Identifying minimal risk factors for adolescent suicidal ideation and suicide attempts: A machine learning-optimized approach

Catherine Park; Beom-Chan Lee

doi:10.1371/journal.pone.0346050

. 2026 Apr 6;21(4):e0346050. doi: 10.1371/journal.pone.0346050

Identifying minimal risk factors for adolescent suicidal ideation and suicide attempts: A machine learning-optimized approach

Catherine Park ^1,², Beom-Chan Lee ^3,^4,^*

Editor: Vincenzo De Luca⁵

PMCID: PMC13052902 PMID: 41941504

Abstract

This study aimed to develop and validate a machine learning (ML) model to identify the minimal risk factors for adolescent suicidal behaviors, including suicidal ideation and suicide attempts. Data from the Korea Youth Risk Behavior Web-based Survey (2022–2023), including 90,813 adolescents aged 12–18 years, were analyzed. Using multidimensional risk factors spanning sociodemographic, physical and mental health, and behavioral domains, we applied a Random Forest model combined with recursive feature elimination to identify a minimal subset of risk factors (optimal features). Model performance for identifying suicidal ideation and predicting suicide attempts was evaluated via area under the curve (AUC), sensitivity, specificity, and accuracy metrics across the validation datasets. Sadness, loneliness, anxiety, and stress were identified as optimal features, achieving a high AUC, sensitivity, specificity, and accuracy in identifying suicidal ideation and predicting suicide attempts. Additional factors further improved the ML model’s predictive performance for suicide attempts, achieving an AUC of 97.28%, sensitivity of 93.49%, specificity of 90.21%, accuracy of 91.85%, a PPV of 90.52%, and an NPV of 93.26%. This study demonstrated the efficacy of ML-driven approaches in identifying critical risk factors for adolescent suicidal behaviors. The findings highlight the potential of ML frameworks to transform suicide prevention strategies and improve mental health outcomes in adolescents.

Introduction

Adolescent suicide is a global public health concern. According to the World Health Organization, more than 720,000 individuals die by suicide annually, comprising a substantial proportion of adolescents and young adults [1]. Although suicide is the third leading cause of death among adolescents worldwide [1], the prevalence of suicidal ideation and behavior among adolescents varies considerably across regions and countries. A study analyzing data from 90 countries reported a high prevalence of suicidal thoughts among adolescents aged 13–17 years [2]. A recent study revealed that while suicide rates among individuals aged 10–24 years generally declined between 1990 and 2021, some regions continue to report alarmingly high rates [3].

The multifaceted nature of adolescent suicides requires a comprehensive understanding of its underlying determinants. Psychological factors, such as depression and anxiety, have long been recognized as primary contributors to adolescent suicide [4,5]. However, recent umbrella reviews and meta-analyses have highlighted the important role of sociodemographic factors (e.g., academic achievement, academic stress, family structure, and family socioeconomic status) and health-related lifestyle factors (e.g., physical activity, smoking, and alcohol consumption) in contributing to suicidal behaviors among adolescents [6–8]. This growing body of evidence underscores the need for methodologies capable of analyzing and integrating these diverse and complex risk factors.

In recent years, machine learning (ML) has revolutionized suicide risk assessments in adolescents by leveraging large-scale electronic health records and survey data to identify the intricate patterns and interactions among risk factors. Recent systematic reviews and meta-analyses have indicated the increasing application of ML models with algorithms such as random forest and extreme gradient boosting to identify and predict suicidal behaviors (i.e., suicidal ideation and suicide attempts) [9,10]. These studies identified a broad spectrum of risk factors, including psychological determinants, social and familial influences, and behavioral and lifestyle attributes [9,10]. Notably, ML-based approaches have outperformed conventional statistical methods, effectively handling high-dimensional data, capturing nonlinear and intricate patterns, and adapting dynamically to diverse data sources [9,10]. Furthermore, their scalability and flexibility enable early detection, enhanced predictive accuracy, and personalized interventions [9,10].

Despite these advancements, the minimal subset of risk factors required for the accurate and reliable identification and prediction of suicidal ideation and suicide attempts is yet to be established. This could potentially reduce computational overheads, enhance the interpretability of ML models, and facilitate clinical integration. To address this gap, this study aimed to 1) develop and validate an ML model that incorporates multidimensional risk factors, including sociodemographic characteristics, physical and mental health, and health-related behaviors, to identify and predict suicidal ideation and suicide attempts and 2) determine the minimal subset of risk factors (i.e., optimal features) required to ensure accurate, reliable, and clinically actionable identification of suicidal ideation and prediction of suicide attempts.

Methods

Data source, participant selection, and multidimensional risk factors

We utilized datasets from the Korea Youth Risk Behavior Web-based Survey (KYRBS) collected from 2022 to 2023. The KYRBS is a nationwide, self-administered survey designed to monitor health behaviors and risk factors among Korean adolescents (aged 12–18 years) and has been conducted annually by the Korea Disease Control and Prevention Agency since 2005 [11]. The 2022 and 2023 KYRBS were conducted as independent, nationally representative cross-sectional surveys using random sampling of schools and students in South Korea; based on direct confirmation from the KYRBS administrators, there was no respondent overlap between the two survey years.

The KYRBS survey consists of 15 independent sections, including 103 questions that address sociodemographic characteristics, physical and mental health, and health-related behaviors. The use of the KYRBS data and study protocol were approved by the Institutional Review Board of Yonsei University (IRB No. 1041849–202411-BM-226–01; approval date: November 7, 2024), and the study was conducted in accordance with the Declaration of Helsinki. Data access for research purposes occurred on November 8, 2024, following IRB approval. Because this retrospective analysis relied exclusively on anonymized data, the need for written informed consent was waived in accordance with national regulations and institutional policies.

Fig 1 illustrates the participant selection and group classification processes. We initially extracted responses from 54,383 adolescents from the 2022 KYRBS and 55,413 adolescents from the 2023 KYRBS, totaling 109,796 participants. Adolescents who did not respond to any of the 103 questions were excluded from the analysis (N = 18,983), resulting in a final sample size of 90,813 adolescents included in this study. A total of 90,813 adolescents were classified into two groups based on their responses to questions regarding suicidal thoughts and suicide planning as well as well-established classification criteria [3]. The non-suicidal ideation group included adolescents who provided negative (or No) responses to both suicidal thoughts and suicide planning (n = 78,975). The suicidal ideation group included adolescents who provided positive (Yes) responses to suicidal thoughts, suicide planning, or both (n = 11,838). Suicidal ideation was assessed by the question asking whether the participant had ever seriously considered suicide during the past 12 months (yes/no) as part of the KYRBS. Participants who answered “yes” were categorized as the suicidal ideation group (n = 11,838), whereas those who answered “no” were categorized as the non-suicidal ideation group (n = 78,975). For the analysis of suicide attempts, participants were further classified according to their responses to the question asking whether they had ever attempted suicide during the past 12 months (yes/no), resulting in 2,191 participants in the suicide attempt group; additionally, based on direct confirmation from the KYRBS administrators, there was no participant overlap between the suicidal ideation group and the suicide attempt subgroup.

Building on the findings of recent systematic reviews and meta-analyses [9,10], we investigated multidimensional risk factors among the 90,813 adolescents included in this study, focusing on their associations with suicidal behaviors, including suicidal ideation and suicide attempts. Specifically, multidimensional risk factors were categorized into four domains: sociodemographic characteristics, physical health, mental health, and health-related behaviors [9,10]. The sociodemographic factors included sex, academic grade, residential area type, academic performance, and family economic status. Physical health factors included physical activity level, sleep quality, and self-reported health status. Mental health factors included stress, anxiety, sadness, and loneliness. Health-related behavioral factors included alcohol consumption and smoking status. A total of 14 factors across these four domains were analyzed. Furthermore, we integrated responses to suicidal behaviors, including suicidal thoughts, planning, and attempts, into statistical analyses and ML models.

Statistical analysis

To evaluate the effects of these 14 factors on suicidal ideation, we conducted statistical analyses using the IBM SPSS Statistics software (IBM Corp., Armonk, NY, USA). Categorical variables are presented as numbers and percentages, whereas continuous variables are reported as means ± standard deviations. Categorical variables were analyzed using the chi-square test to determine significant differences between the non-suicidal ideation and suicidal ideation groups. For continuous variables, the Shapiro–Wilk test was performed to assess normality. As all continuous variables were not normally distributed, the Mann–Whitney U test was used to compare differences in variables between the non-suicidal ideation and suicidal ideation groups.

To ensure a robust interpretation of the results, effect sizes were calculated to address the possibility that statistically significant differences resulted from a large sample size rather than meaningful effects. Specifically, the Cramer’s V value was calculated to measure the strength of the association between categorical variables, with values ranging from 0 (no association) to 1 (perfect association), making it particularly suitable for chi-square analysis [12]. For continuous variables, Cohen’s d was calculated to quantify the magnitude of the differences between group means, expressed in standard deviation units, facilitating consistent interpretation across variables [13]. Statistical significance was set at two-sided p < 0.05.

Machine learning framework and optimal feature selection

The ML framework depicted in Fig 2 incorporates several key processes: 1) training and validation of the ML model, 2) identification of the minimal subset of risk factors (i.e., optimal features) associated with suicidal ideation, and 3) validation of the model’s performance in predicting suicide attempts. We implemented and executed this ML framework using custom scripts created in MATLAB (MathWorks, Natick, MA, USA).

To address the class imbalance between the non-suicidal ideation group (n = 78,975) and the suicidal ideation group (n = 11,838), we applied the synthetic minority oversampling technique (SMOTE) to the dataset. SMOTE generates synthetic samples for the minority class to achieve a balanced distribution within the training and validation datasets [14]. Our ML framework incorporated bootstrapping and recursive feature elimination, both of which are well-established techniques for enhancing ML performance [15,16]. Considering the sample sizes (i.e., n = 78,975 and n = 11,838) and recommendations from previous studies [17,18], 50 bootstrapping iterations were conducted to ensure robustness in assessing the performance of the ML model across varying data distributions.

The Random Forest (RF) model was selected based on the evidence from recent systematic reviews and meta-analyses [9,10], which demonstrate its widespread application and high predictive accuracy in identifying and forecasting suicidal behaviors among adolescents. The RF is an ensemble-based ML approach that combines multiple decision trees to produce robust and precise predictions.

In the RF model, the dependent variable was defined as a categorical variable, with 0 representing the non-suicidal ideation group and 1 representing the suicidal ideation group. The independent variables (i.e., features) were selected from the 14 factors analyzed based on their statistical significance and practical relevance [19]. Specifically, factors were included if they met the criterion of statistical significance (p < 0.05) and if they demonstrated effect sizes ranging from small to large. This rigorous selection methodology was applied to ensure that the RF model incorporated features with both statistically validated and practically meaningful contributions, thereby enhancing its interpretability and robustness in predicting suicidal ideation.

To evaluate feature importance, a permutation feature importance analysis was conducted using out-of-bag (OOB) data [20]. The predictive accuracy was initially assessed using the original feature values to calculate the baseline OOB error. Subsequently, the values of each feature were randomly permuted within the OOB dataset, and the OOB error was recalculated. The difference between the baseline and permuted OOB errors provided an importance score for each feature. The features with the lowest importance scores were iteratively removed, and this recursive process continued until only one feature remained (i.e., recursive feature elimination).

Following the ranking of features through recursive feature elimination, the RF models were re-evaluated using validation datasets to assess their performance with the reduced feature set. During the validation process, performance metrics, such as the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy, were calculated [15,16,21]. The AUC measured the model’s capacity to distinguish between adolescents with and without suicidal ideation, serving as an overall indicator of classification effectiveness. Sensitivity reflected the model’s ability to accurately identify cases of suicidal ideation, whereas specificity quantified its precision in correctly classifying non-suicidal ideation. The accuracy represented the overall classification performance of the model across all categories.

The optimal feature set was determined when all performance metrics (AUC, sensitivity, specificity, and accuracy) consistently exceeded 80%, ensuring robust and reliable model performance. Notably, an AUC of > 80% is generally considered excellent [22].

To evaluate the predictive performance of the different ML models using either the optimal feature set or sets that included high-ranking features beyond the optimal set for predicting suicide attempts, the same dataset comprising adolescents with and without suicide attempts (n = 88,622 and n = 2,191, respectively) was used. SMOTE was applied to the dataset to address the class imbalance and generate validation datasets. Model validation was conducted using 50 bootstrapping iterations, during which AUC, sensitivity, specificity, and accuracy were calculated to assess the model’s predictive capabilities.

Ethics statement

The use of the KYRBS data and study protocol were approved by the Institutional Review Board of Yonsei University (IRB No. 1041849–202411-BM-226–01; approval date: November 7, 2024). Owing to the retrospective design of the study using anonymized data, the need to obtain informed consent was waived by the Institutional Review Board of Yonsei University. All methods in this study were performed in accordance with the guidelines and regulations of the Declaration of Helsinki. Data access for research purposes occurred on November 8, 2024, following IRB approval.

Results

Multidimensional risk factors contributing to suicidal ideation

Table 1 presents the descriptive statistics and results of the statistical analyses of the multidimensional risk factors associated with suicidal ideation. Among the sociodemographic factors, sex, academic performance, and family economic status demonstrated significant differences between the groups. Female adolescents demonstrated a higher prevalence of suicidal ideation than males. Academic performance and family economic status were poorer in the suicidal ideation group than in the non-suicidal ideation group.

Table 1. Multidimensional risk factors associated with suicidal ideation, including sociodemographic characteristics, physical health,mental health, and health-related behaviors.

Variables	Non-suicidal ideation (n = 78,975)	Suicidal ideation (n = 11,838)	p-value	Effect size
Sociodemographic factors
*Sex (female), n (%)	37,887 (47.97)	7,599 (64.19)	< 0.0001	0.11
Academic grade (high school), n (%)	36,679 (46.44)	4,731 (39.96)	< 0.0001	0.04
Residential area type, score (1: big–5: small)	1.64 ± 0.62	1.65 ± 0.60	0.208	0.01
Academic performance, score (1: good–5: bad)	2.86 ± 1.14	3.04 ± 1.21	< 0.0001	0.15
*Family economic status, score (1: good–5: bad)	2.57 ± 0.85	2.75 ± 0.95	< 0.0001	0.20
Physical health factors
Physical activity levels, score (1: none–8: 7 days)	3.19 ± 2.16	3.11 ± 2.15	< 0.0001	0.04
*Sleep quality, score (1: good–5: bad)	3.18 ± 1.09	3.67 ± 1.05	< 0.0001	0.46
*Self-reported health status, score (1: good–5: bad)	2.19 ± 0.87	2.68 ± 0.97	< 0.0001	0.53
Mental health factors
*Sadness, n (%)	15,994 (20.25)	8,255 (69.73)	< 0.0001	0.38
*Stress, score (1: a lot–5: none)	2.83 ± 0.88	1.95 ± 0.81	< 0.0001	1.01
*Anxiety, score (1: minimal–4: severe)	2.60 ± 0.70	3.41 ± 0.69	< 0.0001	1.12
*Loneliness, score (1: none–5: always)	2.42 ± 1.00	3.50 ± 0.96	< 0.0001	1.10
Health-related behavioral factors
Alcohol consumption, n (%)	24,653 (31.22)	5,105 (43.12)	< 0.0001	0.09
Smoking status, n (%)	5,658 (7.16)	1,647 (13.91)	< 0.0001	0.08

Open in a new tab

The asterisk (*) indicates statistically significant and practically relevant effects

Physical and mental health factors were significantly associated with suicidal ideation. Adolescents in the suicidal ideation group reported significantly poorer sleep quality and lower self-reported health status than those in the non-suicidal ideation group. A significantly larger proportion of adolescents in the suicidal ideation group experienced sadness than those in the non-suicidal ideation group. Additionally, stress levels, anxiety, and loneliness scores were significantly higher in the suicidal ideation group.

Health-related behavioral factors, including alcohol consumption and smoking status, were also significantly more prevalent among adolescents with suicidal ideation. The suicidal ideation group demonstrated a significantly higher prevalence of alcohol consumption and smoking than did the non-suicidal group.

Among the 14 multidimensional risk factors examined, 8 factors (sex, family economic status, sleep quality, self-reported health status, sadness, stress, anxiety, and loneliness) were statistically significant (p < 0.05) and practically relevant (i.e., small to large effect sizes) in their association with suicidal ideation. Consequently, these eight factors were included in the machine learning model.

Performance of the machine learning model and optimal feature selection

Table 2 presents the rankings of the eight factors included in the ML model and the validation results for suicidal ideation. The results were evaluated using four performance metrics: AUC, sensitivity, specificity, and accuracy. Fig 3A illustrates the average values of these performance metrics (AUC, sensitivity, specificity, and accuracy) with their corresponding 95% confidence intervals plotted as a function of the ranked factors. Fig 3B presents the traditional receiver operating characteristic (ROC) curves for the eight cumulative models used to identify suicidal ideation, with all curves overlaid in a single graph (sensitivity vs. 1 – specificity). Each curve displays its respective AUC value, demonstrating that all models achieved strong discriminatory power (AUC > 78.86%) and that predictive performance progressively improved as additional features were incorporated.

Table 2. Ranking of the eight factors included in the machine learning model and validation results of the model using 50 validation datasets.

Rank	Factor	AUC (mean)	Sensitivity (mean)	Specificity (mean)	Accuracy (mean)	PPV (mean)	NPV (mean)	Included rank(s)
1	Sadness	78.76%	81.52%	74.48%	78.00%	76.16%	80.13%	1
2	Loneliness	86.77%	81.42%	78.93%	80.17%	79.44%	80.95%	1–2
3	Anxiety	89.05%	83.80%	80.12%	81.96%	80.82%	83.19%	1–3
4	Stress	90.91%	83.94%	83.50%	83.72%	83.57%	83.88%	1–4
5	Family economic status	93.04%	86.90%	84.46%	85.68%	84.83%	86.58%	1–5
6	Sleep quality	94.73%	88.97%	86.80%	87.88%	87.08%	88.73%	1–6
7	Self-reported health status	96.49%	92.06%	89.04%	90.55%	89.36%	91.81%	1–7
8	Sex	97.22%	93.51%	90.06%	91.79%	90.39%	93.28%	1–8

Open in a new tab

Fig 3 — AUC: area under the curve. PPV: Positive Predictive Value. NPV: Negative Predictive Value. (B) Overlaid receiver operating characteristic (ROC) curves for the eight cumulative models used to identify suicidal ideation (sensitivity vs. 1 – specificity). Each curve displays its corresponding AUC value, demonstrating progressive improvement in model performance with additional features. AUC: area under the curve.

The performance of the model stabilized and demonstrated its effectiveness for the four factors (sadness, loneliness, anxiety, and stress). Under these conditions, the model achieved an AUC exceeding 90%, accompanied by a sensitivity, specificity, accuracy, PPV (Positive Predictive Value), and NPV (Negative Predictive Value) exceeding 80%, thereby demonstrating robust predictive capabilities. Consequently, these four factors were determined to be the optimal features (i.e., the minimal subset of risk factors), which are required to maintain high classification performance while enhancing interpretability and computational efficiency.

Predictive performance of machine learning model for suicide attempts

Table 3 presents the predictive performances of the five ML models for suicide attempts, demonstrating progressive enhancement with the inclusion of additional features. Model 1, which incorporated the optimal features (sadness, loneliness, anxiety, and stress), achieved an AUC of 90.81%, a sensitivity of 84.09%, a specificity of 83.27%, an accuracy of 83.68%, a PPV of 83.41%, and an NPV of 83.95%. When family economic status was added to Model 2, the predictive performance improved, with an AUC of 93.06%, sensitivity of 87.16%, specificity of 84.21%, and accuracy of 85.68%, a PPV of 84.67%, and an NPV of 86.77%.

Table 3. Predictive performance of machine learning models for suicide attempts.

Model number	Features	AUC	Sensitivity	Specificity	Accuracy	PPV	NPV
1	Sadness, loneliness, anxiety, stress	90.81 ± 0.06%	84.09 ± 0.13%	83.27 ± 0.20%	83.68 ± 0.09%	83.41 ± 0.16%	83.95 ± 0.10%
2	Model 1 features combined with family economic status	93.06 ± 0.05%	87.16 ± 0.12%	84.21 ± 0.17%	85.68 ± 0.08%	84.67 ± 0.13%	86.77 ± 0.10%
3	Model 2 features combined with sleep quality	94.81 ± 0.04%	89.28 ± 0.09%	86.76 ± 0.15%	88.02 ± 0.06%	87.09 ± 0.12%	89.00 ± 0.08%
4	Model 3 features combined with self-reported health status	96.55 ± 0.03%	92.08 ± 0.08%	89.10 ± 0.10%	90.59 ± 0.05%	89.41 ± 0.09%	91.84 ± 0.07%
5	Model 4 features combined with sex	97.28 ± 0.02%	93.49 ± 0.08%	90.21 ± 0.09%	91.85 ± 0.05%	90.52 ± 0.07%	93.26 ± 0.08%

Open in a new tab

The reported values for area under the curve (AUC), sensitivity, specificity, accuracy, Positive Predictive Value (PPV), and Negative Predictive Value (NPV) represent the mean values ± their corresponding 95% confidence intervals calculated across validation datasets

Further enhancement was observed in Model 3 with the inclusion of sleep quality, resulting in an AUC of 94.81%, sensitivity of 89.28%, specificity of 86.76%, and accuracy of 88.02%, a PPV of 87.09%, and an NPV of 89.00%. Model 4, which included self-reported health status, demonstrated an even higher predictive performance, achieving an AUC of 96.55%, sensitivity of 92.08%, specificity of 89.10%, and accuracy of 90.59%, a PPV of 89.41%, and an NPV of 91.48%. Finally, Model 5, which incorporated sex, showed the highest predictive performance. The AUC reached 97.28%, sensitivity of 93.49%, specificity of 90.21%, and accuracy of 91.85%, a PPV of 90.52%, and an NPV of 93.26%.

Discussion

This study explored the significance of identifying a minimal set of critical risk factors for adolescent suicidal behaviors (i.e., suicidal ideation and suicide attempts) using an ML approach. Consistent with the existing literature, the findings of this study emphasize the multidimensional nature of adolescent suicidal behaviors, which includes sociodemographic characteristics, physical and mental health, and health-related behaviors. Using data from the KYRBS, this study demonstrated the efficacy of ML models in accurately identifying suicidal ideation and predicting suicide attempts. Notably, the ML models identified a minimal subset of risk factors (i.e., optimal features), including sadness, loneliness, anxiety, and stress, which ensured a robust and precise classification of suicidal ideation and the prediction of suicide attempts.

The results of this study highlight the critical role of optimal features as key risk factors for suicidal behaviors among adolescents. These results are consistent with a growing body of evidence emphasizing the psychological dimensions of suicidality in this population, particularly the strong association between elevated stress levels and increased suicide risk [23]. Sadness and loneliness were also identified as the most influential factors, consistent with previous studies indicating that emotional isolation exacerbates vulnerability to suicidal behaviors [24]. These results are particularly significant, given that adolescence represents a developmental period marked by elevated emotional sensitivity and social reliance, during which perceived or actual isolation can profoundly exacerbate psychological distress. Additionally, the role of anxiety is consistent with its established association with cognitive distortions, such as catastrophic thinking and rumination, which have been shown to intensify suicidal behavior [25,26].

In addition to psychological factors, this study identified sex, family economic status, sleep quality, and self-reported health status as significant secondary risk factors for suicidal behaviors. Previous research has consistently indicated that female adolescents exhibit higher rates of suicidal behaviors than their male counterparts. This disparity has been attributed to interrelated factors, including gender differences in emotional expression, coping mechanisms, and heightened exposure to stressors such as interpersonal conflicts and academic pressures [27]. Additionally, hormonal fluctuations during puberty may intensify emotional sensitivity and increase susceptibility to mood disorders, both of which are strongly associated with suicidal behaviors [28]. The association between economic hardship and an elevated risk of suicidal behaviors can be attributed to the cumulative stressors of financial instability, including restricted access to mental health resources and increased familial conflict, which collectively foster an environment conducive to psychological distress [29,30]. Additionally, sleep quality was identified as a significant risk factor, corroborating growing evidence linking sleep disturbances to emotional dysregulation and impaired executive function, both of which play pivotal roles in the regulation of suicidal behaviors [31,32]. Disrupted sleep patterns, such as insomnia and irregular sleep schedules, have been associated with diminished neural plasticity and altered emotional processing, thereby exacerbating vulnerability to suicidal behaviors in adolescents [33]. Similarly, the identification of self-reported health status as a significant risk factor highlights the intricate relationship between perceived physical well-being and mental health. Poor self-rated health often coexists with feelings of hopelessness and despair, which are foundational components of suicidality [34].

The robustness of the RF model in this study demonstrates its applicability in suicide risk assessment, consistent with prior findings demonstrating that ensemble methods surpass traditional statistical approaches and other ML models (e.g., logistic regression), when applied to high-dimensional datasets [9]. These findings could potentially advance the methodological framework for suicide risk assessment by validating the use of recursive feature elimination to identify optimal features. An AUC exceeding 90% with only four features suggests the potential to optimize predictive processes for clinical applications without compromising precision. This efficiency meets the growing demand for interpretable and computationally viable ML models in clinical contexts [35,36]. Moreover, recursive feature elimination enhances model interpretability and facilitates efficient integration into resource-constrained healthcare settings, enabling the early detection of high-risk adolescents. By prioritizing critical risk factors and maintaining a robust predictive performance, this approach can enhance decision-making processes in clinical environments with limited resources. The application of SMOTE to address class imbalances further reinforces the generalizability of the study’s findings [14]. Combined with rigorous validation via bootstrapping, the SMOTE ensured robust performance metrics (sensitivity, specificity, and AUC) across all validation sets. These methodological advancements align with established best practices for reliable ML systems to predict suicide risk behaviors [37,38].

Although this study provides valuable insights into the classification and prediction of suicidal ideation and suicide attempts among adolescents, two limitations must be acknowledged. First, reliance on self-reported survey data may introduce reporting bias, as adolescents may underreport or exaggerate their experiences due to social desirability or recall inaccuracies. Furthermore, although the KYRBS dataset offers comprehensive coverage of Korean adolescents, its generalizability to other cultural contexts remains uncertain, underscoring the need for further validation using datasets from diverse populations. Additionally, the KYRBS dataset used in this study does not include variables explicitly specifically assessing physical or sexual abuse. The absence of such trauma-related variables may limit the comprehensiveness of the risk factor assessment, as exposure to abuse is a well-documented determinant of suicidal behaviors among adolescents. Future studies incorporating datasets that include abuse- or trauma-related items are worthwhile to provide a more complete understanding of the multidimensional contributors to suicidal behaviors.

Nevertheless, identifying a minimal set of predictive factors has significant practical and clinical implications. Clinicians can allocate resources more effectively by prioritizing interventions for adolescents who exhibit high levels of sadness, loneliness, anxiety, and stress. Moreover, the simplicity of the model enhances its feasibility for integration into school-based mental health programs where early detection can play a pivotal role in preventing the escalation of suicidal behaviors. The demonstrated accuracy of our model in predicting suicide attempts further underscores its potential utility in preventive interventions, offering a promising framework for addressing adolescent mental health challenges.

Conclusion

This study emphasizes the critical importance of identifying the minimal risk factors for adolescent suicidal behaviors, including suicidal ideation and suicide attempts, using a rigorous ML approach. Using a nationally representative dataset of Korean adolescents, we developed a robust ML model that demonstrated its capability to accurately identify suicidal ideation and predict suicide attempts. Importantly, the results of this study identify a minimal subset of four risk factors (i.e., sadness, loneliness, anxiety, and stress) that are sufficient to maintain high predictive accuracy while enhancing model interpretability, which can reduce computational overhead and facilitate integration into clinical and school-based mental health programs. Additionally, the inclusion of additional factors, such as sex, family economic status, sleep quality, and self-reported health status, further enhances predictive performance, illustrating the interplay between sociodemographic, lifestyle, and psychological risk factors for suicidal behaviors.

The findings of this study have the potential to offer a scalable and efficient framework for early detection of high-risk adolescents. By focusing on the identified psychological dimensions, mental health professionals can implement precise and resource-efficient intervention strategies. The high accuracy of the ML model suggests its potential as a practical tool across diverse healthcare settings, enabling timely and effective interventions to prevent suicidal behaviors. In addition, these results advocate for the integration of ML-driven frameworks into educational systems to enhance suicide prevention efforts. School-based applications of the proposed model can support routine screenings and facilitate early risk identification, ensuring prompt support for vulnerable adolescents. Moreover, the interpretability of the model equips educators and counselors with actionable insights into critical risk factors, enabling data-driven and informed decision-making.

Acknowledgments

We are grateful to the nationwide Korea Youth Risk Behavior Web-based Survey (KYRBS) registry.

Data Availability

The data used in this study were obtained from the Korea Disease Control and Prevention Agency through the Korea Youth Risk Behavior Web-based Survey (KYRBS). The datasets are publicly available to all researchers via the official KDCA portal (https://kdca.go.kr/yhs/yhs/main.do). To replicate the study findings, interested researchers may download the raw datasets for the relevant survey years (e.g., 2022 and 2023) in SPSS format. The authors had no special access privileges to these data.

Funding Statement

This research was supported by the Ministry of Science and ICT (MSIT), Korea, under the National Program in Medical AI Semiconductor, supervised by the Institute of Information & Communications Technology Planning & Evaluation (IITP) in 2026 (2024-0-00096 to C.P.), and by the Regional Innovation System & Education (RISE) program through the Gangwon RISE Center, funded by the Ministry of Education and Gangwon State, Republic of Korea (2026-RISE-10-006 to C.P.).

References

1.WHO. Suicide. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/suicide. 2024. [Google Scholar]
2.Yan N, Luo Y, Mackay LE, Wang Y, Wang Y, Wang Y, et al. Global patterns and trends of suicide mortality and years of life lost among adolescents and young adults from 1990 to 2021: a systematic analysis for the Global Burden of Disease Study 2021. Epidemiol Psychiatr Sci. 2024;33:e52. doi: 10.1017/S2045796024000532 [DOI] [PMC free article] [PubMed] [Google Scholar]
3.Campisi SC, Carducci B, Akseer N, Zasowski C, Szatmari P, Bhutta ZA. Suicidal behaviours among adolescents from 90 countries: a pooled analysis of the global school-based student health survey. BMC Public Health. 2020;20(1):1102. doi: 10.1186/s12889-020-09209-z [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Ho TC, Gifuni AJ, Gotlib IH. Psychobiological risk factors for suicidal thoughts and behaviors in adolescence: a consideration of the role of puberty. Mol Psychiatry. 2022;27(1):606–23. doi: 10.1038/s41380-021-01171-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Hua LL, Lee J, Rahmandar MH, Sigel EJ. Suicide and suicide risk in adolescents. Pediatrics. 2024;153:e2023064800. doi: 10.1542/peds.2023-064800 [DOI] [PubMed] [Google Scholar]
6.Prades-Caballero V, Navarro-Pérez J-J, Carbonell Á. Factors Associated with Suicidal Behavior in Adolescents: An Umbrella Review Using the Socio-Ecological Model. Community Ment Health J. 2025;61(4):612–28. doi: 10.1007/s10597-024-01368-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Richardson R, Connell T, Foster M, Blamires J, Keshoor S, Moir C, et al. Risk and Protective Factors of Self-harm and Suicidality in Adolescents: An Umbrella Review with Meta-Analysis. J Youth Adolesc. 2024;53(6):1301–22. doi: 10.1007/s10964-024-01969-w [DOI] [PMC free article] [PubMed] [Google Scholar]
8.Zhan Y, Wang P, Zhan Y, Lu Z, Guo Y, Ahmad NA, et al. Clustering of lifestyle risk factors in relation to suicidal thoughts and behaviors in young adolescents: a cross-national study of 45 low- and middle-income countries. BMC Glob Public Health. 2024;2(1):24. doi: 10.1186/s44263-024-00055-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Ehtemam H, Sadeghi Esfahlani S, Sanaei A, Ghaemi MM, Hajesmaeel-Gohari S, Rahimisadegh R, et al. Role of machine learning algorithms in suicide risk prediction: a systematic review-meta analysis of clinical studies. BMC Med Inform Decis Mak. 2024;24(1):138. doi: 10.1186/s12911-024-02524-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Somé NH, Noormohammadpour P, Lange S. The use of machine learning on administrative and survey data to predict suicidal thoughts and behaviors: a systematic review. Front Psychiatry. 2024;15:1291362. doi: 10.3389/fpsyt.2024.1291362 [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Kim Y, Choi S, Chun C, Park S, Khang Y-H, Oh K. Data Resource Profile: The Korea Youth Risk Behavior Web-based Survey (KYRBS). Int J Epidemiol. 2016;45(4):1076–1076e. doi: 10.1093/ije/dyw070 [DOI] [PubMed] [Google Scholar]
12.Gignac GE, Szodorai ET. Effect size guidelines for individual differences researchers. Pers Individ Dif. 2016;102:74–8. doi: 10.1016/j.paid.2016.06.069 [DOI] [Google Scholar]
13.Lovakov A, Agadullina ER. Empirically derived guidelines for effect size interpretation in social psychology. Euro J Social Psych. 2021;51(3):485–504. doi: 10.1002/ejsp.2752 [DOI] [Google Scholar]
14.Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321–57. doi: 10.1613/jair.953 [DOI] [Google Scholar]
15.Park C, Mishra R, Golledge J, Najafi B. Digital Biomarkers of Physical Frailty and Frailty Phenotypes Using Sensor-Based Physical Activity and Machine Learning. Sensors (Basel). 2021;21(16):5289. doi: 10.3390/s21165289 [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Park C, Mishra R, Sharafkhaneh A, Bryant MS, Nguyen C, Torres I, et al. Digital Biomarker Representing Frailty Phenotypes: The Use of Machine Learning and Sensor-Based Sit-to-Stand Test. Sensors (Basel). 2021;21(9):3258. doi: 10.3390/s21093258 [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Efron B, Tibshirani RJ. An introduction to the bootstrap. Chapman & Hall/CRC. 1994. doi: 10.1201/9780429246593 [DOI] [Google Scholar]
18.Varma S, Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics. 2006;7:91. doi: 10.1186/1471-2105-7-91 [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Cohen J. Statistical power analysis for the behavioral sciences. Routledge. 2013. doi: 10.4324/9780203771587 [DOI] [Google Scholar]
20.Altmann A, Toloşi L, Sander O, Lengauer T. Permutation importance: a corrected feature importance measure. Bioinformatics. 2010;26(10):1340–7. doi: 10.1093/bioinformatics/btq134 [DOI] [PubMed] [Google Scholar]
21.Bradley AP. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognition. 1997;30(7):1145–59. doi: 10.1016/s0031-3203(96)00142-2 [DOI] [Google Scholar]
22.Mandrekar JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol. 2010;5(9):1315–6. doi: 10.1097/JTO.0b013e3181ec173d [DOI] [PubMed] [Google Scholar]
23.Heckler WF, de Carvalho JV, Barbosa JLV. Machine learning for suicidal ideation identification: a systematic literature review. Comput Hum Behav. 2022;128:107095. doi: 10.1016/j.chb.2021.107095 [DOI] [Google Scholar]
24.Motillon-Toudic C, Walter M, Séguin M, Carrier J-D, Berrouiguet S, Lemey C. Social isolation and suicide risk: Literature review and perspectives. Eur Psychiatry. 2022;65(1):e65. doi: 10.1192/j.eurpsy.2022.2320 [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Uno A, Nagaoka D, Usami S, Yamaguchi S, Minami R, Tanaka R, et al. Suicidal Thoughts and Trajectories of Psychopathological and Behavioral Symptoms in Adolescence. JAMA Netw Open. 2024;7(1):e2353166. doi: 10.1001/jamanetworkopen.2023.53166 [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Xu S, Ju Y, Wei X, Ou W, Ma M, Lv G, et al. Network analysis of suicide ideation and depression-anxiety symptoms among Chinese adolescents. Gen Psychiatr. 2024;37(2):e101225. doi: 10.1136/gpsych-2023-101225 [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Miranda-Mendizabal A, Castellví P, Parés-Badell O, Alayo I, Almenara J, Alonso I, et al. Gender differences in suicidal behavior in adolescents and young adults: systematic review and meta-analysis of longitudinal studies. Int J Public Health. 2019;64(2):265–83. doi: 10.1007/s00038-018-1196-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Albert PR. Why is depression more prevalent in women?. J Psychiatry Neurosci. 2015;40:219–21. doi: 10.1503/jpn.150205 [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Jeon G-S, Ha Y, Choi E. Effects of Objective and Subjective Socioeconomic Status on Self-Rated Health, Depressive Symptoms, and Suicidal Ideation in Adolescents. Child Ind Res. 2013;6(3):479–92. doi: 10.1007/s12187-013-9180-z [DOI] [Google Scholar]
30.Joseph VA, Kreski NT, Keyes KM. Sleep deprivation and suicide risk among minoritized US adolescents. BMC Psychiatry. 2023;23(1):638. doi: 10.1186/s12888-023-05074-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Gowin JL, Stoddard J, Doykos TK, Sammel MD, Bernert RA. Sleep Disturbance and Subsequent Suicidal Behaviors in Preadolescence. JAMA Netw Open. 2024;7(9):e2433734. doi: 10.1001/jamanetworkopen.2024.33734 [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Cen Y, He J, Zhong Y, Zhou J, Zeng J, Huang G, et al. The mediating role of sleep problems and depressed mood between psychological abuse/neglect and suicidal ideation in adolescent childhood: a multicentred, large sample survey in Western China. BMC Psychiatry. 2024;24(1):64. doi: 10.1186/s12888-024-05503-x [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Hartsock MJ, Strnad HK, Spencer RL. Iterative Metaplasticity Across Timescales: How Circadian, Ultradian, and Infradian Rhythms Modulate Memory Mechanisms. J Biol Rhythms. 2022;37(1):29–42. doi: 10.1177/07487304211058256 [DOI] [PMC free article] [PubMed] [Google Scholar]
34.Nkansah-Amankra S, Walker AD, Clark A. Suicidal Behaviors, Self Rated Health and Multiple Health Risk Behaviors Among Adolescents: Exploring New Perspectives With High School Students in Suicide Prevention Research. Annals of Epidemiology. 2009;19(9):669–70. doi: 10.1016/j.annepidem.2009.07.031 [DOI] [Google Scholar]
35.Nordin N, Zainol Z, Mohd Noor MH, Lai Fong C. A comparative study of machine learning techniques for suicide attempts predictive model. Health Informatics J. 2021;27(1):1460458221989395. doi: 10.1177/1460458221989395 [DOI] [PubMed] [Google Scholar]
36.Staartjes VE, Kernbach JM, Stumpo V, van Niftrik CHB, Serra C, Regli L. Foundations of Feature Selection in Clinical Prediction Modeling. Acta Neurochir Suppl. 2022;134:51–7. doi: 10.1007/978-3-030-85292-4_7 [DOI] [PubMed] [Google Scholar]
37.Roy A, Nikolitch K, McGinn R, Jinah S, Klement W, Kaminsky ZA. A machine learning approach predicts future risk to suicidal ideation from social media data. NPJ Digit Med. 2020;3:78. doi: 10.1038/s41746-020-0287-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Walsh CG, Ribeiro JD, Franklin JC. Predicting risk of suicide attempts over time through machine learning. Clin Psychol Sci. 2017;5:457–69. doi: 10.1177/2167702617691560 [DOI] [Google Scholar]

PLoS One. doi: 10.1371/journal.pone.0346050.r001

Decision Letter 0

Vincenzo De Luca

11 Nov 2025

Dear Dr. Lee,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Dec 26 2025 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

• A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

• A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

• An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at . Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at . Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at . Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols....

We look forward to receiving your revised manuscript.

Kind regards,

Vincenzo De Luca

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf....

2. Please note that PLOS One has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, we expect all author-generated code to be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse.

3. Thank you for stating in your Funding Statement:

“This work was supported by the MSIT (Ministry of Science and ICT), Korea, under the National Program in Medical AI Semiconductor supervised by the IITP (Institute of Information & Communications Technology Planning & Evaluation) [grant number 2024-0-0096]. The funding sources had no role in the study design, methods, data collection and analysis, and submission of the results.”

Please provide an amended statement that declares *all* the funding or sources of support (whether external or internal to your organization) received during this study, as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now. Please also include the statement “There was no additional external funding received for this study.” in your updated Funding Statement.

Please include your amended Funding Statement within your cover letter. We will change the online submission form on your behalf.

4. Please be informed that funding information should not appear in the Acknowledgments section or other areas of your manuscript. We will only publish funding information present in the Funding Statement section of the online submission form. Please remove any funding-related text from the manuscript.

5. Your ethics statement should only appear in the Methods section of your manuscript. If your ethics statement is written in any section besides the Methods, please move it to the Methods section and delete it from any other section. Please ensure that your ethics statement is included in your manuscript, as the ethics statement entered into the online submission form will not be published alongside your manuscript.

6. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #1: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #1: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.-->

Reviewer #1: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #1: Yes

**********

Reviewer #1: The rationale for this study was the fact that “suicide is the 3rd leading cause of death for adolescents world-wide.” It used data from the Korea Youth Risk Behavior Web-based Survey (2022–2023, Korea Disease Control and Prevention Agency, KDCA), including 90,217 adolescents aged 12–18 years. Suicidal: yes=11,838, no=78,975 (6.67%); attempts 2,191.

Data were analyzed via machine learning (Random Forest, RF) model, which they note is superior to meta-analysis. They used multidimensional risk factors (psychological, sociological, physical illness, demographic ) spanning sociodemographic, physical and mental health, and behavioral domains. Sensitivity and specificity were assessed using the AUC method. “Sadness, loneliness, anxiety, and stress were identified as optimal features, achieving strong results in identifying suicidal ideation and predicting suicide attempts” == AUC 92%, Sens 94%, Spec 90%.

Cramer’s V was used to define effect size of categoricals, Cohen’s-d for effect size of continuous variables.

[how was suicidal vs. non-suicidal assessed?]

Of note, statistical risk factors by effect size were: Anxiety > Loneliness > Stress > Sleep > Sadness. These all had AUC >78%. While alcohol and smoking had very low effect size. 8 factors (sex, family economic status, sleep quality, self-reported health status, sadness, stress, anxiety, and loneliness) were statistically significant (p < 0.05) and relevant to clinical practice. Five statistical models comprised various combinations of these factors, with the most predictive (5) having sens=93.5%, spec= 90.3%

One wonders whether physical or sexual abuse was asked about in this large survey, and if so, what effect it would have on the results. Otherwise, the stated limitations are reasonable.

Fig. 3 displays factors that were internal to the machine learning method but are presented in a way that is unfamiliar to readers accustomed to reading AUC curves. I suggest that the authors either add a traditional AUC curve (Sensitivity vs. 1 – Specificity) for each of the 8 factors in their analysis, with all 8 AUCs overlaid on the same graph, or replace Fig. 3 with the latter AUC curve.

Overall, this paper is an excellent guide for those working with suicidal adolescents

**********

what does this mean?). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our For information about this choice, including consent withdrawal, please see our For information about this choice, including consent withdrawal, please see our For information about this choice, including consent withdrawal, please see our Privacy Policy..-->

Reviewer #1: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures

You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation.

NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications.

PLoS One. 2026 Apr 6;21(4):e0346050. doi: 10.1371/journal.pone.0346050.r002

Author response to Decision Letter 1

17 Nov 2025

Reviewer 1’s comments

The rationale for this study was the fact that “suicide is the 3rd leading cause of death for adolescents world-wide.” It used data from the Korea Youth Risk Behavior Web-based Survey (2022–2023, Korea Disease Control and Prevention Agency, KDCA), including 90,217 adolescents aged 12–18 years. Suicidal: yes=11,838, no=78,975 (6.67%); attempts 2,191.

Data were analyzed via machine learning (Random Forest, RF) model, which they note is superior to meta-analysis. They used multidimensional risk factors (psychological, sociological, physical illness, demographic) spanning sociodemographic, physical and mental health, and behavioral domains. Sensitivity and specificity were assessed using the AUC method. “Sadness, loneliness, anxiety, and stress were identified as optimal features, achieving strong results in identifying suicidal ideation and predicting suicide attempts” == AUC 92%, Sens 94%, Spec 90%.

Cramer’s V was used to define effect size of categoricals, Cohen’s-d for effect size of continuous variables.

[how was suicidal vs. non-suicidal assessed?]

One wonders whether physical or sexual abuse was asked about in this large survey, and if so, what effect it would have on the results. Otherwise, the stated limitations are reasonable.

Overall, this paper is an excellent guide for those working with suicidal adolescents

Response:

We sincerely appreciate the reviewer’s thorough evaluation and constructive feedback. Below, we have restated each comment (in bold) and provided detailed responses (in regular text). All corresponding revisions in the manuscript have been incorporated and are highlighted in grey.

Comments:

1. How was suicidal vs. non-suicidal assessed?

Response:

We thank the reviewer for pointing out the need for clearer clarification of the classification criteria.

In the revised manuscript, we have now explicitly described how suicidal ideation and suicide attempts were defined based on the KYRBS survey questions. Specifically, suicidal ideation and non-suicidal groups were classified according to responses to the following KYRBS items (e.g., During the past 12 months, have you ever seriously thought about committing suicide?; During the past 12 months, have you ever made a specific plan to commit suicide?; and During the past 12 months, have you ever attempted suicide?).

Participants who responded “Yes” to either suicidal thoughts or suicide planning were categorized into the suicidal ideation group, whereas those who answered “No” to both questions were classified as the non-suicidal group. For the analysis of suicide attempts, participants who responded “Yes” to the suicide attempt question were classified into the suicide attempt group, and those who answered “No” were classified as non-attempt. This clarification has been added to the Data Source, Participant Selection, and Multidimensional Risk Factors section.

2. One wonders whether physical or sexual abuse was asked about in this large survey, and if so, what effect it would have on the results.

Response:

We examined the available variables from the KYRBS dataset and found no items specifically assessing physical or sexual abuse, such as sexual victimisation or non-partner violence. Because these trauma-related exposures are known risk factors for suicidal ideation and suicide attempts, the absence of such items in the dataset may limit the comprehensiveness of our risk-factor assessment. We have therefore added a statement in the Limitations section to that effect and recommend that future research incorporate datasets which include trauma or abuse-related items.

3. Fig. 3 displays factors that were internal to the machine learning method but are presented in a way that is unfamiliar to readers accustomed to reading AUC curves. I suggest that the authors either add a traditional AUC curve (Sensitivity vs. 1 – Specificity) for each of the 8 factors in their analysis, with all 8 AUCs overlaid on the same graph, or replace Fig. 3 with the latter AUC curve.

Response:

We greatly appreciate this valuable suggestion. Following the reviewer’s recommendation, we have revised Figure 3 to include a traditional receiver operating characteristic (ROC) curve presentation. The updated figure now shows the overlaid ROC curves (Sensitivity vs. 1 – Specificity) for each model. We have also updated the corresponding Results subsection and Figure 3 legend to reflect these changes.

Attachment

Submitted filename: PONE-D-25-47059_Revision 1_Response to Reviewers.pdf

pone.0346050.s002.pdf^{(75.8KB, pdf)}

PLoS One. doi: 10.1371/journal.pone.0346050.r003

Decision Letter 1

Vincenzo De Luca

23 Feb 2026

Dear Dr. Lee,

Please submit your revised manuscript by Apr 09 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

• A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

• A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

• An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

We look forward to receiving your revised manuscript.

Kind regards,

Vincenzo De Luca

Academic Editor

PLOS One

Journal Requirements:

If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

Reviewer #3: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions??>

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Partly

**********

3. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: No

**********

4. Have the authors made all data underlying the findings in their manuscript fully available??>

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

Reviewer #1: All of my questions were answered to my satisfaction. Especially, the multiple AUC graph is now fully understandable, and quite dramatic in its accuracy.

Reviewer #2: The manuscript addresses an important and timely public health issue, namely adolescent suicidal ideation and suicide attempts, and applies a machine learning-optimized approach to identify minimal risk factors. The topic is highly relevant, and the overall structure of the paper is clear and logically organized.

The objectives of the study are well stated, and the use of machine learning methods appears appropriate for the research question. The manuscript is generally well written and understandable, and the results are presented in a coherent manner.

I did not identify any major concerns related to the overall approach, ethical considerations, or presentation that would preclude publication. Minor improvements may be possible in terms of clarity or detail in some sections, but these do not affect the main conclusions of the study.

Overall, the manuscript makes a useful contribution to the literature on adolescent mental health and suicide prevention.

Reviewer #3: The manuscript presents an original investigation aiming to develop and validate a machine learning (ML) model to identify minimal risk factors associated with adolescent suicidal ideation and suicide attempts. Data from a large nationwide online survey (Korea Youth Risk Behavior Web-based Survey, 2022–2023) were used for the analyses. The survey assesses four domains of risk factors: sociodemographic, physical health, mental health, and behavioral domains. The results suggest that the ML model performed well in identifying optimal features associated with suicidal ideation. Subsequently, these selected optimal features, along with additional factors, were used to assess the prediction of suicide attempts. The authors conclude that ML-driven approaches may have potential clinical usefulness in improving adolescent mental health outcomes.

Given the high public health importance of suicide in adolescent populations, the topic is highly relevant. The English language is clear overall, and the structure of the manuscript is appropriate.

However, several methodological questions arose during the review of the manuscript. The comments below are intended as constructive feedback to help strengthen this important and valuable work.

1. Sample characteristics and reporting

Questions arose regarding the analysed sample:

1. The total number of analysed subjects differs across sections of the manuscript (Abstract, ‘Data source, participant selection, and multidimensional risk factors’ section in lines 96–111, and Figure 1; N=90,813 vs. N=90,217). This discrepancy should be clarified.

2. It remains unclear whether respondents could overlap between the 2022 and 2023 KYRBS cohorts (e.g., whether the same individual may have participated in both survey years).

3. Lines 109–111: “For the analysis of suicide attempts, participants were further classified according to their responses to the question asking whether they had ever attempted suicide during the past 12 months (yes/no).”

The total number of suicide attempts (n = 2,191) should also be reported here for clarity (not only later in lines 194–195). Additionally, it would be helpful to report the degree of overlap between this subgroup and the suicidal ideation group.

2. Terminology

I would suggest reconsidering the use of the umbrella term “suicidal behaviors” to describe suicidal ideation and suicide attempts together (e.g., lines 331–332). As the manuscript highlights potential clinical applications of ML-driven approaches, it is important to note that suicidal ideation and suicide attempts are considered distinct constructs with distinct predictors in clinical practice (for example see Klonsky & May, 2013), although they may share some predictors. Clear differentiation may improve conceptual clarity.

3. Methodological consideration regarding feature selection

This conceptual distinction also raises a methodological question: If suicidal ideation and suicide attempts have (partially) different predictors, it may be theoretically problematic to use optimal features derived from a model trained on suicidal ideation to predict suicide attempts. To my knowledge, feature importance in machine learning models is inherently outcome-specific. Therefore, features selected using a model optimized for suicidal ideation may not generalize to suicide attempt prediction. Re-deriving the predictive model using suicide attempt as the target outcome could strengthen the methodological validity of the analysis.

4. Clinical interpretability of model performance

Regarding potential clinical usefulness, the manuscript demonstrates strong model performance primarily using AUC, sensitivity, and specificity metrics. However, for clinical applicability, additional performance measures may be more informative, particularly positive predictive value (PPV) and negative predictive value (NPV). This is especially relevant given the relatively low prevalence of suicide attempts in both the general adolescent population and the analysed dataset.

Overall recommendation

Overall, the study has the potential to make a meaningful contribution to the scientific literature. However, in my opinion, the manuscript would benefit from further methodological clarification and potential additional analyses addressing the concerns outlined above before reconsideration for publication.

**********

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

**********

To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures

You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation.

NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications.

PLoS One. 2026 Apr 6;21(4):e0346050. doi: 10.1371/journal.pone.0346050.r004

Author response to Decision Letter 2

26 Feb 2026

Reviewer 1’s comments

All of my questions were answered to my satisfaction. Especially, the multiple AUC graph is now fully understandable, and quite dramatic in its accuracy.

Response:

Thank you for your positive and constructive feedback. We are pleased to hear that all of your questions were addressed to your satisfaction. In particular, we sincerely appreciate your comment that the multiple AUC figure is now fully understandable and that the accuracy is presented in a compelling manner. We have retained the revised figure and its accompanying explanation to ensure clarity and transparency for readers in the final manuscript.

Reviewer 2’s comments

The manuscript addresses an important and timely public health issue, namely adolescent suicidal ideation and suicide attempts, and applies a machine learning-optimized approach to identify minimal risk factors. The topic is highly relevant, and the overall structure of the paper is clear and logically organized.

Overall, the manuscript makes a useful contribution to the literature on adolescent mental health and suicide prevention.

Response:

Thank you for your thoughtful and encouraging comments on the significance, clarity, and overall rigor of our manuscript. We appreciate your assessment that the objectives are well stated, the machine learning approach is appropriate, and the results are coherently presented. Following your suggestion and incorporating the other reviewers’ comments and suggestions, we revised the manuscript to improve clarity and add methodological and reporting details where needed, without altering the main conclusions. We are grateful for your consideration of our work.

Reviewer 3’s comments

The manuscript presents an original investigation aiming to develop and validate a machine learning (ML) model to identify minimal risk factors associated with adolescent suicidal ideation and suicide attempts. Data from a large nationwide online survey (Korea Youth Risk Behavior Web-based Survey, 2022–2023) were used for the analyses. The survey assesses four domains of risk factors: sociodemographic, physical health, mental health, and behavioral domains. The results suggest that the ML model performed well in identifying optimal features associated with suicidal ideation. Subsequently, these selected optimal features, along with additional factors, were used to assess the prediction of suicide attempts. The authors conclude that ML-driven approaches may have potential clinical usefulness in improving adolescent mental health outcomes. Given the high public health importance of suicide in adolescent populations, the topic is highly relevant. The English language is clear overall, and the structure of the manuscript is appropriate. However, several methodological questions arose during the review of the manuscript. The comments below are intended as constructive feedback to help strengthen this important and valuable work.

Specific Comments:

1. Sample characteristics and reporting

Questions arose regarding the analysed sample:

1-1. The total number of analysed subjects differs across sections of the manuscript (Abstract, ‘Data source, participant selection, and multidimensional risk factors’ section in lines 96–111, and Figure 1; N=90,813 vs. N=90,217). This discrepancy should be clarified

Response:

Thank you for noting this discrepancy. We confirm that N = 90,813 (as shown in Figure 1) is the correct total, and the inconsistency was due to a manual transcription error during manuscript preparation. We have revised the Abstract and the “Data source, participant selection, and multidimensional risk factors” section to consistently reflect N = 90,813 throughout the manuscript. In addition to this correction, all other revisions addressing your comments have been highlighted in yellow for your convenience.

1-2. It remains unclear whether respondents could overlap between the 2022 and 2023 KYRBS cohorts (e.g., whether the same individual may have participated in both survey years).

Response:

We thank the reviewer for pointing this out. For clarification, we contacted the KYRBS administrators directly and confirmed that the 2022 and 2023 KYRBS were conducted as independent, nationally representative cross-sectional surveys (i.e., point-in-time assessments of the population with no longitudinal follow-up) using random sampling of schools and students in South Korea. Based on this survey design, they confirmed that there was no respondent overlap between the 2022 and 2023 survey years. We have added this clarification to the manuscript in the “Data source, participant selection, and multidimensional risk factors” section as follows:

“The 2022 and 2023 KYRBS were conducted as independent, nationally representative cross-sectional surveys using random sampling of schools and students in South Korea; based on direct confirmation from the KYRBS administrators, there was no respondent overlap between the two survey years.”

1-3. Lines 109–111: “For the analysis of suicide attempts, participants were further classified according to their responses to the question asking whether they had ever attempted suicide during the past 12 months (yes/no).” The total number of suicide attempts (n = 2,191) should also be reported here for clarity (not only later in lines 194–195). Additionally, it would be helpful to report the degree of overlap between this subgroup and the suicidal ideation group.

Response:

We appreciate your suggestion. We have revised the text in the “Data source, participant selection, and multidimensional risk factors” section to explicitly report the total number of suicide attempts (n = 2,191) for clarity. In addition, consistent with our response above, we contacted the KYRBS administrators directly and confirmed that there is no participant overlap between the suicidal ideation group and the suicide attempt subgroup, and we have added this clarification to the manuscript. Specifically, we revised the sentence as follows:

“For the analysis of suicide attempts, participants were further classified according to their responses to the question asking whether they had ever attempted suicide during the past 12 months (yes/no), resulting in 2,191 participants in the suicide attempt group; additionally, based on direct confirmation from the KYRBS administrators, there was no participant overlap between the suicidal ideation group and the suicide attempt subgroup.”

2. Terminology: I would suggest reconsidering the use of the umbrella term “suicidal behaviors” to describe suicidal ideation and suicide attempts together (e.g., lines 331–332). As the manuscript highlights potential clinical applications of ML-driven approaches, it is important to note that suicidal ideation and suicide attempts are considered distinct constructs with distinct predictors in clinical practice (for example see Klonsky & May, 2013), although they may share some predictors. Clear differentiation may improve conceptual clarity.

Response:

Thank you for this important conceptual clarification. We agree that suicidal ideation and suicide attempts are related but clinically distinct outcomes that may involve partially different predictors. In response, we revised the manuscript to improve conceptual clarity by explicitly distinguishing these two outcomes in the Discussion and by limiting the use of the term “suicidal behaviors” to a broader descriptive context in the Conclusion, where it is clearly specified as encompassing both suicidal ideation and suicide attempts. We believe this revision better aligns the manuscript with clinical terminology and interpretation.

3. Methodological consideration regarding feature selection: This conceptual distinction also raises a methodological question: If suicidal ideation and suicide attempts have (partially) different predictors, it may be theoretically problematic to use optimal features derived from a model trained on suicidal ideation to predict suicide attempts. To my knowledge, feature importance in machine learning models is inherently outcome-specific. Therefore, features selected using a model optimized for suicidal ideation may not generalize to suicide attempt prediction. Re-deriving the predictive model using suicide attempt as the target outcome could strengthen the methodological validity of the analysis.

Response:

Thank you for this important methodological comment. We agree that suicidal ideation and suicide attempts are distinct outcomes, but we respectfully note that our study was not designed to identify an attempt-specific feature set via outcome-specific feature selection. Rather, the explicit aim of the study was to identify a minimal, clinically actionable core set of risk factors and then test whether this minimally sufficient set could support prediction across both suicidal ideation and suicide attempts. This is also the framework described in the Methods, where optimal features were first identified for suicidal ideation and subsequently evaluated for suicide-attempt prediction. Importantly, the results, reported in Table 3, empirically support this choice: the ideation-derived four-feature model already showed meaningful performance for suicide attempts, and performance improved progressively as additional factors were added in Models 2–5. Accordingly, re-deriving a separate attempt-specific model would address a different objective from that of the present study, rather than correct a flaw in the current design.

4. Clinical interpretability of model performance: Regarding potential clinical usefulness, the manuscript demonstrates strong model performance primarily using AUC, sensitivity, and specificity metrics. However, for clinical applicability, additional performance measures may be more informative, particularly positive predictive value (PPV) and negative predictive value (NPV). This is especially relevant given the relatively low prevalence of suicide attempts in both the general adolescent population and the analysed dataset.

Response:

Thank you for this important suggestion. We agree that positive predictive value (PPV) and negative predictive value (NPV) are important additional metrics for evaluating the potential clinical applicability of our models, particularly given the relatively low prevalence of suicide attempts in the analyzed dataset. In response, we reran the machine learning analyses and updated Figure 1, Tables 2 and 3, as well as the corresponding results reported in the Results section and the Abstract. It is worth noting that, because the updated analyses were conducted using the same overall framework as in the previous version, including SMOTE and bootstrapping, some performance values in the tables changed slightly; however, these changes did not lead to any significant differences in the main findings or conclusions of our Revision 1 manuscript.

Overall recommendation: Overall, the study has the potential to make a meaningful contribution to the scientific literature. However, in my opinion, the manuscript would benefit from further methodological clarification and potential additional analyses addressing the concerns outlined above before reconsideration for publication.

Response:

Thank you for your thoughtful and constructive overall assessment. We sincerely appreciate your recognition of the potential contribution of this study to the literature on adolescent mental health and suicide prevention. In response to your comments, we have carefully revised the manuscript to provide further methodological clarification and additional detail where needed, and we also conducted additional analyses to strengthen the clinical interpretability of the findings. We believe that these revisions have substantially improved the clarity, rigor, and transparency of the manuscript, while the main findings and conclusions remain unchanged.

Attachment

Submitted filename: PONE-D-25-47059_Revision 2_Response to Reviewers.pdf

pone.0346050.s003.pdf^{(87.9KB, pdf)}

PLoS One. doi: 10.1371/journal.pone.0346050.r005

Decision Letter 2

Vincenzo De Luca

16 Mar 2026

Identifying minimal risk factors for adolescent suicidal ideation and suicide attempts: A machine learning-optimized approach

PONE-D-25-47059R2

Dear Dr. Lee,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support....

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Vincenzo De Luca

Academic Editor

PLOS One

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

Reviewer #3: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions??>

Reviewer #3: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #3: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available??>

Reviewer #3: Yes

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #3: Yes

**********

Reviewer #3: All of my previous questions have been satisfactorily addressed.

In the current version of the manuscript, the sample size is consistently reported across the different sections of the text. The concern regarding a potential overlap between the investigated cohorts from 2022 and 2023 has been clarified. Similarly, the possible overlap between the suicide attempt and suicidal ideation groups. Consequently, the revised manuscript provides a clearer description of the study cohort.

The theoretical issue raised during the previous review round has also been clarified: the Authors explain that the KYRBS survey was not originally designed to differentiate between suicide attempts and suicidal ideation. In addition, the terminology in the text has been improved: the umbrella term “suicidal behaviors” used in the previous manuscript version has been revised.

The methodological question raised in the previous review round has been satisfactorily addressed. The positive predictive value (PPV) and negative predictive value (NPV) have now been incorporated into the presentation of the results.

Overall, the revised manuscript makes a valuable contribution to the scientific literature

**********

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Reviewer #3: No

**********

PLoS One. doi: 10.1371/journal.pone.0346050.r006

Acceptance letter

Vincenzo De Luca

PONE-D-25-47059R2

PLOS One

Dear Dr. Lee,

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team.

At this stage, our production department will prepare your paper for publication. This includes ensuring the following:

* All references, tables, and figures are properly cited

* All relevant supporting information is included in the manuscript submission,

* There are no issues that prevent the paper from being properly typeset

You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps.

Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing.

If we can help with anything else, please email us at customercare@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Vincenzo De Luca

Academic Editor

PLOS One

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Attachment

Submitted filename: PONE-D-25-47059_Revision 1_Response to Reviewers.pdf

pone.0346050.s002.pdf^{(75.8KB, pdf)}

Attachment

Submitted filename: PONE-D-25-47059_Revision 2_Response to Reviewers.pdf

pone.0346050.s003.pdf^{(87.9KB, pdf)}

Data Availability Statement

[pone.0346050.ref001] 1.WHO. Suicide. World Health Organization. https://www.who.int/news-room/fact-sheets/detail/suicide. 2024. [Google Scholar]

[pone.0346050.ref002] 2.Yan N, Luo Y, Mackay LE, Wang Y, Wang Y, Wang Y, et al. Global patterns and trends of suicide mortality and years of life lost among adolescents and young adults from 1990 to 2021: a systematic analysis for the Global Burden of Disease Study 2021. Epidemiol Psychiatr Sci. 2024;33:e52. doi: 10.1017/S2045796024000532 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref003] 3.Campisi SC, Carducci B, Akseer N, Zasowski C, Szatmari P, Bhutta ZA. Suicidal behaviours among adolescents from 90 countries: a pooled analysis of the global school-based student health survey. BMC Public Health. 2020;20(1):1102. doi: 10.1186/s12889-020-09209-z [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref004] 4.Ho TC, Gifuni AJ, Gotlib IH. Psychobiological risk factors for suicidal thoughts and behaviors in adolescence: a consideration of the role of puberty. Mol Psychiatry. 2022;27(1):606–23. doi: 10.1038/s41380-021-01171-5 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref005] 5.Hua LL, Lee J, Rahmandar MH, Sigel EJ. Suicide and suicide risk in adolescents. Pediatrics. 2024;153:e2023064800. doi: 10.1542/peds.2023-064800 [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref006] 6.Prades-Caballero V, Navarro-Pérez J-J, Carbonell Á. Factors Associated with Suicidal Behavior in Adolescents: An Umbrella Review Using the Socio-Ecological Model. Community Ment Health J. 2025;61(4):612–28. doi: 10.1007/s10597-024-01368-2 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref007] 7.Richardson R, Connell T, Foster M, Blamires J, Keshoor S, Moir C, et al. Risk and Protective Factors of Self-harm and Suicidality in Adolescents: An Umbrella Review with Meta-Analysis. J Youth Adolesc. 2024;53(6):1301–22. doi: 10.1007/s10964-024-01969-w [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref008] 8.Zhan Y, Wang P, Zhan Y, Lu Z, Guo Y, Ahmad NA, et al. Clustering of lifestyle risk factors in relation to suicidal thoughts and behaviors in young adolescents: a cross-national study of 45 low- and middle-income countries. BMC Glob Public Health. 2024;2(1):24. doi: 10.1186/s44263-024-00055-4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref009] 9.Ehtemam H, Sadeghi Esfahlani S, Sanaei A, Ghaemi MM, Hajesmaeel-Gohari S, Rahimisadegh R, et al. Role of machine learning algorithms in suicide risk prediction: a systematic review-meta analysis of clinical studies. BMC Med Inform Decis Mak. 2024;24(1):138. doi: 10.1186/s12911-024-02524-0 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref010] 10.Somé NH, Noormohammadpour P, Lange S. The use of machine learning on administrative and survey data to predict suicidal thoughts and behaviors: a systematic review. Front Psychiatry. 2024;15:1291362. doi: 10.3389/fpsyt.2024.1291362 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref011] 11.Kim Y, Choi S, Chun C, Park S, Khang Y-H, Oh K. Data Resource Profile: The Korea Youth Risk Behavior Web-based Survey (KYRBS). Int J Epidemiol. 2016;45(4):1076–1076e. doi: 10.1093/ije/dyw070 [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref012] 12.Gignac GE, Szodorai ET. Effect size guidelines for individual differences researchers. Pers Individ Dif. 2016;102:74–8. doi: 10.1016/j.paid.2016.06.069 [DOI] [Google Scholar]

[pone.0346050.ref013] 13.Lovakov A, Agadullina ER. Empirically derived guidelines for effect size interpretation in social psychology. Euro J Social Psych. 2021;51(3):485–504. doi: 10.1002/ejsp.2752 [DOI] [Google Scholar]

[pone.0346050.ref014] 14.Chawla NV, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic minority over-sampling technique. J Artif Intell Res. 2002;16:321–57. doi: 10.1613/jair.953 [DOI] [Google Scholar]

[pone.0346050.ref015] 15.Park C, Mishra R, Golledge J, Najafi B. Digital Biomarkers of Physical Frailty and Frailty Phenotypes Using Sensor-Based Physical Activity and Machine Learning. Sensors (Basel). 2021;21(16):5289. doi: 10.3390/s21165289 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref016] 16.Park C, Mishra R, Sharafkhaneh A, Bryant MS, Nguyen C, Torres I, et al. Digital Biomarker Representing Frailty Phenotypes: The Use of Machine Learning and Sensor-Based Sit-to-Stand Test. Sensors (Basel). 2021;21(9):3258. doi: 10.3390/s21093258 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref017] 17.Efron B, Tibshirani RJ. An introduction to the bootstrap. Chapman & Hall/CRC. 1994. doi: 10.1201/9780429246593 [DOI] [Google Scholar]

[pone.0346050.ref018] 18.Varma S, Simon R. Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics. 2006;7:91. doi: 10.1186/1471-2105-7-91 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref019] 19.Cohen J. Statistical power analysis for the behavioral sciences. Routledge. 2013. doi: 10.4324/9780203771587 [DOI] [Google Scholar]

[pone.0346050.ref020] 20.Altmann A, Toloşi L, Sander O, Lengauer T. Permutation importance: a corrected feature importance measure. Bioinformatics. 2010;26(10):1340–7. doi: 10.1093/bioinformatics/btq134 [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref021] 21.Bradley AP. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pattern Recognition. 1997;30(7):1145–59. doi: 10.1016/s0031-3203(96)00142-2 [DOI] [Google Scholar]

[pone.0346050.ref022] 22.Mandrekar JN. Receiver operating characteristic curve in diagnostic test assessment. J Thorac Oncol. 2010;5(9):1315–6. doi: 10.1097/JTO.0b013e3181ec173d [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref023] 23.Heckler WF, de Carvalho JV, Barbosa JLV. Machine learning for suicidal ideation identification: a systematic literature review. Comput Hum Behav. 2022;128:107095. doi: 10.1016/j.chb.2021.107095 [DOI] [Google Scholar]

[pone.0346050.ref024] 24.Motillon-Toudic C, Walter M, Séguin M, Carrier J-D, Berrouiguet S, Lemey C. Social isolation and suicide risk: Literature review and perspectives. Eur Psychiatry. 2022;65(1):e65. doi: 10.1192/j.eurpsy.2022.2320 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref025] 25.Uno A, Nagaoka D, Usami S, Yamaguchi S, Minami R, Tanaka R, et al. Suicidal Thoughts and Trajectories of Psychopathological and Behavioral Symptoms in Adolescence. JAMA Netw Open. 2024;7(1):e2353166. doi: 10.1001/jamanetworkopen.2023.53166 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref026] 26.Xu S, Ju Y, Wei X, Ou W, Ma M, Lv G, et al. Network analysis of suicide ideation and depression-anxiety symptoms among Chinese adolescents. Gen Psychiatr. 2024;37(2):e101225. doi: 10.1136/gpsych-2023-101225 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref027] 27.Miranda-Mendizabal A, Castellví P, Parés-Badell O, Alayo I, Almenara J, Alonso I, et al. Gender differences in suicidal behavior in adolescents and young adults: systematic review and meta-analysis of longitudinal studies. Int J Public Health. 2019;64(2):265–83. doi: 10.1007/s00038-018-1196-1 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref028] 28.Albert PR. Why is depression more prevalent in women?. J Psychiatry Neurosci. 2015;40:219–21. doi: 10.1503/jpn.150205 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref029] 29.Jeon G-S, Ha Y, Choi E. Effects of Objective and Subjective Socioeconomic Status on Self-Rated Health, Depressive Symptoms, and Suicidal Ideation in Adolescents. Child Ind Res. 2013;6(3):479–92. doi: 10.1007/s12187-013-9180-z [DOI] [Google Scholar]

[pone.0346050.ref030] 30.Joseph VA, Kreski NT, Keyes KM. Sleep deprivation and suicide risk among minoritized US adolescents. BMC Psychiatry. 2023;23(1):638. doi: 10.1186/s12888-023-05074-3 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref031] 31.Gowin JL, Stoddard J, Doykos TK, Sammel MD, Bernert RA. Sleep Disturbance and Subsequent Suicidal Behaviors in Preadolescence. JAMA Netw Open. 2024;7(9):e2433734. doi: 10.1001/jamanetworkopen.2024.33734 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref032] 32.Cen Y, He J, Zhong Y, Zhou J, Zeng J, Huang G, et al. The mediating role of sleep problems and depressed mood between psychological abuse/neglect and suicidal ideation in adolescent childhood: a multicentred, large sample survey in Western China. BMC Psychiatry. 2024;24(1):64. doi: 10.1186/s12888-024-05503-x [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref033] 33.Hartsock MJ, Strnad HK, Spencer RL. Iterative Metaplasticity Across Timescales: How Circadian, Ultradian, and Infradian Rhythms Modulate Memory Mechanisms. J Biol Rhythms. 2022;37(1):29–42. doi: 10.1177/07487304211058256 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref034] 34.Nkansah-Amankra S, Walker AD, Clark A. Suicidal Behaviors, Self Rated Health and Multiple Health Risk Behaviors Among Adolescents: Exploring New Perspectives With High School Students in Suicide Prevention Research. Annals of Epidemiology. 2009;19(9):669–70. doi: 10.1016/j.annepidem.2009.07.031 [DOI] [Google Scholar]

[pone.0346050.ref035] 35.Nordin N, Zainol Z, Mohd Noor MH, Lai Fong C. A comparative study of machine learning techniques for suicide attempts predictive model. Health Informatics J. 2021;27(1):1460458221989395. doi: 10.1177/1460458221989395 [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref036] 36.Staartjes VE, Kernbach JM, Stumpo V, van Niftrik CHB, Serra C, Regli L. Foundations of Feature Selection in Clinical Prediction Modeling. Acta Neurochir Suppl. 2022;134:51–7. doi: 10.1007/978-3-030-85292-4_7 [DOI] [PubMed] [Google Scholar]

[pone.0346050.ref037] 37.Roy A, Nikolitch K, McGinn R, Jinah S, Klement W, Kaminsky ZA. A machine learning approach predicts future risk to suicidal ideation from social media data. NPJ Digit Med. 2020;3:78. doi: 10.1038/s41746-020-0287-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[pone.0346050.ref038] 38.Walsh CG, Ribeiro JD, Franklin JC. Predicting risk of suicide attempts over time through machine learning. Clin Psychol Sci. 2017;5:457–69. doi: 10.1177/2167702617691560 [DOI] [Google Scholar]

PERMALINK

Identifying minimal risk factors for adolescent suicidal ideation and suicide attempts: A machine learning-optimized approach

Catherine Park

Beom-Chan Lee

Roles

Abstract

Introduction

Methods

Data source, participant selection, and multidimensional risk factors

Fig 1. Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) diagram of eligibility determinations and group classification of participants.

Statistical analysis

Machine learning framework and optimal feature selection

Fig 2. The machine learning framework is under the curve (AUC), the area under the receiver operating characteristic curve.

Ethics statement

Results

Multidimensional risk factors contributing to suicidal ideation

Table 1. Multidimensional risk factors associated with suicidal ideation, including sociodemographic characteristics, physical health,mental health, and health-related behaviors.

Performance of the machine learning model and optimal feature selection

Table 2. Ranking of the eight factors included in the machine learning model and validation results of the model using 50 validation datasets.

Fig 3. (A) Performance metrics (accuracy, sensitivity, specificity, and AUC) plotted against the number of included features during model validation, with error bars representing 95% confidence intervals.

Predictive performance of machine learning model for suicide attempts

Table 3. Predictive performance of machine learning models for suicide attempts.

Discussion

Conclusion

Acknowledgments

Data Availability

Funding Statement

References

Decision Letter 0

Vincenzo De Luca

Roles

Author response to Decision Letter 1

Decision Letter 1

Vincenzo De Luca

Roles

Author response to Decision Letter 2

Decision Letter 2

Vincenzo De Luca

Roles

Acceptance letter

Vincenzo De Luca

Roles

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases