-07-14 15:42:55","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":161878,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDLT trajectory output examples shown to participants during user testing.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/414a0be45db124c0f158e95f.png"},{"id":86699167,"identity":"e725f49c-3dad-4ae9-98d0-758a8b7d293e","added_by":"auto","created_at":"2025-07-14 15:58:56","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1985496,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/c3a82215-e816-460b-b60e-d4c32d56b937.pdf"},{"id":86697884,"identity":"92bfbc79-5009-44ea-8a55-d88e1bc01962","added_by":"auto","created_at":"2025-07-14 15:50:55","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":130273,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7062632/v1/228b16279a122a5f60a80036.pdf"}],"financialInterests":"Competing interest reported. IBH is the Co-Director, Health and Policy at the Brain and Mind Centre (BMC) University of Sydney. The BMC operates an early-intervention youth services at Camperdown under contract to headspace. He is the Chief Scientific Advisor to, and a 3.2% equity shareholder in, InnoWell Pty Ltd which aims to transform mental health services through the use of innovative technologies. AP, MKC, HML, AT, MV, and FI have nothing to disclose.","formattedTitle":"Mental health professionals’ perspectives on dynamic learning of individual-level trajectories in youth mental health care: A qualitative study","fulltext":[{"header":"Background","content":"\u003cp\u003eGlobally, youth mental health is deteriorating [1], and youth suicide rates remain alarming [2-4]. Mental ill health is a multifaceted problem impacted by environmental, social, economic, political, and technological change stressors and persistent stigma and discrimination about mental health [1, 5, 6]. Young people are particularly vulnerable as many mental disorders emerge during developmental stages [7], further compounded by contemporary challenges such as climate change, economic instability, intergenerational inequity, and unregulated social media exposure [1, 8, 9].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe high demand for mental health care necessitates more effective strategies and systems for indicated prevention and early intervention [10]. Current systems struggle to identify and respond to risk in a timely manner, relying heavily on static, one-off sampling of clinical state and risk markers (e.g., based on single baseline assessments) to predict clients\u0026rsquo; mental health trajectories, yielding only modest accuracy and replicability [11]. Responding to these multidimensional problems requires a highly personalised and real-time measurement-based approach that incorporates dynamic change into predictive modelling to better understand individual illness trajectories in young people [11, 12]. The continued digitisation of mental health, especially with artificial intelligence (AI) applications such as machine learning, represents a crucial advancement [13]. Digital mental health tools (DMHTs) have been shown to improve mental health outcomes [14] and are positioned to advance the accessibility, efficiency, and scalability of interventions to complement traditional care and services [15-17].\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eYouth suicide remains a significant challenge requiring urgent attention. Current suicidal ideation assessment, conducted primarily through needs-based clinical interviews, has limited temporal validity and frequently raises false positives and negatives [18]. Moreover, the dynamic nature of suicidal ideation makes accurate prediction of trajectories difficult [19]. Clinicians therefore need more objective and precise tools to supplement their clinical judgement and predict individual suicidal ideation trajectories [20, 21]. Big data and machine learning models, applied through DMHTs, offer promising opportunities for dynamic trajectory learning and prediction of suicidal ideation in mental health services. These technologies can help to overcome the highly subjective nature and inaccuracy of human-driven prediction [22, 23] and advance towards greater personalisation through, for instance, gathering user-generated and passive sensor data relevant to suicidal ideation trajectories (e.g., sleep behaviour data collected via smartwatches) [18, 24]. To date, the accuracy of algorithmically modelling suicidal ideation trajectories demonstrates promising results [18, 22, 23, 25, 26]. This data-driven, individual-level, ongoing approach represents an imperative shift from conventional risk stratification-based prediction towards highly personalised and measurement-based care [12, 27-29].\u003c/p\u003e\n\u003cp\u003eThis digital, personalised approach would be a transformative change in traditional mental health care. Therefore, understanding mental health professionals\u0026rsquo; attitudes and requirements concerning DMHTs is a crucial step before implementation [30], particularly for identifying possible barriers [31] and determining factors affecting usefulness, acceptance, and effectiveness [32]. Research on DMHTs reports a diversity of perspectives among mental health professionals [33-37]. In the United Kingdom, a 2019 online survey of child and adolescent mental health services clinical staff found that mental health professionals considered DMHTs helpful in their clinical work, advancing accessibility, convenience, and appeal to young clients, though they expressed little knowledge about the safety, security, privacy, and reliability of DMHTs used [37]. Similarly, a 2023 Portuguese study revealed that while 87.2% of professionals supported prescribing DMHTs to clients, only 43.6% planned to implement them, citing barriers such as a lack of information about DMHTs, required training effort, and the need to adjust clinical processes and records [33]. Professionals also reported key determinants of DMHT adoption that were not adequately met, including information about alignment with health objectives, evidence of DMHT validity, and expert-based recommendations for specific DMHTs [33]. A 2022 study in China by Zhang, Lewis (34) reported that clinicians held generally positive attitudes towards DMHTs, but they expressed concerns about online safety, social and cultural barriers, and limited digital competence among professionals and clients.\u003c/p\u003e\n\u003cp\u003eDespite the broader evidence on DMHTs, key underexplored attitudes and design requirements among mental health professionals about machine learning of predictive trajectories include clinical utility, safety, ethics, effectiveness, permissibility, user interface (UI) design, implementation, and governance [22-24]. Going forward, the development and evaluation of implementable, scalable, and culturally sensitive machine learning models and tools for predicting suicidal ideation trajectories requires a deeper understanding of these factors, particularly regarding acceptability, feasibility, cost-effectiveness, real-world implementation, ethical considerations, and potential impact to ultimately realise the full potential of these technologies [38, 39].\u003c/p\u003e\n\u003cp\u003eThus, this work aims to explore mental health professionals\u0026rsquo; attitudes towards AI (e.g., machine learning models) in youth mental health and determine the design and content requirements for a novel DMHT, the Dynamic Learning Tool (DLT). Principally, the DLT provides predictive trajectories of individual clients\u0026rsquo; suicidal ideation utilising machine learning to inform clinical decision-making by youth mental health professionals and is integrated into an existing platform used for clinical practice in both Australia and Canada. As outlined in an earlier technical paper [40], the DLT operates by: (1) using longitudinal suicidal ideation data collected in a clinical setting, captured with the suicidal ideation attribute scale (SIDAS) [41], (2) generating predicted trajectories of an individual\u0026rsquo;s suicidal ideation level and variability to guide clinical decisions and determine treatment progress (e.g., changes in the frequency that individuals may need to be observed), (3) incorporating a continuous-time modelling framework to handle irregularly time-spaced observations common in clinical data and update predictions as new observations are collected [40]. For the purposes of this study, the tool was incorporated into an online platform, Innowell (see Figure 1), which supports self-reporting, management, and monitoring of mental ill health and maintenance of well-being [42, 43]. As a precursor to the DLT, Innowell also incorporates static categorical suicide risk assessment indicators (low, moderate, or high risk) (see Figures 1a and 1c), based on self-reported Columbia-Suicide Severity Rating Scale (C-SSRS) [44] data. Figure 2 visualises the DLT as integrated into Innowell. Per best practices in developing digital health solutions [45], this study employs a co-design methodology using semi-structured interviews and usability testing to understand mental health professionals\u0026rsquo; attitudes and co-create design and content requirements for the DLT.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003eStudy setting\u003c/p\u003e\n\u003cp\u003eThis study was conducted primarily online by a researcher in Sydney, Australia, from June 20\u003csup\u003eth\u003c/sup\u003e to August 23\u003csup\u003erd\u003c/sup\u003e, 2024, via the Zoom videoconferencing software, enabling participants across Australia, Canada, and the United States to participate.\u003c/p\u003e\n\u003cp\u003eEthics approval\u003c/p\u003e\n\u003cp\u003eThis research received ethical approval from the University of Sydney\u0026rsquo;s Human Research and Ethics Committee (Project number: 2021/HE000680). Participants provided informed consent prior to completing the study.\u003c/p\u003e\n\u003cp\u003eSampling and data collection\u003c/p\u003e\n\u003cp\u003ePurposive sampling was utilised to recruit participants for this study. Eligibility criteria included mental health clinicians who assess and treat young people or adults (including mental health service managers and administrators in this context) and required English language proficiency. The researchers identified potential participants (e.g., known professionals and local professionals advertising services online) and contacted them via email. Additional recruitment was facilitated through mental health organisations, professional associations, and university departments that agreed to advertise the study details. Potential participants joined the study by completing an online form via Qualtrics, which provided them with a participant information sheet and allowed them to consent electronically. Qualtrics automatically notified the researchers when a participant joined the study, sending the participant contact details to schedule the interview. Participants were free to select a convenient interview time and mode of participation (i.e., in-person and remote videoconferencing participation were offered). Two participants elected to participate in person at the Brain and Mind Centre. The average interview length was approximately 47 minutes (shortest=28, longest=64). Interview audio and visuals were recorded and later transcribed verbatim and de-identified. Researcher notes were taken throughout the study to collect rich contextual information and benefit transcription [46]. Data collection stopped when the sample size was within a pre-determined range (15-25 participants for single case studies [47]) following the information power principle [48], which takes into consideration the study aim, sample specificity, theoretical background, quality of dialogue, and strategy for analysis. Following this principle, a study needs the least amount of participants when the study primarily addresses a narrow aim (i.e., attitudes about AI in mental health), the participants represent similar experiences and knowledge (i.e., mental health clinicians), the study is grounded in existing theory (i.e., personalised and measurement-based care and co-design), an experienced interviewer conducts it, and it addresses a single case analysis.\u003c/p\u003e\n\u003cp\u003eFollowing a co-design methodology, the interviews involved questioning as well as a think-aloud activity as a usability test [49, 50] of the DLT, adhering to the following procedure. Participants were informed about the study aim, interview procedure, DLT, and Innowell. Consent to record the audio and visuals of the interview was reaffirmed. The researcher shared their computer screen to demonstrate how to access the DLT online and answer any related participant questions. Participants freely tested the DLT, completing a think-aloud activity that involved them verbalising what they see, do, think, and feel while testing the tool [49, 50], with researcher support (e.g., answering participant questions about the tool) provided throughout. A semi-structured interview guide of open-ended questions and possible follow-up questions was developed for this study and followed. The questions were adapted to ensure a natural conversational flow and account for numerous other factors, such as a participant\u0026rsquo;s digital literacy. Unguided usability testing without extensive instruction and training simulated real-world first-time use of the DLT, eliciting data relating to tool implementation and governance. Box 1 presents the interview agenda topics that guided the interview (see Additional file 1 for the semi-structured interview guide).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBox 1. Interview agenda topics.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 100%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAcceptability\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eInitial thoughts\u003c/li\u003e\n \u003cli\u003eAppropriateness\u003c/li\u003e\n \u003cli\u003eIntended users\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eClinical utility\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eFit for the intended purpose\u003c/li\u003e\n \u003cli\u003ePotential uses\u003c/li\u003e\n \u003cli\u003eFunction, utility, and outputs\u003c/li\u003e\n \u003cli\u003eImpact on clinical decisions, practice, services, and clients\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eDesign\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eUsability\u003c/li\u003e\n \u003cli\u003eEngagement\u003c/li\u003e\n \u003cli\u003eAesthetics\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eClinical workflow and implementation\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eIntegration into practice\u003c/li\u003e\n \u003cli\u003eIntegration into mental health services\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eEthics and safety\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eEthical and safety considerations\u003c/li\u003e\n \u003cli\u003eTrust\u003c/li\u003e\n \u003cli\u003eTransparency\u003c/li\u003e\n \u003cli\u003ePrivacy\u003c/li\u003e\n \u003cli\u003eProfessional codes of conduct\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eDMHTs\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eExisting tools used\u003c/li\u003e\n \u003cli\u003ePersonalised and measurement-based care\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003e\u003cstrong\u003eNext steps\u003c/strong\u003e\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eFeatures wanted\u003c/li\u003e\n \u003cli\u003eFeedback on proposed future features\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eIn total, 21 mental health professionals were interviewed between June\u0026ndash;August 2024 in Australia (n=18), the United States (n=2), and Canada (n=1). Participants comprised clinical psychologists (n=12), clinical neuropsychologists (n=3), and a clinical nurse consultant, mental health nurse, occupational therapist, social worker, service manager, and registered psychologist. A total of 19 participants (mean age=38 years; range=25-63; 58% female; 42% male) provided further demographic data. Summarised demographics are presented in Table 1. Notably, most participants worked in metropolitan areas (n=16), and a few reported working in a regional area (n=2) or rural area (n=1). Furthermore, most participants working in multiple sectors also worked in private practice (5/6).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Participants\u0026rsquo; demographic data.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVariable\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNumber (n)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePercentage (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u003cem\u003e(n=21)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCountry\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eAustralia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e85.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eUnited States\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e9.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eCanada\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eOccupation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePsychologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e57.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eClinical neuropsychologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e14.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eSocial worker\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eService manager\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eClinical nurse consultant\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eMental health nurse\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eOccupational therapist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eRegistered psychologist\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u003cem\u003e(n=19)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e25-29\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e26.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e30-34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e21.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e35-39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e40-44\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e45-49\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e50-54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e55-59\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e60-64\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGender\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e11\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e57.9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e42.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHighest educational attainment\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eDoctorate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e12\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e63.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eMaster\u0026rsquo;s degree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e21.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eBachelor\u0026rsquo;s degree\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eWork sector\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eAcademia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eCommunity organisation/NGO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePublic\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eHospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePrivate\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eMultiple\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePrivate and academia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePrivate, community organisation/NGO, and academia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePrivate and community organisation/NGO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003ePublic and academia\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eYears in profession\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e0-5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e15.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e6-10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e42.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e11-20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e31.6\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e21+\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGeographical area\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eMetropolitan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e84.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eRegional\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e10.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 65.1685%;\"\u003e\n \u003cp\u003eRural\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0883%;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 19.7432%;\"\u003e\n \u003cp\u003e5.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAnalysis\u003c/p\u003e\n\u003cp\u003eInductive reflexive thematic analysis was applied in this research to develop and report themes based on patterns of shared meaning (or meaning-based interpretative stories [51]) in the interview data without a pre-determined set of themes [52, 53]. This study adhered to conventional thematic techniques, following six iterative analytic phases: data familiarisation, generating initial codes, generating candidate themes, reviewing candidate themes and establishing final themes, defining and naming final themes, and producing the report [54]. NVivo 14 was utilised to organise and code the interview data.\u003c/p\u003e\n\u003cp\u003ePer best practices for reflexive thematic analysis [55], the remainder of this section reports theoretical assumptions that underpin the analysis and details how thematic analysis was implemented in this study. Given that reflexive analysis never neatly fits into either an inductive or deductive approach and it instead combines both to some extent as researchers must make some assumptions (e.g., what pieces of information are worth coding), one key theoretical assumption underpins the analysis. That is, clinical utility, acceptance, and design are key determinants of DMHT use, and thus these factors were considered foci during the analysis process.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eResearcher reflexivity was crucial throughout the analysis to confront and interrogate assumptions [52]. Here, thematic analysis was implemented as follows. Partway through data collection, two coauthors (AP and HML) independently read and engaged in familiarisation of three interview transcripts and generated initial codes. The initial codes were reviewed and discussed, resulting in numerous codes being merged, discarded, and revised and new codes being created. Following this review, noting an adaption of the reflexive thematic analysis method in this study [55], a mutable coding glossary was developed by AP and HML to create a shared language for the remaining analysis processes. When data collection stopped, three coauthors (AP, MKC, and AT) each independently reviewed and coded a portion of the remaining transcripts with reference to the flexible coding glossary. Then, AP, MKC, and AT met to iteratively generate, review, and label candidate themes and subthemes in consultation with the senior author (FI). After that, AP independently examined the candidate themes and subthemes, leading to the themes being renamed and several subthemes being merged. Finally, AP and FI established the set of final themes and subthemes and their labels.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eThree themes and seven subthemes were developed through reflexive thematic analysis of the interview data. Following best practices for reporting thematic analysis [54], Figure 3 presents the final thematic map. Figure 4 illustrates the combined initial and final thematic maps, indicating four additional candidate subthemes that were merged into the final set of themes. Table 2 reports the themes, subthemes, associated descriptions, and illustrative participant quotes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Themes, subthemes, and illustrative participant quotes.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTheme\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSubtheme\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eIllustrative participant quote\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eValue-add of predictive trajectories in mental health care\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eClinical utility in the \u0026lsquo;care as usual\u0026rsquo; milieu\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;This is the problem with services in acute care. It\u0026apos;s all about symptom reduction, not resolution. That\u0026apos;s where I come in, and I, trust me, I had to headbutt the brick wall a lot with other mental health services that are supposed to jump in and help us, but they don\u0026apos;t. So, if I\u0026apos;ve got this, I can then actually provide evidence to a service we can cut and paste this into a report per se, and say, \u0026ldquo;This is what we\u0026apos;ve seen. These were the events\u0026rdquo;, and it provides some historical data to the service to say, \u0026ldquo;This has been your [a client\u0026rsquo;s] input. This is where they\u0026apos;ve come from, this is where they\u0026apos;re going to, and we don\u0026apos;t know where this person\u0026apos;s going to go\u0026rdquo;\u0026rsquo; (P2 Mental health nurse)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eAccuracy: Clinical judgement versus machine learning\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;And seeing some of these predictions, the predictions aren\u0026apos;t accurate enough. [\u0026hellip;] We\u0026apos;re not very good at saying what level of confidence we have for this individual yet. I do think it\u0026apos;s an area of great interest and importance for clinicians. But, it still feels like there\u0026apos;s a lot to do in terms of the data itself, the accuracy and the validity of that data and the predictions because clinicians really need to be convinced if they\u0026apos;re going to take an action that is against what they would have taken without that data. [\u0026hellip;] If you\u0026apos;ve got data and you\u0026apos;ve got ways to show people that they can increase their faith in these predictions, they\u0026apos;re now likely to change their clinical decision. If not, they might consider it in amongst 100 other things they\u0026apos;re considering about what to do with this person\u0026rsquo; (P6 Clinical psychologist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eIlluminating and magnifying the trajectory landscape\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eMore transparency is needed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;I\u0026apos;d want to know a bit more about this algorithm. And, beyond the black box of how it is actually making this prediction, how seriously I need to take it and what the margin for error is. I can see the margin error, I guess, I\u0026apos;m inferring that the bigger the [trajectory] range in the display over time infers that confidence. [\u0026hellip;] In terms of AI clinical decision tools, there\u0026apos;s a lot written around the transparency for a clinician about how it is calculating this and how it is arriving at this conclusion. So, taking that information out of the black box a bit; otherwise, you just won\u0026apos;t have trust in the model. I don\u0026apos;t know why that\u0026apos;s telling me that there\u0026apos;s a high risk of this person getting worse or a low risk of this person getting worse\u0026rsquo; (P6 Clinical psychologist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eIncorporating real-world considerations\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;I\u0026apos;d love to know if Innowell was able to take into account things like age, marital status [\u0026hellip;] social connection, or if there\u0026apos;s any extra bits because I think that there might be people who would say, \u0026quot;No, I haven\u0026apos;t had any [suicidal] thoughts\u0026quot;, or they wouldn\u0026apos;t want to talk about it. But, they might still be someone who [\u0026hellip;] statistically is at a greater risk, or is not. They\u0026apos;re coming in, and they\u0026apos;re really starting to neglect their personal hygiene, and you notice those things, or just those other factors that I would be taking into account with the risk assessment on top of these [SIDAS and C-SSRS] questions. [\u0026hellip;] I think demographic stuff, and that could be built into a tool like this. And I know there are subsets of the population who are at increased risk, and there are determinants that we know of or that we know make people at increased risk. So that would be cool if that was built in as well\u0026rsquo; (P9 Registered psychologist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 13px;\"\u003e\n \u003cp\u003eMoving towards implementation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eMachine learning amplifies ethical challenges\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;I remember being trained and specifically being told, \u0026lsquo;Cover your arse.\u0026rsquo; That is your job. Don\u0026apos;t expose yourself to risk. [\u0026hellip;] If this person goes into hospital tomorrow, have you done all the steps? Have you done all the right things? Did you miss something that you should have not missed? And so, for a tool like this [the DLT] for me, it is about providing the guidelines to ensure that I have done those steps. I have checked in. I have delivered a safety plan. I have cross-validated that right now, I understand that the guidelines tell me that this is probably the best recommended care. This is the evidence-based care. Have I done that, and have I box ticked that? Yes, I have. Did I miss anything? No, I have not\u0026rsquo; (P7 Clinical psychologist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eClear guidance on use is necessary\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026lsquo;Clinicians will need to be trained on how to read it for sure. I think once you\u0026apos;ve explained that to me, now I can kind of see what you\u0026apos;re talking about. It\u0026apos;s not evident very clearly at the beginning what the dots and the lines mean, and the grey parts [DLT trajectory output] mean? [\u0026hellip;] When does this come in handy? [\u0026hellip;] Is it to help support us in risk assessment? Or is it to help us support in what we could do to bump it up higher towards the more healthy kind of aspect?\u0026rsquo; (P17 Occupational therapist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 12px;\"\u003e\n \u003cp\u003eEngaging the right users\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 74px;\"\u003e\n \u003cp\u003e\u003cem\u003e\u0026apos;I think there\u0026apos;s a potential that people see their computer [DLT trajectory output] and say, \u0026ldquo;I\u0026apos;m feeling suicidal and I\u0026apos;m not going to feel less suicidal, and so why bother. It\u0026apos;s never going to get better.\u0026rdquo; I think that\u0026apos;s definitely a risk. That potentially could be interpreted very negatively by clients. [\u0026hellip;] That could be, that could be pretty counter therapeutic\u0026apos; (P16 Clinical psychologist)\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eValue-add of predictive trajectories in mental health care\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eClinical utility in the \u0026lsquo;care as usual\u0026rsquo; milieu\u003c/em\u003e\u003c/strong\u003e. During the interviews, participants described various possible uses of the DLT in clinical practice. Primarily, the discussion centred on the DLT playing an added role in coordinating care and data-driven trajectory learning and prediction. Coordinating care with the DLT may involve using it as a collaborative tool between clinicians and clients, documenting clinical support roles, evidencing referrals, and supporting triage processes. On data-driven trajectory learning and prediction, most participants indicated that the DLT adequately indicates risk, timeliness of data, trends and change, and volatility, positioning it to support risk mitigation, decision-making, care planning, streamlining work, and a shift towards more targeted and personalised interventions. Some participants also suggested that the DLT could be used as a training or supervisory tool for junior clinicians or peer workers.\u003c/p\u003e\n\u003cp\u003eDiscussions often situated the DLT in regular mental health care settings, with participants reflecting on typical processes, failings, existing tools, and the fit of the DLT among these components. For the most part, participants noted that assessing suicidal ideation is very difficult, and they reported varying, standardised and non-standardised ways of assessing suicidal ideation, including measures (e.g., SIDAS), interviewing, and observations. Furthermore, each participant prioritised different interventions, depending on the population and mental health disorders they typically encounter in clinical practice. Despite a diversity of approaches, all participants emphasised the importance of collaborative care and sought to understand client goals and values in developing care decisions and plans, including those relating to suicide. When discussing where the DLT could fit within existing services, participants remarked on current issues within services to highlight the difficulty in introducing new digital technologies. These include high demand and low resources, long waitlists, a perceived focus on symptom reduction rather than resolution, slow adoption of digital technology, non-interoperable software across services, limited time available to integrate and use new tools, and bureaucratic processes and administrative tasks burdening clinicians. Most participants used digital tools in their practice. However, this was mostly limited to electronic health records for mandatory data collection procedures, and a few participants reported using AI tools to support report writing and notetaking, but none were for making predictions.\u003c/p\u003e\n\u003cp\u003eFraming the DLT within real care settings created a focused space in which participants suggested utilising the DLT as a tool to enhance and accommodate existing processes, diverse approaches, and technology use, namely centralising data collected via measures and supplementing non-standardised assessments with quantitative data. Furthermore, reflecting on current issues, participants highlighted the need to ensure that the DLT is non-disruptive and meaningful.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eAccuracy: Clinical judgement versus machine learning\u003c/em\u003e\u003c/strong\u003e. Accuracy was a salient topic. During user testing, many participants flagged concerns about the accuracy of the DLT trajectory output, which is chiefly displayed in a graph that visualises historical SIDAS scores and a trajectory range for the next SIDAS score indicated by a grey-coloured area on the graph (see Figure 5). Specifically, in instances when the trajectory range of the client\u0026rsquo;s predicted future score was wide (indicating uncertainty, informed by either volatility in clients\u0026rsquo; historical scores [see Figure 5a] or a sudden recent change after a long period of stability [see Figure 5b]), the accuracy and clinical utility was in question. Furthermore, in instances when the trajectory range was narrow (indicating high confidence, informed by a long period of consistent stability [see Figure 5c]), participants expressed that the DLT trajectory output is not evidently any more accurate than a clinician who could determine this by looking at the same data trend, thus limiting the added value of the DLT. In sum, it was not readily apparent to participants how to accurately interpret the trajectory ranges as indicating, in part, uncertainty (wide range) and confidence (narrow range).\u003c/p\u003e\n\u003cp\u003eViews on clinicians\u0026rsquo; perceived ability to change their approach to suicidal ideation assessment, poor adoption of digital interventions, and the superiority of clinical judgement compared to algorithmic prediction in mental health play a role in the perceived accuracy of the DLT. Specifically, as reported in the interview data, inspiring a change in approach to suicidal ideation assessment among clinicians is difficult due to the challenging nature of predicting suicide, reliance on favoured approaches for assessment, and the perceived burden and incongruence between digital interventions and trained human-driven approaches. Participants expressed that suicide prediction requires, in some part, the human touch and that they would more likely rely on a combination of their clinical judgement and the DLT compared to the DLT alone in decision-making, framing the DLT as an instrument to either validate or invalidate a clinician\u0026rsquo;s assessment. Yet, only a minority of participants expressed a willingness to use the DLT as a tool to enhance existing practice, and none were willing to completely delegate suicidal ideation assessment to AI, due to lingering uncertainties about accuracy and the perceived advantages of clinical judgement to draw on additional contextual data in collaboration with a client in situ.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIlluminating and magnifying the trajectory landscape\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMore transparency is needed\u003c/em\u003e\u003c/strong\u003e. The DLT allows users to browse each client\u0026rsquo;s responses to each SIDAS question to provide transparency. During user testing, most participants appreciated this feature as it allowed them to investigate individual assessment items closely to better understand clients\u0026rsquo; overall scores. Nevertheless, most participants sought greater transparency about the DLT model to improve their understanding of how it works, improve their interpretation of trends across clients\u0026rsquo; scores and trajectories, and, ultimately, advance trust. Specifically, they wanted to know what data is incorporated into the DLT model, including the data utilised during model training and operation, what constituents of SIDAS are incorporated in the predictive trajectories, and whether static factors (e.g., demographics) and population data are integrated. Furthermore, participants wanted more details about DLT trajectory outputs\u0026apos; calculations and confidence intervals. To that end, participants made various redesign recommendations to ensure that key information about the DLT model was plainly communicated on the UI. These included a confidence interval indicator for the DLT trajectory outputs; overtly flagging key driving predictors for each client\u0026rsquo;s trajectory; highlighting and elaborating on different psychiatric constructs underlying SIDAS that inform the DLT trajectory output; creating links between DLT trajectory outputs and the research evidence base and proof of the DLT model efficacy (e.g., providing external links to academic papers explaining the efficacy and importance of measures used, mental health dimensions considered, and the DLT model itself); and identifying the data used to generate outputs.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eIncorporating real-world considerations\u003c/em\u003e\u003c/strong\u003e. The various existing approaches to assessing suicidal ideation demand individual-level, mutable, and diverse contextual data. This includes, for instance, client demographics, social dimensions (e.g., family or romantic relationships), major life stressors (e.g., financial or legal challenges), and mental health history (e.g., past suicide plans or attempts). Some of this data also relies on being face-to-face with clients, such as observations of body hygiene and posture. Thus, participant feedback on the data used by the DLT to model suicidal ideation trajectories (i.e., SIDAS questionnaire responses) was modest and mixed. In short, whereas some noted that SIDAS is the gold standard for assessing suicidal ideation, others reported that the measure lacked nuance and missed crucial context. The same sentiment was expressed about C-SSRS (although it is not integrated into the DLT trajectory outputs).\u003c/p\u003e\n\u003cp\u003eDuring user testing, participants recommended incorporating real-world considerations that they currently rely on or would like to have access to for assessing suicidal ideation into the DLT in two key ways. First, feature key client demographics next to the DLT trajectory output to help clinicians recall key client details in situ. One participant highlighted that this feature might prove essential when assessing risk, especially when dealing with high-risk populations, noting that what appears to be a high-risk score for the general population might represent an average score among specific high-risk populations, such as Indigenous Australian brotherboys or sistergirls. Second, the variables considered by the DLT model should be expanded. Vital data to incorporate into suicidal ideation modelling included client demographics (e.g., age, sex, relationship status) and measures of impulsivity, substance use (alcohol, tobacco, cannabis), mental health (depression, anxiety, post-traumatic stress, psychosis-like and mania-like experiences), social factors (connectedness, occupational functioning, quality of life), and behavioural risks (self-harm, homicidal ideation, oppositionality, rule-breaking, sleep disturbances, distress, and violence). One participant recommended determining average SIDAS scores based on population data (e.g., sex, age, educational attainment level, and risk factors) and showing a comparison between a client\u0026rsquo;s score and their population group average. Another participant suggested integrating key events in a client\u0026rsquo;s life or mental health treatment (e.g., the start dates of the school year, new therapy intervention or medication, and a suicide attempt) to both inform the DLT trajectory output and provide greater context and perhaps help to interpret a client\u0026rsquo;s historical data and trajectory. To that end, participants discussed sourcing additional data to advance the contextualisation of the DLT model from client notes, health records, police records, coronial reports, and client session transcripts, leveraging data mining techniques and large language models to parse data.\u003c/p\u003e\n\u003cp\u003eBeyond suicidal ideation, participants also recommended modelling and generating predictive trajectories for numerous mental health dimensions. These included\u003cem\u003e\u0026nbsp;\u0026lsquo;more classic, major aspects of symptomatology\u0026rsquo; (P10 Clinical neuropsychologist)\u003c/em\u003e such as psychological distress, alcohol, tobacco, and cannabis use, sleep, anxiety, psychosis-like experiences, and depression. However, emphasising the importance of context, participants noted that the utility of these mental health dimensions trajectories (including suicidal ideation) depends on individual client presentation and a service\u0026rsquo;s typical population.\u003c/p\u003e\n\u003cp\u003eMoving towards implementation\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eMachine learning amplifies ethical challenges\u003c/em\u003e\u003c/strong\u003e. Two key ethical considerations about machine learning in mental health and the DLT were discussed: privacy and accountability. Participants were very concerned about privacy relating to DMHTs. Many noted the importance of DMHT platforms ensuring that sensitive personal information was protected, remained confidential, and managed according to national digital health standards governing privacy, especially given the sensitive nature of data regarding suicide. They raised questions about data ownership, security, governance, and storage in digital mental health and conveyed that the integration of AI obfuscates these issues further. When discussing the use of AI scribes in mental health, i.e., speech recognition technology that automatically converts voice-to-text for clinical documentation, and the possibility of integrating the same functionality with the DLT, participants flagged that the data should be deidentified before being utilised and that appropriate procedures to obtain client consent to use their data is necessary. Some participants said explicit adherence to Australian privacy and cybersecurity laws and standards is essential.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAccountability was a significant concern. Many reflected on existing accountability for acting upon assessments that indicate a high risk of suicide and wondered if the DLT added additional complexities to practice. For instance, one participant noted that if the DLT trajectory output indicates a negative future trend and the clinician does not expect the same trend, the accountability for action based on conflicting assessments is unclear. Furthermore, the accountability for any subsequent negative consequences to the client is confused. Multiple participants made explicit requests for guidelines on using and actioning the DLT.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eClear guidance on use is necessary\u003c/em\u003e\u003c/strong\u003e. During user testing, most participants found it challenging to discern the DLT trajectory output (see Figure 5) intuitively. This presented an initial barrier to clinical use, and most participants needed comprehensive instructions to interpret the output. Participants reported that the output was hard to understand for several reasons. These were (1) data plotted, and the trajectory itself was inadequately labelled; (2) the y-axis directions were confusing and did not meet expectations (i.e., whereas the top of the y-axis represents \u0026lsquo;healthy\u0026rsquo; scores, participants expected the opposite); and (3) a tooltip (i.e., a text box that appears when a user hovers over a specific UI element to describe the element) integrated into the graph was considered frustrating as it sometimes covered elements of the graph and did not provide helpful, applicable, and tailored information.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMoreso than resolving the aforementioned UI problems that created confusion, the participants indicated that the DLT should be implemented with adequate guidelines and associated onboarding training. This includes formalised comprehensive guidelines, training procedures, and helpful tips featured on the UI to provide simple, ad hoc guidance. Specifically, guidelines and training are needed to advise how to interpret the outputs and subsequently develop an appropriate care plan based on the outputs. Participants indicated that guidance is crucial to ensure correct use among junior clinicians who may experience more difficulties interpreting and generating relevant treatment options than a more experienced clinician.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eEngaging the right users\u003c/em\u003e\u003c/strong\u003e. Identifying target user groups for the DLT was a salient, divisive point of discussion. Participants agreed that mental health professionals were the appropriate primary target end users. Those working in private psychology practices or smaller service centres were more interested in using the DLT than those working in public and large service centres (e.g., headspace). Participants representing the latter group reported disinterest due to the existing burden of large clinical caseloads and administrative obligations in public and large psychology service centres.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThere were strong attitudes about whether young people (i.e., clients in youth mental health services) should have access to the DLT. Some participants conveyed that young people should have access to information about their care, especially to track and observe trends and create opportunities to discuss data with clinicians. Thus, they should have access to the DLT. Other participants expressed the opposite attitude, noting some risks of young clients seeing the DLT trajectory outputs, including misinterpreting the outputs, creating a sense of hopelessness, and generating negative expectation-induced behaviours (i.e., if a client observes that the DLT indicates a likely negative trend, then they may be disempowered or withdraw from treatment participation). As pointed out by the participants, depending on the client and their current mental health experiences and trajectory, accessing the DLT risks empowerment or disempowerment.\u003c/p\u003e\n\u003cp\u003eDuring the interviews, to further explore the possibility of young clients accessing the DLT, a feasible design feature was presented to gather participant feedback: clinician direct control over which aspects of the DLT each client can access. Most participants rejected this potential feature, suggesting it negatively impacts young people\u0026rsquo;s autonomy and may create a sense of infantilisation. Ultimately, it was determined that for the sake of safety and to avoid these negative impacts, it is best to avoid clients accessing the DLT altogether. As the next best alternative that aligns with how clinicians typically share data with clients, participants favoured an approach that enabled clinicians and clients to collaboratively access the DLT during a client\u0026rsquo;s session at the clinician\u0026rsquo;s discretion. Participants also noted that it is difficult to encourage young people to engage with DMHTs for self-reporting purposes, and thus, enabling clients to access the DLT may be ineffective and require substantial integration into service delivery.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis qualitative study employed semi-structured, in-depth interviews, usability testing, and reflexive thematic analysis to explore youth mental health professionals\u0026rsquo; attitudes towards AI and determine the design and content requirements for the DLT. Most participants were open to utilising AI and the dynamic learning of illness trajectories across mental health dimensions in clinical practice as a tool to enhance and support their decision-making. Furthermore, most found that the DLT indicates risk, volatility, trends and change, and timeliness of data. Yet, they expressed scepticism about predictive trajectories of clients\u0026rsquo; suicidal ideation, compared to other mental health dimensions, citing fears of perceived low accuracy and AI tools replacing clinical judgement. Participants emphasised that suicidal ideation assessment largely remains a human-centric task, leveraging existing quantitative measures, interviewing, observations, understanding client goals and values, and collaborative care. They viewed AI and the DLT as tools to enhance, but not replace, care coordination and data-driven trajectory learning and prediction in Australian youth mental health, emphasising that implementation should support existing processes and technology use without adding to existing issues, including administrative burden and slow technology adoption.\u003c/p\u003e\n\u003cp\u003eFollowing the co-design methodology, participants served a vital role in informing the redesign of the DLT model and its implementation. Participants called for greater model transparency to inspire trust among users. Users should be able to examine what data is used to develop DLT trajectory outputs and train the model, the confidence intervals of the outputs, key driving predictors of each trajectory, and supporting evidence validating the model. Moreover, participants wanted the DLT model to be expanded to incorporate individual-level, mutable, and diverse contextual data that they find useful in assessing suicidal ideation, including mental health history, client demographics, and major life stressors. Beyond suicidal ideation, participants supported modelling trajectories for other mental health dimensions, such as psychological distress, anxiety, and depression. Although suicidal ideation was the focus of this study, the DLT could be adapted to learn trajectories of other mental health dimensions with appropriate evidence-based measurements. Table 3 summarises the key, in-scope recommendations for the DLT co-designed during interviews, including new and redesigned features.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Additional requested features for the DLT.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSource of frustration\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e(Re)design recommendations\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelated theme (subtheme)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eLack of content personalisation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cul\u003e\n \u003cli\u003eContent associated with the static risk indicator should include more personal insights into a client\u0026rsquo;s risk factors, with reference to significant results from questionnaires\u003c/li\u003e\n \u003cli\u003eAny instance referring to the client should use their proper name\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eValue-add of predictive trajectories in mental health care (Clinical utility in the \u0026lsquo;care as usual\u0026rsquo; milieu)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNo link between trajectory and recommended treatment options/care plan\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eGenerate recommended potential treatment options based on predictions for clinician users to consider and action if deemed appropriate. For example, if the client\u0026rsquo;s trajectory range is beyond an uncertainty threshold, inform the clinician user and recommend a higher frequency of client follow-up appointments and review of their treatment plan\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNo flags of significant risk factors\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eInstances that could be flagged and conveyed to the mental health professional include:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eSignificant areas of risk related to client questionnaire responses\u003c/li\u003e\n \u003cli\u003eThe probability of reaching a high-risk category, as indicated by an integrated high-ideation probability (IHIP) score greater than 20 (of 50, i.e., 40+%) [40]\u003c/li\u003e\n \u003cli\u003eA new data point that is higher or lower than the last expected trajectory range\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNo discrete indicator of client questionnaire response volatility to differentiate volatility from risk. At present, the DLT trajectory output indicates three factors: volatility, uncertainty, and timeliness of the data.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eEstablish an appropriate numerical boundary that flags instances of volatility. Add a volatility indicator that does one of the following when volatility is flagged:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eDeactivate and restrict access to the DLT trajectory output (i.e., the graph) and display the message: \u0026lsquo;Volatility flagged in clients\u0026rsquo; questionnaire responses, speak to the client\u0026rsquo;. Once volatility is no longer detected, return access rights\u003c/li\u003e\n \u003cli\u003eAdd a discrete volatility indicator to distinguish the three factors, consigning the DLT trajectory output to indicate only the uncertainty and timeliness of the data\u003c/li\u003e\n \u003cli\u003eTo not inadvertently convey uncertainty with the DLT trajectory output at baseline and risk confusing people that it is conveying volatility at baseline, do not display trajectories at baseline\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"5\" valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eValue-add of predictive trajectories in mental health care (Accuracy: Clinical judgement versus machine learning)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eTimeliness of the data is not explicitly communicated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eAdd a data timeliness indicator. For example, divide the DLT trajectory output into 1, 2, 3, and 4-week timepoints segments, indicating that as the time increases and the output widens, the timeliness of the data degrades\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eGraph is hard to read (i.e., data points overlap)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eAdd a zoom function on the graph\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eScepticism about the DLT replacing clinical judgement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eFeature a reminder on the UI: \u0026lsquo;This tool is not intended to replace clinical judgement. It should be used to support clinical judgement\u0026rsquo;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eDLT trajectory output does not sufficiently convey uncertainty and confidence; some DLT trajectory output ranges are too wide or too narrow, affecting perceived accuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cul\u003e\n \u003cli\u003eA wide output range primarily indicates uncertainty based on either volatility in clients\u0026rsquo; historical scores or a sudden recent change after a long period of stability. Revise to determine an uncertainty threshold, label the output with this information for clarity, and advise a potential treatment option based on predictions for clinician users to consider and action if deemed appropriate (in this instance, recommend more regular client follow-up appointments and treatment plan review)\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eA narrow output range primarily indicates confidence based on a long period of consistent stability, which participants thought could be interpreted intuitively from the data points. Revise to add multiple confidence levels to the output to evidence the trajectory and possibilities for clarity and advise a subsequent potential treatment option (in this instance, recommend normal frequency follow-up with the client and typical review of their treatment plan)\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNeed more details on the DLT model and data incorporated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cul\u003e\n \u003cli\u003eFlag key driving predictors for DLT trajectory outputs\u003c/li\u003e\n \u003cli\u003eProvide external links to the research evidence base (e.g., academic papers explaining the efficacy and importance of each measure used, mental health dimensions considered, and the model)\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eIlluminating and magnifying the trajectory landscape (More transparency is needed)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNeed more details on the DLT trajectory output calculations and confidence intervals\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cul\u003e\n \u003cli\u003eHighlight and explicate psychiatric constructs underlying SIDAS that inform outputs\u003c/li\u003e\n \u003cli\u003eAdd a confidence interval indicator for the outputs. Include a feature to enable users to switch between different confidence intervals (e.g., 68% and 95%)\u003c/li\u003e\n \u003cli\u003eAdd a tooltip connected to data points to label significant links between the client\u0026rsquo;s historical data and trajectory and individual items answered on the SIDAS questionnaire\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNeed to show contextual data typically used for assessing suicidal ideation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eFeature key client demographics on the UI to help clinicians recall key details about clients in situ and interpret the output (e.g., population group, age, and relationship status). Consider collating this data by specific decisions to be made by the clinician (e.g., data relating to developing a treatment plan, introducing an intervention, or choosing an appropriate assessment scale or measure)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"4\" valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eIlluminating and magnifying the trajectory landscape (Incorporating real-world considerations)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNeed to expand the variables incorporated in the DLT model for assessing suicidal ideation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eConsider incorporating the following data from in-scope measures already captured in Innowell that may be incorporated into the DLT model:\u0026nbsp;\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eFor \u003cem\u003ebaseline\u003c/em\u003e predictive trajectories: Client demographics (e.g., age, sex, and relationship status), Innowell onboarding questionnaires (e.g., Quick Inventory of Depressive Symptomatology), and relationships between parameters\u003c/li\u003e\n \u003cli\u003eFor \u003cem\u003econtinuous-time\u003c/em\u003e predictive trajectories: social connectedness and social and occupational functioning. Incorporating this data may prove useful for data imputation to resolve instances when clients did not self-report suicidal ideation data but provided data on other mental health dimensions to support prediction.\u003c/li\u003e\n \u003c/ul\u003e\n \u003cp\u003eIn the future, consider incorporating data not yet captured in Innowell, including:\u003c/p\u003e\n \u003cul\u003e\n \u003cli\u003eRelevant life events in a client\u0026rsquo;s life or mental health treatment (e.g., the start dates of the school year, new therapy intervention or medication, and a suicide attempt)\u003c/li\u003e\n \u003cli\u003eMeasures of impulsivity, quality of life, substance use (alcohol, tobacco, cannabis), behavioural risks (self-harm, homicidal ideation, oppositionality, rule-breaking, sleep disturbances, distress, and violence), and other mental health dimensions (depression, anxiety, post-traumatic stress, psychosis-like and mania-like experiences)\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eNeed comparison scores against clients\u0026rsquo; population group\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eTo contextualise individual-level trajectories in a population to benefit interpretation, add a population group average line on the graph evidencing average SIDAS scores based on population data relevant to the client (e.g., sex, age, educational attainment level, and risk factors).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eExpand data sources\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eSource additional contextual data using data mining techniques and large language models to search, for instance, client notes, health records, police records, coronial reports, and client session transcripts. Adhere to Australian standards and laws on health data usage, retention, consent, and privacy\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eConcerns about data privacy and protection\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eOvertly communicate adherence to Australian standards and laws concerning data privacy, protection, and cybersecurity on the UI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eMoving towards implementation (Machine learning amplifies ethical challenges)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eData points and the DLT trajectory output were inadequately labelled\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cul\u003e\n \u003cli\u003eAdd relevant labels to data point labels (e.g., date data gathered, score, and whether the score has increased or decreased since the last data point)\u003c/li\u003e\n \u003cli\u003eRevise DLT trajectory outputs with the abovementioned labels\u003c/li\u003e\n \u003c/ul\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eMoving towards implementation (Clear guidance on use is necessary)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eGraph y-axis is confusing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eReframe and revise y-axis directions so that the top represents \u0026lsquo;unhealthy\u0026rsquo; or undesirable scores (i.e., suicidal ideation is increasing), and the bottom represents \u0026lsquo;healthy\u0026rsquo; or desirable scores (but do not label the bottom to avoid additional confusion)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eThe existing tooltip connected to data points is frustrating and blocks key information\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eRemove and revise as outlined above\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 179px;\"\u003e\n \u003cp\u003eThere is risk associated with showing outputs to clients\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 567px;\"\u003e\n \u003cp\u003eDo not develop a client view of the DLT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 117px;\"\u003e\n \u003cp\u003eMoving towards implementation (Engaging the right users)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003ePrior to implementing the DLT, participants recommended resolving ethical issues, providing clear guidance on how to use the DLT, and identifying and accommodating the appropriate target user groups. Privacy and accountability were key ethical considerations intensified by the application of machine learning in the study context. Personal data privacy and data security, storage, governance, and ownership and accountability for DLT-supported decisions and subsequent client outcomes need to be addressed and made explicit to the user in the DLT UI. Guidance, training, and dynamic tips featured on the UI to support the interpretation of the DLT trajectory outputs are crucial to avoid misinterpretation and associated risks. Participants agreed that the primary users of the DLT should be mental health clinicians. In response to mixed attitudes on whether young people should have access to the DLT and to alleviate risks associated with clients misinterpreting the DLT, client access to the DLT should be at the discretion of their clinician, and a tailored client portal of the DLT should be avoided. As the next steps for this work, co-design of the DLT with key stakeholders in youth mental health services is crucial to ensure value-add, increase system transparency and data incorporated into algorithmic decision-making, and progress appropriate implementation with the right safeguards in place, with the right users in the right environments, and with the right know-how to utilise the DLT in a way that minimises risks and maximises benefits.\u003c/p\u003e\n\u003cp\u003eThese findings have broader implications for the ethics, design, development, implementation, governance, and evaluation of machine learning of predictive trajectories and AI in youth mental health. While evidencing prediction accuracy to health practitioners is an enduring problem generally [56], digital health systems face additional concerns about trust and reliability as health data and decision-making are increasingly delegated to digital solutions [57]. Research suggests several methods to alleviate the problem, including conveying system accuracy information with simple, non-technical language [56] and visual representations [57]. Although participants understood the DLT\u0026rsquo;s language and were presented with a line graph visualisation, difficulties among participants interpreting the graph indicate a need for visualisation redesign. Schneider, Wayrauther (57) suggest that visualising deviations or algorithmic errors between historical health data and predictive trajectories improves understanding by providing reference to real past data, and so redesigning the DLT in this way may be appropriate. While health research advocates for communicating system accuracy by presenting natural frequencies rather than probability [58] (e.g., 1 out of 10 people are positive versus the probability that one person is positive is 10%), human-computer interaction research shows that such analogy-based explanations like natural frequencies do not increase perceived accuracy among expert users (e.g., mental health clinicians) unless adjusted to their needs [59]. Greater stakeholder needs assessment and subsequent personalisation of user experience are necessary to ensure that the DLT accuracy is communicated to the target users in the way they need it to be communicated. To that end, further co-design is needed. Digital health literature highlights the importance of incorporating highly personalised data and communicating to the user what personal data is integrated into algorithmic decision-making to dispel fears about perceived system inaccuracies [60]. This affirms the study finding that incorporating real-world considerations into the DLT and transparently and simplistically communicating this process and data to the users is essential in future iterations of the DLT and other state-of-the-art AI in youth mental health broadly.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe findings presented here support existing research that emphasises the need to frame and design AI in youth mental health as a tool to enhance but not replace clinical judgement to ensure value-add going forward [61, 62]. This framing is particularly important to alleviate concerns about potential conflicts between clinical judgement and algorithmic prediction. Furthermore, this framing is key to curbing the evident dichotomy between the realistic amount of data that a mental health professional can reasonably consider in their day-to-day decision-making and their high expectations on the volume of data and data sources to be incorporated into machine learning of predictive trajectories. In this way, data-driven trajectory learning and prediction (such as the DLT\u0026rsquo;s predictive trajectories) are positioned as another information source to enhance a clinician\u0026rsquo;s decision, rather than a replacement and improve personalised care services. For the DLT, this could include generating recommendations for follow-up appointment frequency (e.g., advising whether to aim to see a client next week or next month) or indicating treatment effectiveness based on a client\u0026rsquo;s volatility or stability (see Table 3 for more details), addressing lasting unanswered questions in mental health regarding determining appointment frequency [63] and treatment effectiveness [64]. More broadly, emphasising system simplicity by design to effectively convey system accuracy, clearly highlighting key information incorporated into data-driven decisions and relevant information to inform a clinician\u0026rsquo;s decisions, and ensuring system ubiquity to work alongside existing systems, workflows, and different therapy approaches will play a key role in establishing the dynamic learning of individual-level trajectories, and AI more broadly, in youth mental health as enhancing rather than replacing human decision-making.\u003c/p\u003e\n\u003cp\u003eCalls for greater transparency of AI in mental health are well-documented in the literature [65]. Yet, limited studies report practical design recommendations relevant to AI systems intended for dynamic trajectory learning and prediction to best communicate explainability to users and, ultimately, improve user experience, trust, comprehension of underlying machine learning models, and interpretation of system outputs. To best address this gap, this study found that AI generally, and data-driven trajectory learning and prediction tools specifically, operating in mental health should directly communicate and evidence, at least, the following:\u003c/p\u003e\n\u003col\u003e\n \u003cli\u003eVisually indicate confidence intervals for the trajectory outputs (e.g., explicitly label each output displayed with the appropriate confidence interval).\u003c/li\u003e\n \u003cli\u003eOvertly flag key driving predictors for each output (e.g., list driving predictors alongside associated outputs).\u003c/li\u003e\n \u003cli\u003eHighlight and elucidate different psychiatric constructs underlying clinical assessment measured that are incorporated into outputs (e.g., list constructs alongside associated outputs).\u003c/li\u003e\n \u003cli\u003eCreate and show links between outputs and the research knowledge base and proof of the underlying machine learning model efficacy (e.g., feature external links to academic papers evidencing the efficacy and importance of measures used, mental health dimensions considered, and the model used).\u003c/li\u003e\n \u003cli\u003eIdentify all data used to generate outputs (e.g., list appropriate data alongside associated outputs, ranked by their corresponding weight).\u003c/li\u003e\n\u003c/ol\u003e\n\u003cp\u003eOther studies about consumer-facing AI in health generally have identified and warned of \u0026ldquo;inappropriate explainability,\u0026rdquo; i.e., excessive or improper levels of transparency of explanations [66]. Existing research remarks that inappropriate explainability may lead to information overload among users, negative user experiences, user harm caused by ignoring system prompts, accidental disclosure and privacy violation of sensitive consumer data, and counteracting the interests of AI providers [66]. Findings from the present study identify additional dimensions of potential user harm in the youth mental health space. Based on participant feedback, some interactions with AI in mental health and their outputs should not be made available to young clients in mental health services despite potential opportunities for empowerment due to risks associated with misinterpreting outputs, developing negative expectation-induced behaviours, and creating a sense of hopelessness, all of which threaten disempowerment, care service disengagement, low digital tool uptake among clients. This empowerment-disempowerment dichotomy must inevitably be dealt with at the design, development, implementation, and governance stages with an approach that integrates a highly personalised approach to care and AI in mental health design to best tailor outputs to different user groups based on user capacity, needs, values, preferences, expectations, and role in the caring relation [67].\u003c/p\u003e\n\u003cp\u003eSome limitations of this research should be noted. First, the majority of participants were located in Australian metropolitan areas, and thus, the findings are not generalisable to mental health care delivery practices and services in other countries or regional and rural areas. Second, most participants were psychologists, and therefore, the study lacks a salient representation of other mental health professionals that should be accounted for in the research, design, implementation, use, and governance of DMHTs.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eIdentifying risk and predicting future trajectories in youth mental health using conventional measures produces only modest accuracy and replicability, failing to confront the worsening global youth mental health crisis. Although the use of AI applications, such as machine learning, demonstrates a promising alternative, there are key underexplored attitudes and design requirements among mental health professionals in this space. This work addresses this gap by developing three themes, \u003cem\u003eValue-add of predictive trajectories in mental health care\u003c/em\u003e, \u003cem\u003eIlluminating and magnifying the trajectory landscape\u003c/em\u003e, and \u003cem\u003eMoving towards implementation\u003c/em\u003e, reporting a rich understanding of mental health professionals\u0026rsquo; attitudes towards AI in youth mental health generally and their design and content requirements for a novel prediction tool, the DLT, specifically. Professionals are willing to use AI and predictive trajectories, but only to enhance their decision-making and clinical judgement. Moreover, they require considerable machine learning model transparency, evident privacy and accountability standards adherence, use guidance, and contextualisation. To advance the design, development, implementation, ethics, governance, and evaluation of AI and prediction tools in youth mental health more broadly, this study also provides stakeholder-informed practical design recommendations to best communicate system explainability to users and thus benefit trust, comprehension, and user experience.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cdiv class=\"DefinitionList\"\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eAI\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eArtificial intelligence\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eC-SSRS\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eColumbia-Suicide Severity Rating Scale\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eDLT\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eDynamic Learning Tool\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eDMHTs\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eDigital mental health tools\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eSIDAS\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eSuicidal ideation attribute scale\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv class=\"DefinitionListEntry\"\u003e\u003cdiv class=\"Term\"\u003eUI\u003c/div\u003e\u003cdiv class=\"Description\"\u003e\u003cp\u003eUser interface\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics approval and consent to participate\u003c/p\u003e\n\u003cp\u003eThis research received ethical approval from the University of Sydney\u0026rsquo;s Human Research and Ethics Committee (Project number: 2021/HE000680). Clinical trial number: not applicable. Participants provided informed consent prior to completing the study. Participant informed consent, participant privacy, confidentiality, and deidentification, and secure data transmission, storage, and handling procedures were carried out in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003eConsent for publication\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials\u003c/p\u003e\n\u003cp\u003eRaw data are not available to the public due to data protection and privacy reasons. Anonymised metadata (e.g., coding glossary and code frequency) that support the findings are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eIBH is the Co-Director, Health and Policy at the Brain and Mind Centre (BMC) University of Sydney. The BMC operates an early-intervention youth services at Camperdown under contract to headspace. He is the Chief Scientific Advisor to, and a 3.2% equity shareholder in, InnoWell Pty Ltd which aims to transform mental health services through the use of innovative technologies. AP, MKC, HML, AT, MV, and FI have nothing to disclose.\u003c/p\u003e\n\u003cp\u003eFunding\u003c/p\u003e\n\u003cp\u003eIBH is supported by a NHMRC L3 Investigator Grant (GNT2016346). FI is supported by an NHMRC EL1 Investigator Grant (GNT2018157). MV was supported by philanthropic funding from The Johnston Fellowship and from other donor(s) who are families affected by mental illness who wish to remain anonymous. This work was supported by the Medical Research Future Fund Grant (MRFCRI000279, Using AI to personalise treatment decisions in youth mental health services). The funder played no role in study design, data collection, analysis and interpretation of data, or the writing of this manuscript.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions: AP:\u003c/strong\u003e Conceptualization, Methodology, Formal analysis, Investigation, Data Curation, Writing\u0026ndash;Original Draft, Writing\u0026ndash;Review \u0026amp; Editing, Visualization. \u003cstrong\u003eIBH:\u003c/strong\u003e Resources, Writing\u0026ndash;Review \u0026amp; Editing, Supervision, Project administration, Funding acquisition. \u003cstrong\u003eMKC:\u003c/strong\u003e Formal analysis, Writing\u0026ndash;Original Draft, Writing\u0026ndash;Review \u0026amp; Editing. \u003cstrong\u003eHML:\u003c/strong\u003e Formal analysis, Writing\u0026ndash;Review \u0026amp; Editing. \u003cstrong\u003eAT:\u003c/strong\u003e Formal analysis, Writing\u0026ndash;Review \u0026amp; Editing. \u003cstrong\u003eMV:\u003c/strong\u003e Software, Writing\u0026ndash;Review \u0026amp; Editing. \u003cstrong\u003eFI:\u003c/strong\u003e Conceptualization, Methodology, Formal analysis, Resources, Writing\u0026ndash;Review \u0026amp; Editing.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eMcGorry PD, Mei C, Dalal N, Alvarez-Jimenez M, Blakemore S-J, Browne V, et al. The Lancet Psychiatry Commission on youth mental health. Lancet Psychiatry. 2024;11(9):731-74. doi: 10.1016/S2215-0366(24)00163-9.\u003c/li\u003e\n \u003cli\u003eBertuccio P, Amerio A, Grande E, La Vecchia C, Costanza A, Aguglia A, et al. Global trends in youth suicide from 1990 to 2020: an analysis of data from the WHO mortality database. eClinicalMedicine. 2024;70:1-12. doi: 10.1016/j.eclinm.2024.102506.\u003c/li\u003e\n \u003cli\u003eLovero KL, Dos Santos PF, Come AX, Wainberg ML, Oquendo MA. Suicide in Global Mental Health. Current Psychiatry Reports. 2023;25(6):255-62. doi: 10.1007/s11920-023-01423-x.\u003c/li\u003e\n \u003cli\u003eUddin R, Burton NW, Maple M, Khan SR, Khan A. Suicidal ideation, suicide planning, and suicide attempts among adolescents in 59 low-income and middle-income countries: a population-based study. The Lancet Child \u0026amp; Adolescent Health. 2019;3(4):223-33. doi: 10.1016/S2352-4642(18)30403-6.\u003c/li\u003e\n \u003cli\u003eAhad AA, Sanchez-Gonzalez M, Junquera P. Understanding and Addressing Mental Health Stigma Across Cultures for Improving Psychiatric Care: A Narrative Review. Cureus. 2023;15(5):1-8. doi: 10.7759/cureus.39549.\u003c/li\u003e\n \u003cli\u003ePizzol D, Trott M, Butler L, Barnett Y, Ford T, Neufeld SAS, et al. Relationship between severe mental illness and physical multimorbidity: a meta-analysis and call for action. BMJ Mental Health. 2023;26:1-5. doi: 10.1136/bmjment-2023-300870.\u003c/li\u003e\n \u003cli\u003ePatel V, Saxena S, Lund C, Kohrt B, Kieling C, Sunkel C, et al. Transforming mental health systems globally: principles and policy recommendations. The Lancet. 2023;402(10402):656-66. doi: 10.1016/S0140-6736(23)00918-2.\u003c/li\u003e\n \u003cli\u003eIorfino F, Carpenter JS, Cross SPM, Crouse J, Davenport TA, Hermens DF, et al. Social and occupational outcomes for young people who attend early intervention mental health services: a longitudinal study. Medical Journal of Australia. 2022;216(2):87-93. doi: 10.5694/mja2.51308.\u003c/li\u003e\n \u003cli\u003eHrabok M, Delorme A, Agyapong VIO. Threats to Mental Health and Well-Being Associated with Climate Change. Journal of Anxiety Disorders. 2020;76:1-5. doi: 10.1016/j.janxdis.2020.102295.\u003c/li\u003e\n \u003cli\u003eMcGorry P, Gunasiri H, Mei C, Rice S, Gao CX. The youth mental health crisis: analysis and solutions. Frontiers in Psychiatry. 2025;15:1-20. doi: 10.3389/fpsyt.2024.1517533.\u003c/li\u003e\n \u003cli\u003eNelson B, McGorry PD, Wichers M, Wigman JTW, Hartmann JA. Moving From Static to Dynamic Models of the Onset of Mental Disorder: A Review. JAMA Psychiatry. 2017;74(5):528-34. doi: 10.1001/jamapsychiatry.2017.0001.\u003c/li\u003e\n \u003cli\u003eHickie IB, Scott EM, Cross SP, Iorfino F, Davenport TA, Guastella AJ, et al. Right care, first time: A highly personalised and measurement-based care model to manage youth mental health. Medical Journal of Australia. 2019;211(S9):S3-S46. doi: 10.5694/mja2.50383.\u003c/li\u003e\n \u003cli\u003eHornstein S, Forman-Hoffman V, Nazander A, Ranta K, Hilbert K. Predicting therapy outcome in a digital mental health intervention for depression and anxiety: A machine learning approach. DIGITAL HEALTH. 2021;7:1-11. doi: 10.1177/20552076211060659.\u003c/li\u003e\n \u003cli\u003eWies B, Landers C, Ienca M. Digital Mental Health for Young People: A Scoping Review of Ethical Promises and Challenges. Front Digit Health. 2021;3:1-11. doi: 10.3389/fdgth.2021.697072.\u003c/li\u003e\n \u003cli\u003eTorous J, Bucci S, Bell IH, Kessing LV, Faurholt-Jepsen M, Whelan P, et al. The growing field of digital psychiatry: current evidence and the future of apps, social media, chatbots, and virtual reality. World Psychiatry. 2021;20(3):318-35. doi: 10.1002/wps.20883.\u003c/li\u003e\n \u003cli\u003eUhlhaas P, Torous J. Digital tools for youth mental health. npj Digital Medicine. 2019;2:1-3. doi: 10.1038/s41746-019-0181-2.\u003c/li\u003e\n \u003cli\u003eMcGorry PD, Mei C, Chanen A, Hodges C, Alvarez-Jimenez M, Killackey E. Designing and scaling up integrated youth mental health care. World Psychiatry. 2022;21(1):61-76. doi: 10.1002/wps.20938.\u003c/li\u003e\n \u003cli\u003eBarrigon ML, Romero-Medrano L, Moreno-Mu\u0026ntilde;oz P, Porras-Segovia A, Lopez-Castroman J, Courtet P, et al. One-Week Suicide Risk Prediction Using Real-Time Smartphone Monitoring: Prospective Cohort Study. J Med Internet Res. 2023;25:1-11. doi: 10.2196/43719.\u003c/li\u003e\n \u003cli\u003eHughes JL, Horowitz LM, Ackerman JP, Adrian MC, Campo JV, Bridge JA. Suicide in young people: screening, risk assessment, and intervention. BMJ. 2023;381:1-16. doi: 10.1136/bmj-2022-070630.\u003c/li\u003e\n \u003cli\u003e\u0026Oslash;sterg\u0026aring;rd Ole K, Gr\u0026oslash;nneb\u0026aelig;k L, Nilsson Kristine K. Do Therapists Know When Their Clients Deteriorate? An Investigation of Therapists\u0026apos; Ability to Estimate and Predict Client Change During and After Psychotherapy. Clinical Psychology \u0026amp; Psychotherapy. 2024;31(6):1-10. doi: 10.1002/cpp.70015.\u003c/li\u003e\n \u003cli\u003eRoza TH, Salgado TA, Machado CS, Watts D, Bebber J, Freitas T, et al. Prediction of Suicide Risk Using Machine Learning and Big Data. In: Passos IC, Rabelo-da-Ponte FD, Kapczinski F, editors. Digital Mental Health: A Practitioner\u0026apos;s Guide. Cham: Springer; 2023. p. 173-88.\u003c/li\u003e\n \u003cli\u003eFonseka TM, Bhat V, Kennedy SH. The utility of artificial intelligence in suicide risk prediction and the management of suicidal behaviors. Australian \u0026amp; New Zealand Journal of Psychiatry. 2019;53(10):954-64. doi: 10.1177/0004867419864428.\u003c/li\u003e\n \u003cli\u003eDaniel DH, Erwin L. AI enabled suicide prediction tools: a qualitative narrative review. BMJ Health Care Inform. 2020;27(3):1-10. doi: 10.1136/bmjhci-2020-100175.\u003c/li\u003e\n \u003cli\u003eHaines-Delmont A, Chahal G, Bruen AJ, Wall A, Khan CT, Sadashiv R, et al. Testing Suicide Risk Prediction Algorithms Using Phone Measurements With Patients in Acute Mental Health Settings: Feasibility Study. JMIR Mhealth Uhealth. 2020;8(6):1-15. doi: 10.2196/15901.\u003c/li\u003e\n \u003cli\u003eMcHugh CM, Ho N, Iorfino F, Crouse JJ, Nichles A, Zmicerevska N, et al. Predictive modelling of deliberate self-harm and suicide attempts in young people accessing primary care: a machine learning analysis of a longitudinal study. Social Psychiatry and Psychiatric Epidemiology. 2023;58(6):893-905. doi: 10.1007/s00127-022-02415-7.\u003c/li\u003e\n \u003cli\u003eIorfino F, Ho N, Carpenter JS, Cross SP, Davenport TA, Hermens DF, et al. Predicting self-harm within six months after initial presentation to youth mental health services: A machine learning study. PloS one. 2021;15(12):1-16. doi: 10.1371/journal.pone.0243467.\u003c/li\u003e\n \u003cli\u003eFranklin JC, Ribeiro JD, Fox KR, Bentley KH, Kleiman EM, Huang X, et al. Risk factors for suicidal thoughts and behaviors: A meta-analysis of 50 years of research. Psychol Bull. 2017;143(2):187-232. doi: 10.1037/bul0000084.\u003c/li\u003e\n \u003cli\u003eWang SB, Van Genugten RDI, Yacoby Y, Pan W, Bentley KH, Bird SA, et al. Building personalized machine learning models using real-time monitoring data to predict idiographic suicidal thoughts. Nature Mental Health. 2024;2(11):1382-91. doi: 10.1038/s44220-024-00335-w.\u003c/li\u003e\n \u003cli\u003eLutz W, Schwartz B, Delgadillo J. Measurement-Based and Data-Informed Psychological Therapy. Annu Rev Clin Psychol. 2022;18:71-98. doi: 10.1146/annurev-clinpsy-071720-014821.\u003c/li\u003e\n \u003cli\u003eSinclair C, Holloway K, Riley G, Auret K. Online Mental Health Resources in Rural Australia: Clinician Perceptions of Acceptability. J Med Internet Res. 2013;15(9):1-10. doi: 10.2196/jmir.2772.\u003c/li\u003e\n \u003cli\u003eGanapathy A, Clough BA, Casey LM. Organizational and Policy Barriers to the Use of Digital Mental Health by Mental Health Professionals. Telemedicine and e-Health. 2021;27(12):1332-43. doi: 10.1089/tmj.2020.0455.\u003c/li\u003e\n \u003cli\u003eLamela D, Cabral J, Coelho S, Jongenelen I. Personal stigma, determinants of intention to use technology, and acceptance of internet-based psychological interventions for depression. International Journal of Medical Informatics. 2020;136:1-6. doi: 10.1016/j.ijmedinf.2020.104076.\u003c/li\u003e\n \u003cli\u003eNogueira-Leite D, Diniz JM, Cruz-Correia R. Mental Health Professionals\u0026rsquo; Attitudes Toward Digital Mental Health Apps and Implications for Adoption in Portugal: Mixed Methods Study. JMIR Hum Factors. 2023;10:1-19. doi: 10.2196/45949.\u003c/li\u003e\n \u003cli\u003eZhang X, Lewis S, Chen X, Berry N, Bucci S. Mental health professionals views and the impact of COVID-19 pandemic on implementing digital mental health in China: A nationwide survey study. Internet Interv. 2022;30:1-11. doi: 10.1016/j.invent.2022.100576.\u003c/li\u003e\n \u003cli\u003eRifkin-Zybutz R, Turner N, Derges J, Bould H, Sedgewick F, Gooberman-Hill R, et al. Digital Technology Use and Mental Health Consultations: Survey of the Views and Experiences of Clinicians and Young People. JMIR Ment Health. 2023;10:1-14. doi: 10.2196/44064.\u003c/li\u003e\n \u003cli\u003eBucci S, Berry N, Morris R, Berry K, Haddock G, Lewis S, et al. \u0026ldquo;They Are Not Hard-to-Reach Clients. We Have Just Got Hard-to-Reach Services.\u0026rdquo; Staff Views of Digital Health Tools in Specialist Mental Health Services. Frontiers in Psychiatry. 2019;10:1-14. doi: 10.3389/fpsyt.2019.00344.\u003c/li\u003e\n \u003cli\u003eCliffe B, Croker A, Denne M, Stallard P. Clinicians\u0026apos; use of and attitudes towards technology to provide and support interventions in child and adolescent mental health services. Child and Adolescent Mental Health. 2020;25(2):95-101. doi: 10.1111/camh.12362.\u003c/li\u003e\n \u003cli\u003eAtmakuru A, Shahini A, Chakraborty S, Seoni S, Salvi M, Hafeez-Baig A, et al. Artificial intelligence-based suicide prevention and prediction: A systematic review (2019\u0026ndash;2023). Information Fusion. 2025;114:1-29. doi: 10.1016/j.inffus.2024.102673.\u003c/li\u003e\n \u003cli\u003eD\u0026rsquo;Hotman D, Loh E, Savulescu J. AI-enabled suicide prediction tools: ethical considerations for medical leaders. BMJ Leader. 2021;5(2):102-7. doi: 10.1136/leader-2020-000275.\u003c/li\u003e\n \u003cli\u003eVaridel M, Hickie IB, Prodan A, Skinner A, Marchant R, Cripps S, et al. Dynamic learning of individual-level suicidal ideation trajectories to enhance mental health care. npj Mental Health Research. 2024;3(1):26. doi: 10.1038/s44184-024-00071-0.\u003c/li\u003e\n \u003cli\u003evan Spijker BAJ, Batterham PJ, Calear AL, Farrer L, Christensen H, Reynolds J, et al. The Suicidal Ideation Attributes Scale (SIDAS): Community-Based Validation Study of a New Scale for the Measurement of Suicidal Ideation. Suicide and Life-Threatening Behavior. 2014;44(4):408-19. doi: 10.1111/sltb.12084.\u003c/li\u003e\n \u003cli\u003eIorfino F, Cross SP, Davenport T, Carpenter JS, Scott E, Shiran S, et al. A Digital Platform Designed for Youth Mental Health Services to Deliver Personalized and Measurement-Based Care. Front Psychiatry. 2019;10:1-9. doi: 10.3389/fpsyt.2019.00595.\u003c/li\u003e\n \u003cli\u003eCapon W, Hickie IB, McKenna S, Varidel M, Richards M, LaMonica HM, et al. Characterising variability in youth mental health service populations: A detailed and scalable approach using digital technology. Australasian Psychiatry. 2023;31(3):295-301. doi: 10.1177/10398562231167681.\u003c/li\u003e\n \u003cli\u003ePosner K, Brown GK, Stanley B, Brent DA, Yershova KV, Oquendo MA, et al. The Columbia\u0026ndash;Suicide Severity Rating Scale: Initial Validity and Internal Consistency Findings From Three Multisite Studies With Adolescents and Adults. American Journal of Psychiatry. 2011;168(12):1266-77. doi: 10.1176/appi.ajp.2011.10111704.\u003c/li\u003e\n \u003cli\u003ePapoutsi C, Wherton J, Shaw S, Morrison C, Greenhalgh T. Putting the social back into sociotechnical: Case studies of co-design in digital health. Journal of the American Medical Informatics Association. 2021;28(2):284-93. doi: 10.1093/jamia/ocaa197.\u003c/li\u003e\n \u003cli\u003ePhillippi J, Lauderdale J. A guide to field notes for qualitative research: Context and conversation. Qualitative Health Research. 2018;28(3):381-8. doi: 10.1177/1049732317697102.\u003c/li\u003e\n \u003cli\u003eMarshall B, Cardon P, Poddar A, Fontenot R. Does Sample Size Matter in Qualitative Research?: A Review of Qualitative Interviews in is Research. Journal of Computer Information Systems. 2013;54(1):11-22. doi: 10.1080/08874417.2013.11645667.\u003c/li\u003e\n \u003cli\u003eMalterud K, Siersma VD, Guassora AD. Sample Size in Qualitative Interview Studies: Guided by Information Power. Qualitative Health Research. 2015;26(13):1753-60. doi: 10.1177/1049732315617444.\u003c/li\u003e\n \u003cli\u003eChristie GI, Shepherd M, Merry SN, Hopkins S, Knightly S, Stasiak K. Gamifying CBT to deliver emotional health treatment to young people on smartphones. Internet Interv. 2019;18:1-9. doi: 10.1016/j.invent.2019.100286.\u003c/li\u003e\n \u003cli\u003eBevan Jones R, Stallard P, Agha SS, Rice S, Werner-Seidler A, Stasiak K, et al. Practitioner review: Co-design of digital mental health technologies with children and young people. Journal of Child Psychology and Psychiatry. 2020;61(8):928-40. doi: 10.1111/jcpp.13258.\u003c/li\u003e\n \u003cli\u003eBraun V, Clarke V. Toward good practice in thematic analysis: Avoiding common problems and be(com)ing a knowing researcher. International Journal of Transgender Health. 2023;24(1):1-6. doi: 10.1080/26895269.2022.2129597.\u003c/li\u003e\n \u003cli\u003eBraun V, Clarke V. Reflecting on reflexive thematic analysis. Qualitative Research in Sport, Exercise and Health. 2019;11(4):589-97. doi: 10.1080/2159676X.2019.1628806.\u003c/li\u003e\n \u003cli\u003eBraun V, Clarke V. One size fits all? What counts as quality practice in (reflexive) thematic analysis? Qualitative Research in Psychology. 2021;18(3):328-52. doi: 10.1080/14780887.2020.1769238.\u003c/li\u003e\n \u003cli\u003eBraun V, Clarke V. Thematic analysis. In: Cooper H, Coutanche MN, McMullen LM, Panter AT, Rindskopf D, Sher KJ, editors. APA handbook of research methods in psychology: Research designs: Quantitative, qualitative, neuropsychological, and biological. 2nd ed. Washington, D.C.: American Psychological Association; 2023. p. 65-81.\u003c/li\u003e\n \u003cli\u003eBraun V, Clarke V. A critical review of the reporting of reflexive thematic analysis in Health Promotion International. Health Promot Int. 2024;39(3):1-12. doi: 10.1093/heapro/daae049.\u003c/li\u003e\n \u003cli\u003eSteurer J, Fischer JE, Bachmann LM, Koller M, ter Riet G. Communicating accuracy of tests to general practitioners: a controlled study. BMJ. 2002;324:824-6. doi: 10.1136/bmj.324.7341.824.\u003c/li\u003e\n \u003cli\u003eSchneider H, Wayrauther J, Hassib M, Butz A. Communicating Uncertainty in Fertility Prognosis. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. Glasgow, Scotland UK: ACM; 2019. p. 1-11.\u003c/li\u003e\n \u003cli\u003eWhiting PF, Davenport C, Jameson C, Burke M, Sterne JAC, Hyde C, et al. How well do health professionals interpret diagnostic information? A systematic review. BMJ open. 2015;5(7):1-8. doi: 10.1136/bmjopen-2015-008155.\u003c/li\u003e\n \u003cli\u003eHe G, Buijsman S, Gadiraju U. How Stated Accuracy of an AI System and Analogies to Explain Accuracy Affect Human Reliance on the System. Proc ACM Hum-Comput Interact. 2023;7(CSCW2):1-29. doi: 10.1145/3610067.\u003c/li\u003e\n \u003cli\u003eHe X, Hong Y, Zheng X, Zhang Y. What Are the Users\u0026rsquo; Needs? Design of a User-Centered Explainable Artificial Intelligence Diagnostic System. International Journal of Human\u0026ndash;Computer Interaction. 2023;39(7):1519-42. doi: 10.1080/10447318.2022.2095093.\u003c/li\u003e\n \u003cli\u003eXian X, Chang A, Xiang Y-T, Liu MT. Debate and Dilemmas Regarding Generative AI in Mental Health Care: Scoping Review. Interact J Med Res. 2024;13:1-18. doi: 10.2196/53672.\u003c/li\u003e\n \u003cli\u003eRamadan OME, Alruwaili MM, Alruwaili AN, Elsehrawy MG, Alanazi S. Facilitators and barriers to AI adoption in nursing practice: a qualitative study of registered nurses\u0026apos; perspectives. BMC Nursing. 2024;23:1-16. doi: 10.1186/s12912-024-02571-y.\u003c/li\u003e\n \u003cli\u003eCousineau M, Verter V, Turecki G. Impact of Psychiatric Follow-Up Frequency on Outcomes and Waiting Times. Am J Manag Care. 2024;30(2):e52-e8. doi: 10.37765/ajmc.2024.89501.\u003c/li\u003e\n \u003cli\u003eHawthorne SCC, Williams-Wengerd A. Is treatment helping? How providers gauge effectiveness in treating serious mental illness. SSM - Mental Health. 2022;2:1-8. doi: 10.1016/j.ssmmh.2022.100110.\u003c/li\u003e\n \u003cli\u003eJoyce DW, Kormilitzin A, Smith KA, Cipriani A. Explainable artificial intelligence for mental health through transparency and interpretability for understandability. npj Digital Medicine. 2023;6:1-7. doi: 10.1038/s41746-023-00751-9.\u003c/li\u003e\n \u003cli\u003eHe X, Zheng X, Ding H. Existing Barriers Faced by and Future Design Recommendations for Direct-to-Consumer Health Care Artificial Intelligence Apps: Scoping Review. J Med Internet Res. 2023;25:1-22. doi: 10.2196/50342.\u003c/li\u003e\n \u003cli\u003eMoggia D, Lutz W, Brakemeier E-L, Bickman L. Treatment Personalization and Precision Mental Health Care: Where are we and where do we want to go? Administration and Policy in Mental Health and Mental Health Services Research. 2024;51:1-6. doi: 10.1007/s10488-024-01407-w.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-health-services-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bhsr","sideBox":"Learn more about [BMC Health Services Research](http://bmchealthservres.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/BHSR/default.aspx","title":"BMC Health Services Research","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Artificial intelligence, machine learning, suicide prediction, practitioner attitudes, interviews, qualitative research","lastPublishedDoi":"10.21203/rs.3.rs-7062632/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7062632/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground:\u003c/strong\u003e Assessment of suicidal ideation in youth mental health remains largely static, subjective, and inaccurate. Digital tools present a possible solution. This research examines a novel digital mental health tool, the Dynamic Learning Tool, which applies a machine learning model for individual-level continuous-time predictive trajectories of clients’ suicidal ideation to inform clinical care and was developed for an existing platform used in Australian and Canadian clinical practice. This work aims to explore professionals’ attitudes towards machine learning of predictive trajectories and artificial intelligence and thereafter determine the Dynamic Learning Tool’s design and content requirements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e Following a co-design methodology, semi-structured interviews and usability testing were conducted with 21 mental health professionals in Australia, Canada, and the United States. Data analysis employed inductive reflexive thematic analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003c/strong\u003e Findings indicate that professionals are open to using predictive trajectories strictly to enhance their decision-making but not replace clinical judgement. Furthermore, they emphasised the need for greater machine learning model transparency, incorporation of real-world contextual data, clear guidelines for use, and overt adherence to privacy and accountability standards.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions:\u003c/strong\u003e The findings presented here have broader implications for the ethics, design, development, implementation, governance, and evaluation of machine learning of predictive trajectories and artificial intelligence in youth mental health.\u003c/p\u003e","manuscriptTitle":"Mental health professionals’ perspectives on dynamic learning of individual-level trajectories in youth mental health care: A qualitative study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-14 15:42:50","doi":"10.21203/rs.3.rs-7062632/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2025-11-30T01:08:16+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"180587170332839588641944006782516036976","date":"2025-11-14T04:05:25+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-02T04:47:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"323239235603059738806854497451124188908","date":"2025-08-28T07:08:18+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-26T16:17:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"62815923523900841461296478359140811423","date":"2025-08-26T16:03:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-10T00:46:24+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-10T00:35:11+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-07-09T07:26:59+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-09T06:38:43+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Health Services Research","date":"2025-07-09T06:35:03+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-health-services-research","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bhsr","sideBox":"Learn more about [BMC Health Services Research](http://bmchealthservres.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/BHSR/default.aspx","title":"BMC Health Services Research","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e48e98de-0d3a-4352-9716-b5e16c7d98ab","owner":[],"postedDate":"July 14th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-07-14T15:42:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-14 15:42:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7062632","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7062632","identity":"rs-7062632","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}