Suicide attempt is defined as engaging in self-harm behavior with at least some intention to die (1) and its prevalence during adolescence is estimated to be up to 10% with a peak around the age of 16 (2). Cross-national studies also emphasize the high prevalence of suicide attempts during adolescence and early adulthood (1, 3). Thus, identifying adolescents at risk of attempting suicide is a key step toward offering early interventions and mitigating its substantial public health impact. Yet, identifying at-risk adolescents is a challenging task, because suicide attempt risk is tied to a variety of individual, societal, cultural, and environmental factors (4) and despite decades of extensive research, the predictive power of identified risk factors remains weak and inconsistent (5). A systematic meta-analysis on clinical instruments revealed a low combined positive predictive value of 5.5% for suicidal behavior, concluding that the “high-risk” classification was not clinically useful (6). Adolescents’ suicide attempts can also be unpremeditated or compulsive, making risk estimation even more difficult (7–9). As a result, recent reviews noted that prediction of suicidal behavior needs to account for the complex interplay of a multitude of risk factors, shifting the focus from risk factors to risk algorithms (5, 10).

Not only is the prediction of suicide attempt risk a challenging endeavor, but the sensitive subject matter also poses difficulties for assessment in the general population (11–13). Therefore, suicidal behavior is often not measured in population-based studies (12, 14, 15). Research has shown that prevalence of self-reported suicidal behavior in anonymous surveys is up to three times higher than non-anonymous surveys, further stressing sensitivity of suicide-related items (16, 17). Therefore, recent studies have stressed the importance of identifying individuals at risk for suicide attempts when suicide-related items, such as suicidal ideating and self-harm behavior are not assessed (18).

Recently, the potential of machine learning techniques to improve suicide attempt classification has been explored. Machine learning algorithms can make use of a multitude of variables and evaluate their complex interplay, resulting in promising improvements in the field of suicide research (19–21). However, these studies have focused on variables directly related to suicide, such as suicidal ideation and self-harm behaviors, and consistently identified them as the strongest predictors (19, 22–26). Moreover, there seems to be insufficient focus on the practicality of implementing machine learning algorithms at a population level, especially when suicide-related data are lacking. Although machine learning is regarded as a promising tool for predicting suicide attempts, a strong focus on suicide-related data may serve as a barrier to its implementation at scale. As a result, less is known about the performance of machine learning models in predicting suicide attempts when suicide-related items are not included (18).

The prevalence of self-reported recent suicide attempt in the sample was 4.66% (n = 8,090), with a male prevalence of 3.37% (n = 2,824) and a female prevalence of 5.87% (n = 5,040). There were 2.28% missing data on gender, and a Chi-square test revealed a significant difference between genders [χ² (1) = 579.52, p < 0.001]. According to Table 1, students in their third year of junior high school (age 15) and first year of senior high school (age 16) reported the highest prevalence of self-reported recent suicide attempts, both at a rate of 5.21%. Conversely, the lowest rate was observed among third-year senior high school students, at a rate of 3.20%. Students who reported no parental higher education had higher suicide attempt rates compared to others, as detailed in Table 1. Additionally, the majority of adolescents who reported a recent suicide attempt had not met with a psychologist (56.22%) or visited a health clinic for adolescents (69.67%), a school nurse (42.89%), or an accident and emergency ward (51.42%) in the past 12 months.

The aim of this study was to identify adolescents who had attempted suicide during the past 12 months without utilizing other sensitive suicide-related survey items as predictors. In doing so, we employed population-based national data to fine-tune the XGBoost machine learning algorithm based on a multitude of psychological, social, and environmental survey items. The fine-tuned model reached an excellent AUC of 92.1% (61). This figure is near the highest AUC reported in existing literature, ranging from 0.716 to 0.925 (23, 25, 26, 28). With a high specificity of 0.90 and a high sensitivity of 0.77, the model could correctly classify 77% of adolescents reporting a recent suicide attempt as well as 90% of adolescents not reporting a recent suicide attempt. Such performance is promising and indicates the feasibility of classifying adolescents’ suicide attempts at scale with substantial accuracy, yet without utilizing any suicide-related items. Consequently, our findings answer our first research question, providing substantial evidence that high classification accuracy for adolescents’ recent suicide attempts at the population scale can indeed be achieved without the need to rely on suicide-related items. This result has important implications for the broader field and could guide future efforts in developing scalable, sensitive, and effective tools for identifying and addressing suicide risk in young populations.

To address our second research question, we pinpointed the most important predictors and provisionally grouped them within four main domains of internalizing problems, interpersonal relationships, victimization, and substance use. Our findings are in line with existing literature showing that internalizing problems, particularly depressive symptoms, are key predictors of suicidal behavior among adolescents (7, 8, 62, 63). Similar to our findings, Carballo et al., (62) also identified conflicts in interpersonal relationships to be among the major risk factors of suicidality among adolescents [see also (64)]. Moreover, a review by Calati et al. (65) underscores the importance of interpersonal factors for suicidal outcomes. Reviews have also repeatedly emphasized the importance of childhood maltreatment and bullying experiences for understanding suicidal behavior (7, 9, 66, 67), thereby supporting the role of victimization experiences. Finally, substance use and more specifically, smoking, was also among the most important predictors of adolescents’ suicide attempt. This aligns with numerous studies that recognize smoking as a general risk factor, independent of other socio-demographic variables (68–71), although the association between smoking and suicide has been debated (72). Our analysis accentuates that frequent smoking is indeed related to adolescent suicide attempts, even when accounting for hundreds of other items.

In our analysis, while focusing on the important predictors of suicide attempts, we also observed that certain mental health-related factors were not as influential in the model as might have been expected based on prior literature. This warrants a closer examination of these seemingly inconsistent findings. For example, looking at Tables 1 and 2, despite significant variations in the suicide attempt rate across socio-demographic variables such as age, gender, and socio-economic status, only age emerged as one of the important predictors. Moreover, among the top 15 predictors, age contributed among the least in terms of SHAP values. The fact that gender was not identified as an important predictor is particularly noteworthy. This contrasts with existing literature, including findings by Esang and Ahmed (69), where gender has been recognized as a significant factor. Given the marked prevalence differences between boys and girls in our dataset, the lack of gender’s role in the model requires careful consideration. A plausible explanation for this inconsistency may lie in the complex interplay of internalizing problems, interpersonal problems, and victimization experiences. These domains may subsume much of the demographic differences related to suicide attempts, rendering gender a comparatively less important predictor. Finally, in line with the existing literature, externalizing problems and delinquency did not emerge among the most important predictors of adolescent suicide attempts (69).

The utility of models requiring suicide-specific information for risk estimation of suicide attempts is questionable because such data are challenging to collect on a population scale, especially from adolescents. Our study demonstrates that it is possible to classify adolescents reporting recent suicide attempts with high accuracy without depending on such sensitive items. A novel contribution of our study is the identification of key factors—namely, internalizing problems, interpersonal relationships, victimization, and substance use—as important predictors of suicide attempts in this population, while accounting for a myriad of other items across different domains. A focus on these aspects is warranted when identifying at-risk adolescents, particularly from longitudinal research, which can offer invaluable insights into the causal dynamics between identified factors and adolescence suicide attempts.

Future research should address the limitations of our current study and align its findings with existing theoretical frameworks on suicidal behavior. Subsequent investigations should also examine the latent factorial structure of these predictors and consider whether different machine learning algorithms would identify similar important items. There is also an opportunity to explore the use of stacked ensemble meta-learners, constructed by combining multiple machine learning models, for achieving superior classification performance.

The data used in the current research is not open-access. However, researchers can apply to access the dataset via Ungdata.no website. Further inquiries can be directed to the corresponding author.

The studies involving humans were approved by Ethical committee at the Department of Psychology, University of Oslo. The studies were conducted in accordance with the local legislation and institutional requirements. Informed consent for participation in this study was provided by the participants.

EFH conceptualized the study and the research methods, carried out the data cleaning, missing data imputation, data analysis and visualization, as well as literature review and writing of the draft. NC and TVS were engaged in all steps of the study and provided input regarding multiple theoretical, technical, practical challenges, provided feedback on several drafts, and helped with revising the manuscript. All authors approved the final version of the manuscript for submission.