Skip to contents Skip to Global Navigation Menu
  • KDCA
  • Contact us
  • E-Submission

PHRP : Osong Public Health and Research Perspectives

OPEN ACCESS. pISSN: 2210-9099. eISSN: 2233-6052
Search
Advanced Search

Articles

Page Path

Original Article

A machine learning approach for predicting suicidal ideation among family members of persons with disabilities: a cross-sectional study in the Republic of Korea

Jin Hyuk Leeorcid
Osong Public Health and Research Perspectives 2025;16(6):560-574.
Published online: December 11, 2025

Department of Social Welfare, Baekseok Culture University, Cheonan, Republic of Korea

Corresponding author: Jin Hyuk Lee Department of Social Welfare, Baekseok Culture University, 1 Baekseokdaehak-ro, Dongnam-gu, Cheonan 31065, Republic of Korea E-mail: gene2you@gmail.com
• Received: July 8, 2025   • Revised: November 6, 2025   • Accepted: November 7, 2025

© 2025 Korea Disease Control and Prevention Agency.

This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

  • 1,212 Views
  • 79 Download
prev next

Full Article

Download PDF Download PDF
  • Objectives
    Although family members of persons with disabilities face elevated suicide risk, predictive models remain underdeveloped in Korean contexts. This study aimed to develop machine learning–based predictive models for suicidal ideation among family members of persons with disabilities and examine differential risk patterns by disability onset type.
  • Methods
    This cross-sectional study analyzed 124,783 adult family members (59.9% spouses, 20.3% parents/ascendants, 14.6% adult children, 5.2% extended family) from the 2018 Korean Disability and Life Dynamics Panel using survey weights. Four predictive models, including machine learning approaches, were compared using 31 variables. The dataset was divided into training (70%) and test (30%) sets, with stratified analyses comparing congenital and acquired disability groups.
  • Results
    Among the 124,783 family members analyzed, least absolute shrinkage and selection operator (LASSO) with cross-validation achieved optimal performance (area under the receiver operating characteristic curve, 0.875 training; 0.853 test). LASSO selected 16 of 31 variables for the total sample, with family members’ depression as the strongest predictor (β=0.554), followed by disabled persons’ suicidal ideation (β=0.425). Stratified LASSO analyses revealed that national basic livelihood beneficiary status was the strongest predictor for families with congenital disability (β=0.541), while family members’ depression was the strongest predictor for families with acquired disability (β=0.562), demonstrating distinct predictive patterns by disability onset.
  • Conclusion
    These findings show that predictive factors differ substantially by disability onset type, indicating the need for tailored intervention approaches and offering an evidence-based foundation for targeted suicide prevention strategies.
  • Keywords: Age of onset; Family; Machine learning; Persons with disabilities; Suicidal ideation
Suicide is a major global public health concern and one of the leading causes of death worldwide. Each year, approximately 703,000 individuals die by suicide [1]. In the Republic of Korea, the suicide rate is 24.6 per 100,000, which is more than twice the average rate reported by the Organisation for Economic Co-operation and Development (OECD) [2]. Although suicide rates in many OECD countries have declined [3], the Republic of Korea’s rate has remained persistently high [4], underscoring the urgent need for accurate prediction tools that can identify high-risk groups and facilitate early intervention.
In the Republic of Korea, families of persons with disabilities face intensified challenges due to unique sociocultural dynamics. Confucian cultural norms emphasizing familial responsibility position caregiving as a moral obligation rather than a discretionary role [5]. Despite government programs such as disability pensions and personal assistance services, these supports often prove inadequate, leaving families with primary caregiving responsibilities while facing limited institutional support [6,7]. Persistent social stigma—particularly surrounding mental or developmental disabilities—contributes to family isolation and psychological strain, as families may withdraw from social activities to avoid discrimination [8,9]. These culturally specific pressures, combined with universal caregiving challenges, create a uniquely vulnerable population requiring targeted intervention strategies [10–12].
Family members of persons with disabilities represent one of the most vulnerable groups for suicide risk. Research indicates that suicide ideation rates among family caregivers reach 24.9%, roughly 5 times higher than the general population rate of 4.7% [6]. Depression scores in this population average 20 points, which is substantially higher than scores seen in other high-risk groups such as older adults [9]. Multiple stressors contribute to this elevated risk, including caregiving burden, economic hardship, social isolation, and family conflict, which together create complex vulnerability patterns requiring comprehensive prediction strategies [10–12]. Despite this pronounced risk, existing suicide prediction research has focused primarily on the general population or other vulnerable groups such as older adults and individuals with mental illness, leaving a critical gap in evidence-based prediction tools for family members of persons with disabilities [13,14].
Furthermore, families of persons with disabilities do not represent a uniform population, and the timing of disability onset may create distinct vulnerability profiles that require different predictive approaches. Prior studies have shown that persons with congenital versus acquired disabilities exhibit different suicide risks and contributing factors [15–17]. Likewise, families of persons with congenital disabilities face challenges that differ fundamentally from those confronting acquired disabilities. Families dealing with congenital disabilities often experience chronic strains and long-term adaptation processes due to lifelong caregiving responsibilities [5]. In contrast, families facing acquired disabilities may undergo acute adjustment crises following sudden onset, sharing experiences of grief, trauma, and disruption of anticipated life trajectories with the disabled individual [18]. These differing contexts suggest that effective predictive models should consider disability onset type. However, no studies to date have developed separate predictive algorithms for families of persons with congenital versus acquired disabilities.
Moreover, existing approaches to suicide prediction among families of persons with disabilities face methodological limitations that hinder comprehensive risk assessment. Most studies have relied on traditional statistical methods that examine limited numbers of variables because of sample size constraints and multicollinearity issues [14,15]. Additionally, the small body of prior research on this population makes it difficult to identify optimal predictors for different subgroups. Machine learning approaches can address these challenges by analyzing multiple risk factors simultaneously while identifying new patterns without prior assumptions [19]. These methods are particularly valuable for comparing predictive factors across distinct groups and generally demonstrate superior predictive performance [20,21], making them well suited for developing evidence-based prediction models.
This study is guided by 2 complementary theoretical frameworks. First, the stress process model posits that disability-related primary stressors (e.g., functional limitations) give rise to secondary stressors (e.g., economic strain, family conflict) that accumulate over time, with differential impacts depending on disability onset timing [22]. This model helps explain why families dealing with congenital disabilities may experience chronic strain due to lifelong caregiving, whereas families confronting acquired disabilities often face acute adaptation crises. Second, ecological systems theory emphasizes that suicidal ideation arises from complex interactions across multiple system levels—individual factors (depression, self-esteem), microsystem factors (family relationships, caregiving burden), and exosystem factors (social support, economic resources) [23]. This multi-level perspective informed our comprehensive variable selection across individual, family, and social domains, allowing us to account for the interconnected nature of risk factors rather than examining isolated predictors. These frameworks suggest that effective prediction models must consider both the temporal dynamics of stress accumulation and the multi-systemic nature of suicide risk.
Based on these theoretical foundations, we developed machine learning predictive models for suicidal ideation among family members of persons with disabilities using nationally representative Korean data. We incorporated a comprehensive assessment of risk factors across individual, family, and social dimensions and conducted stratified analyses to investigate differential predictive patterns between families of persons with congenital versus acquired disabilities. The study aimed to (1) develop machine learning models for predicting suicidal ideation among family members of persons with disabilities and compare their performance with traditional statistical approaches, (2) examine differences in risk factors and predictive patterns between congenital and acquired disability groups through stratified analysis, and (3) identify the strongest predictor variables within each disability onset group to guide evidence-based suicide prevention interventions and inform resource allocation strategies.
Study Population and Datasets
This research drew upon data from the 2018 Disability and Life Dynamics Panel (DLDP), a comprehensive nationwide assessment of living conditions among people with disabilities in the Republic of Korea. The DLDP was established to generate data on the life circumstances of disabled individuals and their families and to provide essential information for policy development.
The DLDP employed stratified sampling based on disability type, severity, sex, and other relevant characteristics to ensure adequate representation of registered disabled persons, excluding those residing in institutional facilities. Potential participants were initially contacted by telephone to explain the purpose of the research, and data collection proceeded only after informed consent was obtained. Face-to-face structured interviews were conducted using Tablet Assisted Personal Interviewing, with interviewers trained to approach individuals with disabilities ethically and respectfully. Detailed methodological information is available through the Korea Disabled People’s Development Institute website (https://www.koddi.or.kr/data/research01_view.jsp?brdNum=7422083).
From an initial pool of 6,121 disabled individuals, 284 declined participation, resulting in 5,837 respondents who consented to the survey. During data processing, 236 cases were identified as having missing values on key variables, yielding 5,601 respondents with complete data. Complete case analysis was applied, given the low proportion of missing data (4.0%), which provides unbiased estimates when missingness is below 20% and is unrelated to the outcome [24]. For this study, we identified 3,214 disabled individuals living with family members after excluding those who lived alone. Our primary research subjects were the 2,839 adult family caregivers who had lived with these disabled individuals (aged 14 or older) for at least 6 months. Using household identifiers, we linked data from family caregivers with information about the disabled individuals they lived with, enabling us to analyze how the characteristics of disabled persons were associated with the experiences of their family members.
Adjusted sample weights provided by the DLDP were used to account for non-responses and stratified sampling of specific subpopulations, allowing the estimates to be representative of the disabled population in Korea. Weighting procedures began with a base sampling weight, followed by non-response adjustments and post-stratification to match demographic distributions for region, disability type, severity, sex, and age. Descriptive statistics incorporated survey weights to reflect the complex sampling design, whereas penalized regression analyses were performed on unweighted data using STATA ver. 16.0 (StataCorp LLC).
Variables
The dependent variable was suicidal ideation among family members of persons with disabilities, assessed with a single dichotomous item: “Have you seriously thought about suicide during the past year from the date of the survey?” (0=no, 1=yes).
Based on previous research, factors associated with suicidal ideation were categorized into 4 domains [7–9,11,17,18,24–27]. Individual characteristics of family members included sex (female=0, male=1), age (≤20 to 71–80 years), educational level (none to college or higher), depression (Center for Epidemiologic Studies Depression Scale [CESD]-11), self-esteem, employment status (unemployed=0, employed=1), marital status (single/divorced/widowed=0, married=1), and chronic illness (no=0, yes=1).
Individual characteristics of persons with disabilities included suicidal ideation (0=no, 1=yes), sex (female=0, male=1), age (21–30 to ≥81 years), depression (CESD-11), disability grade (mild=0, severe=1), disability type (external physical=0, internal physical=1, mental=2), disability acceptance, activities of daily living (ADL), and meal companionship (eating with others=0, eating alone=1).
Family characteristics included the number of family members (2, 3, ≥4), family type (ascendant relatives=0, spouse=1, descendants=2, extended family=3), national basic livelihood recipient status (no=0, yes=1), log of family income, financial crisis level, family strengths, interfamilial conflict (no=0, yes=1), and caregiving burden.
Social characteristics included social relationship satisfaction, frequency of contact with friends, frequency of professional contact, and leisure activity companion patterns (non-participant=0, participates alone=1, participates with family=2, participates with acquaintances=3).
Data Analysis
To identify factors associated with suicidal ideation among family members of persons with disabilities, we applied machine learning approaches for predictive modeling. The dataset was randomly divided into a training set (70%) and a test set (30%) to evaluate model performance.
We compared 4 predictive techniques: multivariate logistic regression as the baseline model; least absolute shrinkage and selection operator (LASSO) with cross-validation (CV) selection (α=1, 10-fold CV for optimal λ); LASSO with adaptive selection (α=1, varying penalty weights); and ridge regression (α=0). These penalized regression methods were selected for their ability to perform simultaneous variable selection and regularization, yielding predictive performance comparable to more computationally intensive algorithms such as random forest or support vector machines in datasets of this dimensionality (31 variables), while maintaining superior interpretability and efficiency [19–21]. Because our primary aim was to identify actionable predictors for policy and clinical interventions, interpretable models with clear coefficient estimates were prioritized over black-box approaches such as neural networks or tree-based ensembles. Although further hyperparameter tuning beyond cross-validated λ could be performed, penalized linear models typically show limited additional gains relative to non-linear algorithms, and our CV approach sufficiently optimized the main regularization parameter.
Model performance was evaluated using established metrics: area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1-score. AUC values between 0.7–0.8 indicate acceptable discrimination, 0.8–0.9 excellent discrimination, and values above 0.9 outstanding discrimination [25]. Sensitivity and specificity assessed the accuracy of identifying true positive and true negative cases, respectively. PPV and NPV measured the proportion of correct positive and negative predictions, while the F1-score represented the harmonic mean of precision and recall. The optimal model was selected based on having the highest AUC, PPV, NPV, and F1-score in both the training and test datasets [28,29]. To support clinical interpretability, we presented standardized coefficients as indicators of relative importance and conducted stratified analyses to identify group-specific risk patterns that would provide practical screening insights without requiring advanced computational tools.
After determining the best-performing model, we calculated penalized coefficients from the LASSO models and compared them with coefficients from the multivariate logistic regression to assess the effects of regularization. Standardized coefficients were calculated to compare relative variable importance across groups, rather than to estimate effect sizes. We also performed stratified analyses comparing risk factors between family members of persons with congenital disabilities and those with acquired disabilities to identify potential differences in predictive patterns. Standardized coefficients were compared across 3 groups (total sample, congenital disability, and acquired disability) to determine the most influential predictors for each group.
Ethics Statement
The survey conducted by the Korea Disabled Persons Development Institute received approval from Korean Public Institutional Review Board (No: P01-2018-09-22-004). Informed consent was waived because of the retrospective nature of this study.
Characteristics of the Study Participants
Based on the complex sample analysis, Table 1 presents the characteristics of the study participants in the weighted sample, which included 124,783 family members of persons with disabilities. Among them, 113,403 (90.9%) were family members of persons with acquired disabilities, and 11,380 (9.1%) were family members of persons with congenital disabilities.
Development of a Predictive Model for Suicidal Ideation among Family Members of Persons with Disabilities
To develop predictive models for suicidal ideation among family members of persons with disabilities, we randomly divided the dataset into a training set (70%) and a test set (30%). Four different machine learning approaches were evaluated: multivariate logistic regression, LASSO with CV selection, LASSO with adaptive selection, and ridge regression.
Table 2 presents the performance metrics of the 4 predictive models in both the training and test sets. In the training set, all models demonstrated good discriminative ability, with AUC values ranging from 0.869 to 0.875. The LASSO with CV selection model showed the highest AUC (0.875), followed closely by multivariate logistic regression (0.872), ridge regression (0.870), and LASSO with adaptive selection (0.869). ROC curves for all four models in the training set are shown in Figure S1. When applied to the test set, all models showed the expected slight decreases in performance, with AUC values ranging from 0.848 to 0.853, indicating good generalizability. ROC curves for the test set are presented in Figure S2. For sensitivity, which reflects the ability to correctly identify individuals with suicidal ideation, the training set values ranged from 0.415 to 0.487, with LASSO with CV selection performing best (0.487). In the test set, both multivariate logistic regression and LASSO with CV selection achieved the highest sensitivity (0.461), followed by LASSO with adaptive selection (0.438) and ridge regression (0.388). All models exhibited high specificity in both the training (0.955–0.965) and test sets (0.948–0.957), with ridge regression yielding the highest specificity in both datasets (0.965 and 0.957, respectively). PPV values ranged from 0.393 to 0.421 in the training set and from 0.371 to 0.396 in the test set, with LASSO with CV selection consistently showing the highest PPV. NPV was consistently high across all models in both the training (0.964–0.968) and test sets (0.959–0.964). In terms of overall accuracy, values ranged from 0.925 to 0.930 in the training set and from 0.917 to 0.922 in the test set. The highest F1-scores were observed for LASSO with CV selection in both the training (0.451) and test sets (0.426). Taken together, the LASSO with CV selection model offered the most balanced predictive performance, showing the highest AUC, PPV, NPV, and F1-score across both datasets. Model calibration was also assessed using Brier scores and calibration plots. All 4 models demonstrated good calibration (Brier scores, 0.049–0.051), with the LASSO-CV model performing best. Calibration plots are shown in Figure S3.
Factors Associated with Suicidal Ideation among Family Members of Persons with Disabilities
Table 3 presents factors associated with suicidal ideation among family members of persons with disabilities in the total sample using 3 analytical approaches: traditional multivariate logistic regression, LASSO regression, and multivariate logistic regression using LASSO-selected variables.
In the traditional multivariate logistic regression model including all 31 variables, 6 factors were statistically significant at the p<0.05 level: depression of family members (coefficient=0.099, p<0.001), self-esteem of family members (coefficient=−0.086, p=0.002), chronic illness of family members (coefficient=0.925, p<0.001), suicidal ideation of persons with disabilities (coefficient=1.283, p<0.001), interfamilial conflict (coefficient=0.721, p<0.001), and caregiving burden (coefficient=0.208, p=0.009).
The LASSO regression retained 16 of the original 31 variables. These included sex of family members, age of family members, depression of family members, self-esteem of family members, chronic illness of family members, suicidal ideation of persons with disabilities, age of persons with disabilities, disability acceptance of persons with disabilities, ADL of persons with disabilities, number of family members, family type (spouse), family type (extended family members), beneficiary of national basic livelihood status, financial crisis, family strengths, interfamilial conflict, and caregiving burden.
To validate the statistical significance of the LASSO-selected predictors, we conducted a subsequent multivariate logistic regression including only these 16 variables. Among these variables, 8 were statistically significant at the p<0.05 level: depression of family members (coefficient=0.087, p<0.001), self-esteem of family members (coefficient=−0.085, p=0.002), chronic illness of family members (coefficient=0.950, p<0.001), suicidal ideation of persons with disabilities (coefficient=1.217, p<0.001), age of family members (coefficient=−0.275, p<0.001), age of persons with disabilities (coefficient=−0.188, p=0.010), family type—spouse (coefficient=0.435, p=0.046), interfamilial conflict (coefficient=0.700, p<0.001), and caregiving burden (coefficient=0.212, p=0.007).
Factors Associated with Suicidal Ideation among Family Members of Persons with Disabilities: Comparison between Congenital and Acquired Disability
Table 4 presents the results of the LASSO regression analyses identifying factors associated with suicidal ideation among family members of persons with disabilities, stratified by congenital and acquired disability groups.
In the congenital disability group, the LASSO regression retained 10 of the original 31 variables. The selected factors included sex of family members, depression of family members, suicidal ideation of persons with disabilities, age of persons with disabilities, depression of persons with disabilities, eating alone, beneficiary of national basic livelihood status, log of family income, financial crisis, interfamilial conflict, and frequency of professional contact.
In the acquired disability group, the LASSO regression retained 19 of the 31 original variables. These included sex of family members, age of family members, depression of family members, self-esteem of family members, chronic illness of family members, suicidal ideation of persons with disabilities, age of persons with disabilities, internal physical disabilities, mental disabilities, disability acceptance of persons with disabilities, ADL of persons with disabilities, eating alone, number of family members, family type (spouse), family type (extended family members), family strengths, interfamilial conflict, caregiving burden, frequency of professional contact, and leisure activity companion type (participates alone).
Variables with the Strongest Predictive Associations for Suicidal Ideation in Family Members of Persons with Disabilities
Figure 1 presents the 10 variables with the strongest predictive associations for suicidal ideation among family members of persons with disabilities, based on standardized coefficients derived from the LASSO regression analyses. These standardized coefficients reflect relative importance within the regularized predictive model rather than unbiased effect sizes.
In the total sample, the variables showing the strongest predictive associations were depression of family members (β=0.554), suicidal ideation of persons with disabilities (β=0.425), chronic illness of family members (β=0.371), interfamilial conflict (β=0.310), and self-esteem of family members (β=0.261). Additional important predictors included age of family members (β=0.260), caregiving burden (β=0.249), age of persons with disabilities (β=0.165), sex of family members (β=0.083), and family type—spouse (β=0.069).
For the congenital disability group, the variables with the strongest predictive associations were beneficiary of national basic livelihood status (β=0.541), interfamilial conflict (β=0.488), depression of family members (β=0.332), suicidal ideation of persons with disabilities (β=0.321), and sex of family members (β=0.309). These were followed by depression of persons with disabilities (β=0.209), financial crisis (β=0.112), log of family income (β=0.075), eating alone (β=0.058), and frequency of professional contact (β=0.050).
For the acquired disability group, the variables with the strongest predictive associations were depression of family members (β=0.562), suicidal ideation of persons with disabilities (β=0.422), chronic illness of family members (β=0.392), interfamilial conflict (β=0.282), and age of family members (β=0.273). Additional significant predictors included self-esteem of family members (β=0.270), caregiving burden (β=0.224), age of persons with disabilities (β=0.156), frequency of professional contact (β=−0.107), and ADL of persons with disabilities (β=0.091).
Comparative analysis of shared predictors across groups revealed substantial differences in predictive strength. Interfamilial conflict showed a stronger association in congenital families (β=0.488) than in acquired families (β=0.282), whereas depression of family members displayed the opposite pattern, with a stronger association in acquired families (β=0.562) than in congenital families (β=0.332). Suicidal ideation of persons with disabilities demonstrated a stronger association in acquired families (β=0.422) compared with congenital families (β=0.321). Age of persons with disabilities showed a more pronounced negative association in acquired families (β=−0.156) than in congenital families (β=−0.026). Frequency of professional contact showed a negative association in acquired families (β=−0.107) and a minimal positive association in congenital families (β=0.050). Sex of family members exhibited a weaker association in acquired families (β=0.032) than in congenital families (β=0.309).
This study developed machine learning models for predicting suicidal ideation among family members of persons with disabilities using nationally representative DLDP data. Three key findings emerged. First, multivariate logistic regression identified 6 significant predictors, with depression of family members (coefficient=0.099), suicidal ideation of persons with disabilities (coefficient=1.283), and chronic illness (coefficient=0.925) showing strongest associations (all p<0.001). Second, strong evidence of suicide clustering emerged, with disabled persons' suicidal ideation as the second strongest predictor (β=0.425). Third, stratified analyses revealed distinct patterns by disability onset: economic factors dominated in congenital families (β=0.541 for national basic livelihood status), while psychological factors were strongest in acquired families (β=0.562 for depression). Shared predictors also showed differential strength, with interfamilial conflict more salient in congenital families and depression stronger in acquired families.
We compared our findings with prior studies examining suicidal ideation in disability populations. The strong predictive role of depression is consistent with previous research conducted in general populations and among disability caregivers [11,14]. However, our results advance this literature by demonstrating that multiple factors—including chronic illness, interfamilial conflict, and caregiving burden—contribute simultaneously to suicide risk. Most notably, the strong association between suicidal ideation among persons with disabilities and their family members (β=0.425) provides the first empirical evidence of suicide clustering within Korean disability households. Although suicide clustering has been documented in various contexts [30,31], it has rarely been investigated in Asian populations or specifically within disability family systems.
The differential predictive patterns between congenital and families with acquired disability align with theoretical expectations while offering novel empirical validation. Prior research has documented chronic economic strain among families caring for individuals with developmental disabilities and psychological adaptation challenges following the sudden onset of disability [5,18]. Our findings extend this literature by showing that disability onset timing is associated with distinct risk mechanisms requiring tailored intervention strategies. Economic predictors were most prominent among families with congenital disability, consistent with the long-term accumulation of financial and caregiving burdens. In contrast, psychological predictors predominated among families with acquired disability, reflecting acute adjustment crises following unexpected disability onset.
The significance of this investigation is threefold. First, to our knowledge, this is the first study to develop machine learning predictive models for suicidal ideation among family members of persons with disabilities using nationally representative Korean data, achieving superior predictive performance compared with traditional statistical methods. Second, we conducted a comprehensive analysis incorporating 31 variables across individual, family, and social dimensions, offering an integrated understanding that extends beyond single-factor approaches. Third, we empirically demonstrated differential risk patterns between congenital and families with acquired disability, establishing an evidence-based foundation for developing tailored intervention strategies based on disability onset characteristics.
Our findings provide support for multiple theoretical frameworks. The identification of multidimensional predictors aligns with ecological systems theory [23], showing that suicidal ideation among family members results from interactions among individual factors (depression, chronic illness), family factors (interfamilial conflict, caregiving burden), and interpersonal factors (suicidal ideation of persons with disabilities). The strong association between suicidal ideation of persons with disabilities and their family members supports the assortative relation mechanism of suicide clustering by Joiner [31]. Disability families are characterized by geographical proximity, psychosocial closeness, and overlapping stress exposures, creating what Hawton et al. [32] described as circles of vulnerability. The differential patterns between congenital and families with acquired disability are consistent with the stress process model [22], in which congenital disabilities are associated with chronic secondary stressors such as economic hardship, whereas acquired disabilities are associated with acute adaptation responses, including psychological distress.
These findings have direct practice implications. The multidimensional predictive factors indicate need for integrated assessment approaches rather than single-factor screening. Healthcare systems should establish comprehensive frameworks assessing individual, family, and interactive factors simultaneously. The observed clustering effect necessitates family-system–based evaluations rather than separate individual assessments. Our model provides technical foundation for developing such screening instruments.
Differential patterns require tailored interventions by disability onset type. For congenital disability families, prioritize structural interventions addressing economic vulnerabilities: expanding national basic livelihood programs, increasing disability allowances, and developing comprehensive financial support systems. For acquired disability families, emphasize therapeutic interventions focused on psychological adaptation: routine depression screening, immediate counseling access, and specialized adaptation programs for sudden disability onset. The clinical applicability is demonstrated through clear risk hierarchies directly applicable to assessment protocols, with emphasis varying by onset type.
Resource allocation should follow disability-specific sequences. For congenital families: economic support, family conflict mediation, then psychological interventions. For acquired families: psychological support, crisis intervention, then social support mechanisms. Healthcare systems should use these predictive hierarchies to develop risk stratification protocols enabling efficient allocation of intensive services to highest-risk families.
This study has several limitations. First, the cross-sectional design prevents the establishment of causal relationships between variables, particularly regarding the mutual influence processes between persons with disabilities and their family members. Although our machine learning models demonstrated strong predictive accuracy, the identified variables should be interpreted as predictive indicators rather than causal determinants of suicidal ideation. Longitudinal investigations are therefore needed to examine the temporal development of suicide clustering phenomena within disability households. Second, the single-item measurement of suicidal ideation, although standard in national surveys such as Korea National Health and Nutrition Examination Survey, cannot capture the complexity or severity of suicidal thoughts. Despite this limitation, such measures have demonstrated acceptable predictive validity and may reduce respondent burden in vulnerable populations [33]. Third, the relatively smaller sample size for families with congenital disability (n=11,380) compared to families with acquired disability (n=113,403) may have affected the stability of stratified analyses. Fourth, common issues associated with survey-based research, including non-response bias and social desirability bias, cannot be ruled out, and thus the findings should be interpreted with caution. Fifth, this study used 2018 data, which predates the coronavirus disease 2019 pandemic. Although more recent 2023 data are available, substantial participant attrition between waves raised concerns about selection bias, particularly because dropout patterns may correlate with caregiving burden and mental health outcomes. The 2018 data therefore provide more complete pre-pandemic baseline patterns, though future research should examine how pandemic-related stressors may have altered these predictive relationships. Sixth, we did not examine sex concordance between caregivers and persons with disabilities, nor did we assess differential impacts by relationship type (e.g., spouse versus parent caregivers), which may influence caregiving dynamics and associated stress. Future research should explore whether sex-matched or mismatched dyads exhibit different vulnerability patterns. Seventh, although we included social characteristic variables (frequency of professional contact, social relationship satisfaction, leisure activity patterns), these serve as proxy indicators rather than direct measures of structural factors. While professional contact frequency may reflect regional service availability and social satisfaction may approximate community support, these individual-level perceptions cannot capture objective features such as regional service density, geographic accessibility of mental health resources, or disability infrastructure investments. Future studies should incorporate community-level indicators to examine how structural disparities influence suicide risk. Eighth, our findings are based solely on Korean data without external validation. The unique cultural context of Korean disability families—including Confucian caregiving obligations, specific social welfare systems, and culturally embedded stigma patterns—may limit generalizability to other populations. Additionally, temporal validation using different DLDP waves was not feasible because of substantial panel attrition, which may introduce selection bias. Future studies should validate these predictive patterns in other cultural contexts and assess model stability across different time periods. Ninth, although we provided global model interpretability through standardized coefficients and stratified analyses, we did not implement local explainability methods such as Shapley additive explanations or local interpretable model-agnostic explanations that could yield patient-specific risk explanations. These methods, not currently available in Stata’s LASSO implementation, would require additional computational frameworks. While our population-level insights are appropriate for policy planning, clinical decision-support tools would benefit from individual-level interpretability in future applications. Tenth, the model’s moderate sensitivity likely reflects the inherent difficulty of predicting low base-rate events using distal, population-level predictors, a pattern consistently reported in suicide prediction research [34]. The high specificity suggests that the identified risk profiles remain meaningful for population-level risk stratification and resource allocation. Nevertheless, future research should explore threshold optimization or the incorporation of proximal risk indicators to enhance sensitivity for clinical screening. Finally, although our machine learning approach provided superior predictive accuracy, it has inherent methodological limitations. The LASSO coefficients presented in this study reflect penalized estimates representing relative contribution to prediction under regularization rather than unbiased effect sizes. Because of the shrinkage property of LASSO, these coefficients tend to be smaller than those from traditional regression models and should not be interpreted as measures of actual effect magnitude. Moreover, this approach provides limited insight into complex interaction effects. Despite these limitations, the findings offer the first evidence-based foundation for developing tailored suicide prevention strategies for Korean disability families based on disability onset characteristics. Future research should address these limitations by employing longitudinal study designs and more comprehensive suicide assessment instruments [35].
This study provides 3 fundamental contributions to understanding suicidal ideation among family members of persons with disabilities. First, the identification of multidimensional predictors demonstrated that suicide risk arises from complex interactions among multiple factors, underscoring the need for integrated assessment approaches. Second, the differential predictive patterns between congenital and families with acquired disability established that disability onset timing is associated with distinct vulnerability pathways requiring tailored prevention strategies. Third, the analysis of the most strongly predictive variables offered strategic guidance for efficient resource allocation based on disability-specific intervention priorities. Together, these findings provide a critical empirical foundation for researchers, mental health professionals, and policymakers to develop evidence-based, targeted interventions for suicide prevention within Korean disability family systems.
• This study developed, to our knowledge, the first machine learning model designed to predict suicide risk among family members caring for persons with disabilities in the Republic of Korea.
• Family members’ depression was the strongest overall predictor, and suicide risk was closely linked between disabled persons and their families.
• Predictive factors differed by disability onset. Economic hardship showed the strongest predictive association for families with congenital disabilities, whereas depression was the strongest predictor for families dealing with acquired disabilities.
• These findings offer guidance for developing tailored suicide prevention strategies for families of persons with congenital versus acquired disabilities.

Ethics Approval

The survey conducted by the Korea Disabled Persons Development Institute received approval from a Korean public institution’s Institutional Review Board (No: P01-2018-09-22-004), and this study was performed in accordance with the principles of the Declaration of Helsinki. The informed consent was waived because of the retrospective nature of this study.

Conflicts of Interest

The author has no conflicts of interest to declare.

Funding

None.

Availability of Data

The datasets generated and/or analyzed during the current study are available in (the Disability Life Panel Survey of Korea) repository, https://koddi.or.kr/stat/html/user/main/main.

Supplementary data are available at https://doi.org/10.24171/j.phrp.2025.0261.
Figure S1.
Receiver operating characteristic (ROC) curves for four predictive models evaluated on the training dataset (70% of total sample): (A) multivariate logistic regression (AUC=0.872), (B) LASSO with cross-validation (LASSO-CV; AUC=0.875), (C) LASSO with adaptive selection (LASSO-Adapt; AUC=0.869), and (D) ridge regression (AUC=0.870). The diagonal reference line represents random classification (AUC=0.5). AUC, area under the receiver operating characteristic curve.
j-phrp-2025-0261-Supplementary-Figure-S1.pdf
Figure S2.
Receiver operating characteristic (ROC) curves for four predictive models evaluated on the independent test dataset (30% of total sample): (A) multivariate logistic regression (AUC=0.852), (B) LASSO with cross-validation (LASSO-CV; AUC=0.853), (C) LASSO with adaptive selection (LASSO-Adapt; AUC=0.848), and (D) ridge regression (AUC=0.852). The diagonal reference line represents random classification (AUC=0.5). AUC, area under the receiver operating characteristic curve.
j-phrp-2025-0261-Supplementary-Figure-S2.pdf
Figure S3.
Calibration plots for predictive models. Calibration plots displaying the relationship between predicted probabilities (x-axis) and observed frequencies (y-axis) for each risk decile in the independent test dataset: (A) multivariate logistic regression (LOGIT), (B) LASSO with cross-validation (LASSO-CV), (C) LASSO with adaptive selection (LASSO-ADAPT), and (D) ridge regression (RIDGE). Each point represents the mean predicted and observed probability for one decile. The diagonal red dashed line indicates perfect calibration.
j-phrp-2025-0261-Supplementary-Figure-S3.pdf
Figure 1.
Predictive variables for suicidal ideation in family members of persons with disabilities: standardized coefficients from least absolute shrinkage and selection operator (LASSO) regression. Coefficients represent relative importance for prediction under regularization, not unbiased effect estimates.
Figure 1. Predictive variables for suicidal ideation in family members of persons with disabilities: standardized coefficients from least absolute shrinkage and selection operator (LASSO) regression. Coefficients represent relative importance for prediction under regularization, not unbiased effect estimates.  
A machine learning approach for predicting suicidal ideation among family members of persons with disabilities: a cross-sectional study in the Republic of Korea
Table 1.
Characteristics of the study participants
Table 1.
Variable Total (n=124,783) Acquired (n=113,403) Congenital (n=11,380)
Individual characteristics of family members
 Sex of family members
  Female 76,517 (61.320) 77,216 (61.880) 70,303 (56.340)
  Male 48,266 (38.680) 47,567 (38.120) 54,480 (43.660)
 Age of family members (y)
  ≤20 661 (0.530) 686 (0.550) 437 (0.350)
  21–30 8,311 (6.660) 8,298 (6.650) 8,348 (6.690)
  31–40 9,933 (7.960) 9,920 (7.950) 10,107 (8.100)
  41–50 15,598 (12.500) 14,849 (11.900) 22,411 (17.960)
  51–60 37,447 (30.010) 37,360 (29.940) 38,221 (30.630)
  61–70 34,590 (27.720) 34,914 (27.980) 31,632 (25.350)
  71–80 18,243 (14.620) 18,755 (15.030) 13,626 (10.920)
 Education level of family members
  None 4,617 (3.700) 4,829 (3.870) 2,633 (2.110)
  Elementary 19,516 (15.640) 19,628 (15.730) 18,455 (14.790)
  Middle school 18,817 (15.080) 18,455 (14.790) 21,974 (17.610)
  High school 54,068 (43.330) 53,919 (43.210) 55,366 (44.370)
  ≤College 27,777 (22.260) 27,939 (22.390) 26,367 (21.130)
 Depression of family members 20.660±6.490 20.800±6.530 19.410±6.000
 Self-esteem of family members 28.410±3.620 28.400±3.620 28.520±3.660
 Employment status of family members
  Unemployed 68,830 (55.160) 69,741 (55.890) 60,632 (48.590)
  Employed 55,953 (44.840) 55,042 (44.110) 64,151 (51.410)
 Marital status of family members
  Single/divorced/widowed 29,586 (23.710) 28,862 (23.130) 36,025 (28.870)
  Married 95,197 (76.290) 95,921 (76.870) 88,758 (71.130)
 Chronic illness of family members
  No 72,561 (58.150) 72,574 (58.160) 72,499 (58.100)
  Yes 52,222 (41.850) 52,209 (41.840) 52,284 (41.900)
 Suicidal ideation of family members
  No 116,959 (93.730) 116,722 (93.540) 119,068 (95.420)
  Yes 7,824 (6.270) 8,061 (6.460) 5,715 (4.580)
Individual characteristics of persons with disabilities
 Suicidal ideation of persons with disabilities
  No 104,606 (83.830) 103,932 (83.290) 110,720 (88.730)
  Yes 20,177 (16.170) 20,851 (16.710) 14,063 (11.270)
 Sex of persons with disabilities
  Female 56,527 (45.300) 55,479 (44.460) 65,910 (52.820)
  Male 68,256 (54.700) 69,304 (55.540) 58,873 (47.180)
 Age of persons with disabilities (y)
  21–30 1,622 (1.300) 1,123 (0.900) 6,152 (4.930)
  31–40 9,496 (7.610) 6,738 (5.400) 34,265 (27.460)
  41–50 9,234 (7.400) 8,011 (6.420) 20,215 (16.200)
  51–60 15,298 (12.260) 15,236 (12.210) 15,822 (12.680)
  61–70 37,672 (30.190) 38,733 (31.040) 28,126 (22.540)
  71–80 40,168 (32.190) 42,788 (34.290) 16,696 (13.380)
  ≥81 11,293 (9.050) 12,166 (9.750) 3,519 (2.820)
 Depression of persons with disabilities 22.810±7.210 23.060±7.260 20.560±6.360
 Grade of disability of persons with disabilities
  Mild 56,352 (45.160) 57,824 (46.340) 43,063 (34.510)
  Severe 68,431 (54.840) 66,959 (53.660) 81,720 (65.490)
 Type of disability of persons with disabilities
  External physical disabilities 79,025 (63.330) 81,658 (65.440) 55,366 (44.370)
  Internal physical disabilities 36,299 (29.090) 34,141 (27.360) 55,803 (44.720)
  Mental disabilities 9,446 (7.570) 8,984 (7.200) 13,626 (10.920)
 Disability acceptance of persons with disabilities 27.450±5.430 27.340±5.430 28.470±5.290
 ADL of persons with disabilities 23.730±10.650 24.020±10.800 21.100±8.850
 Eating alone
  No 86,362 (69.210) 86,687 (69.470) 83,480 (66.900)
  Yes 38,421 (30.790) 38,096 (30.530) 41,303 (33.100)
Family characteristics of households with disability
 No. of family members
  2 60,133 (48.190) 62,030 (49.710) 43,063 (34.510)
  3 34,365 (27.540) 33,205 (26.610) 44,822 (35.920)
  ≤4 30,285 (24.270) 29,549 (23.680) 36,911 (29.580)
 Family type
  Ascendant relatives 25,269 (20.250) 21,338 (17.100) 60,632 (48.590)
  Spouse 74,770 (59.920) 78,389 (62.820) 42,177 (33.800)
  Descendants 18,243 (14.620) 19,042 (15.260) 10,981 (8.800)
  Extended family members 6,501 (5.210) 6,002 (4.810) 10,981 (8.800)
 Beneficiary of national basic livelihood
  No 106,976 (85.730) 107,837 (86.420) 99,302 (79.580)
  Yes 17,807 (14.270) 16,946 (13.580) 25,481 (20.420)
 Log of family income 4.900±3.850 4.890±3.900 5.050±3.430
 Financial crisis 16.910±5.470 16.940±5.470 16.630±5.490
 Family strengths 64.400±10.500 64.440±10.500 64.020±10.440
 Interfamilial conflict
  No 87,373 (70.020) 87,860 (70.410) 83,043 (66.550)
  Yes 37,410 (29.980) 36,923 (29.590) 41,740 (33.450)
 Caregiving burden 1.790±1.400 1.830±1.420 1.520±1.260
Social characteristics of households with disability
 Social relationship satisfaction 5.430±1.730 5.420±1.710 5.490±1.880
 Frequency of contact with friends 3.290±2.010 3.300±2.010 3.170±2.030
 Frequency of professional contact 2.030±1.080 2.030±1.080 2.050±1.070
 Leisure activity companion
  Non-participant 16,609 (13.310) 16,946 (13.580) 13,626 (10.920)
  Participates alone 59,422 (47.620) 58,798 (47.120) 65,024 (52.110)
  Participates with family 37,135 (29.760) 37,897 (30.370) 30,322 (24.300)
  Participates with acquaintances 11,605 (9.300) 11,131 (8.920) 15,822 (12.680)

Data are presented as n (%) or mean±standard deviation.

ADL, activities of daily living.

Table 2.
Development of a predictive model for suicidal ideation among family members of persons with disabilities
Table 2.
Multivariate logistic LASSO with CV selection LASSO with adaptive selection Ridge
Training set Test set Training set Test set Training set Test set Training set Test set
AUC 0.872 0.852 0.875 0.853 0.869 0.848 0.870 0.852
Sensitivity 0.485 0.461 0.487 0.461 0.465 0.438 0.415 0.388
Specificity 0.955 0.948 0.96 0.953 0.958 0.951 0.965 0.957
PPV 0.393 0.371 0.421 0.396 0.397 0.375 0.408 0.377
NPV 0.967 0.963 0.968 0.964 0.966 0.962 0.964 0.959
Accuracy 0.925 0.917 0.93 0.922 0.927 0.919 0.93 0.922
F1-score 0.434 0.411 0.451 0.426 0.428 0.404 0.411 0.382

LASSO, least absolute shrinkage and selection operator; CV, cross-validation; AUC, area under the receiver operating characteristic curve; PPV, positive predictive value; NPV, negative predictive value.

Table 3.
Factors associated with suicidal ideation among family members of persons with disabilities
Table 3.
Variable Multivariate logistic LASSO regression Multivariate logistic using the LASSO-selected variables
Coef. SE p>z Penalized coef. Std. coef. Coef. SE p>z
Individual characteristics of family members
 Sex of family members (ref.=female) 0.289 0.240 0.228 0.170 0.083 0.325 0.183 0.075
 Age of family members –0.187 0.141 0.185 –0.185 –0.260 –0.275 0.073 0.000
 Education level of family members 0.027 0.096 0.780 x x - - -
 Depression of family members 0.099 0.017 0.000 0.085 0.554 0.087 0.016 0.000
 Self-esteem of family members –0.086 0.028 0.002 –0.072 –0.261 –0.085 0.027 0.002
 Employment status of family members (ref.=unemployed) –0.068 0.199 0.734 x x - - -
 Marital status of family members (ref.=single/divorced/ widowed) –0.402 0.327 0.218 x x - - -
 Chronic illness of family members (ref.=none) 0.925 0.202 0.000 0.752 0.371 0.950 0.199 0.000
Individual characteristics of persons with disabilities
 Suicidal ideation of persons with disabilities (ref.=no) 1.283 0.197 0.000 1.153 0.425 1.217 0.186 0.000
 Sex of persons with disabilities (ref.=female) 0.034 0.229 0.883 x x - - -
 Age of persons with disabilities –0.257 0.142 0.070 –0.117 –0.165 –0.188 0.072 0.010
 Depression of persons with disabilities –0.023 0.016 0.148 x x - - -
 Grade of disability of persons with disabilities –0.068 0.212 0.749 x x - - -
 Type of disability of persons with disabilities (ref.=external physical disabilities)
  Internal physical disabilities –0.037 0.220 0.867 x x - - -
  Mental disabilities 0.323 0.331 0.330 x x - - -
 Disability acceptance of persons with disabilities –0.015 0.019 0.432 –0.001 –0.007 –0.005 0.018 0.767
 ADL of persons with disabilities 0.015 0.012 0.210 0.004 0.047 0.009 0.010 0.382
 Eating alone (ref.=no) –0.146 0.200 0.466 x x - - -
Family characteristics of households with disability
 No. of family members 0.160 0.121 0.186 0.040 0.032 0.125 0.115 0.279
 Family type (ref.=ascendant relatives)
  Spouse 0.870 0.484 0.072 0.141 0.069 0.435 0.218 0.046
  Descendants 0.293 0.788 0.710 x x - - -
  Extended family members –0.131 0.592 0.825 –0.102 –0.023 –0.365 0.444 0.411
  Beneficiary of national basic livelihood (ref.=no) 0.190 0.241 0.429 0.101 0.035 0.233 0.227 0.305
  Log of family income –0.002 0.025 0.945 x x - - -
  Financial crisis 0.012 0.017 0.491 0.000 0.003 0.006 0.017 0.730
  Family strengths –0.005 0.008 0.539 –0.003 –0.032 –0.006 0.008 0.484
  Interfamilial conflict (ref.=no) 0.721 0.192 0.000 0.677 0.310 0.700 0.189 0.000
  Caregiving burden 0.208 0.080 0.009 0.177 0.249 0.212 0.078 0.007
Social characteristics of households with disability
 Social relationship satisfaction 0.014 0.060 0.810 x x - - -
 Frequency of contact with friends 0.039 0.050 0.431 x x - - -
 Frequency of professional contact –0.137 0.086 0.113 –0.063 –0.067 0.016 0.056 0.769
 Leisure activity companion (ref.=non-participant) - - -
  Participates alone 0.227 0.274 0.408 0.117 0.058 0.273 0.183 0.135
  Participates with family –0.085 0.297 0.775 x x - - -
  Participates with acquaintances 0.027 0.438 0.951 x x - - -
Constants –2.449 1.461 0.094 –2.888 –3.403 –2.872 1.338 0.032

Variables marked with “x” were excluded from the final prediction model after least absolute shrinkage and selection operator (LASSO) regression.

Coef., coefficient; SE, standard error; std. LASSO, least absolute shrinkage and selection operator; coef., standardized coefficient; ref., reference; ADL, activities of daily living; -, variable not included in the model.

Table 4.
Factors associated with suicidal ideation among family members of persons with disabilities: comparison between congenital and acquired disability
Table 4.
Variable Congenital disability Acquired disability
Penalized coef. Std. coef. Penalized coef. Std. coef.
Individual characteristics of family members
 Sex of family members (ref.=female) 0.623 0.309 0.067 0.032
 Age of family members x x –0.193 –0.273
 Education level of family members x x x x
 Depression of family members 0.055 0.332 0.086 0.562
 Self-esteem of family members x x –0.075 –0.270
 Employment status of family members (ref.=unemployed) x x x x
 Marital status of family members (ref.=single/divorced/widowed) x x x x
 Chronic illness of family members (ref.=none) x x 0.795 0.392
Individual characteristics of persons with disabilities
 Suicidal ideation of persons with disabilities (ref.=no) 1.016 0.321 1.131 0.422
 Sex of persons with disabilities (ref.=female) x x x x
 Age of persons with disabilities –0.016 –0.026 –0.119 –0.156
 Depression of persons with disabilities –0.033 –0.209 x x
 Grade of disability of persons with disabilities x x x x
 Type of disability of persons with disabilities (fef.=external physical disabilities)        
 Internal physical disabilities x x –0.018 –0.008
 Mental disabilities x x 0.077 0.020
 Disability acceptance of persons with disabilities x x –0.002 –0.012
 ADL of persons with disabilities x x 0.008 0.091
 Eating alone (ref.=no) 0.124 0.058 –0.006 –0.003
Family characteristics of households with disability
 No. of family members x x 0.064 0.052
 Family type (ref.=ascendant relatives)
 Spouse x x 0.146 0.071
 Descendants x x x x
 Extended family members x x –0.316 –0.068
 Beneficiary of national basic livelihood (ref.=no) 1.343 0.541 x x
 Log of family income –0.022 –0.075 x x
 Financial crisis 0.020 0.112 x x
 Family strengths x x –0.006 –0.063
 Interfamilial conflict (ref.=no) 1.034 0.488 0.617 0.282
 Caregiving burden x x 0.158 0.224
Social characteristics of households with disability
 Social relationship satisfaction x x x x
 Frequency of contact with friends x x x x
 Frequency of professional contact 0.046 0.050 –0.099 –0.107
 Leisure activity companion (ref.=non-participant)
 Participates alone x x 0.154 0.077
 Participates with family x x x x
 Participates with acquaintances x x x x
Constants –5.238 –3.526 –2.560 –3.387

Variables marked with “x” were excluded from the final prediction model after LASSO regression.

Coef., coefficient; std. coef., standardized coefficient; ref., reference; ADL, activities of daily living.

  • 1. World Health Organization (WHO) (CH). Suicide worldwide in 2019: global health estimates [Internet]. WHO; 2021 [cited 2025 Jun 20]. Available from: https://www.who.int/publications/i/item/9789240026643.
  • 2. Organisation for Economic Co-operation and Development (OECD) (FR). Health at a glance 2021 [Internet]. OECD Publishing; 2021 [cited 2025 Jun 20]. Available from: https://www.oecd.org/en/publications/2021/11/health-at-a-glance-2021_cc38aa56.html.
  • 3. Naghavi M. Global, regional, and national burden of suicide mortality 1990 to 2016: systematic analysis for the Global Burden of Disease Study 2016. BMJ 2019;364:l94.
  • 4. Kim HH, Lee JH, Song IH, et al. Characteristics and risk factors of suicide among people who attempted self-harm in South Korea: a longitudinal national cohort study in South Korea. Psychiatry Res 2023;330:115613.
  • 5. Park SK, Lee SW, Lee JJ. The relationship between child’s activities of daily living and depression in parents of children with developmental disabilities. J Disabil Welf 2021;52:55−84. Korean.
  • 6. Kim SH, Lee YH, Oh WC, et al. 2017 Survey on persons with disabilities. Korea Institute for Health and Social Affair; 2017. Korean.
  • 7. Ha JH, Hong J, Seltzer MM, et al. Age and gender differences in the well-being of midlife and aging parents with children with mental health or developmental problems: report of a national study. J Health Soc Behav 2008;49:301−16.
  • 8. Joling KJ, O'Dwyer ST, Hertogh CMPM, et al. The occurrence and persistence of thoughts of suicide, self-harm and death in family caregivers of people with dementia: a longitudinal data analysis over 2 years. Int J Geriatr Psychiatry 2018;33:263−70.
  • 9. Song SY, Lee YP. A study on the factors affecting suicidal ideation in people with mental disabilities. J Korea Contents Assoc 2021;21:765−75.
  • 10. Anderson CS, Linto J, Stewart-Wynne EG. A population-based assessment of the impact and burden of caregiving for long-term stroke survivors. Stroke 1995;26:843−9.
  • 11. Huang YC, Hsu ST, Hung CF, et al. Mental health of caregivers of individuals with disabilities: relation to suicidal ideation. Compr Psychiatry 2018;81:22−7.
  • 12. Marlow NM, Xie Z, Tanner R, et al. Association between disability and suicide-related outcomes among U.S. adults. Am J Prev Med 2021;61:852−62.
  • 13. Russell D, Turner RJ, Joiner TE. Physical disability and suicidal ideation: a community-based study of risk/protective factors for suicidal thoughts. Suicide Life Threat Behav 2009;39:440−51.
  • 14. Lee SI. The integrated approach of the social and psychological factors affecting suicidal ideation. Democr Policy Res 2016;30:104−39. Korean.
  • 15. Khazem LR. Physical disability and suicide: recent advancements in understanding and future directions for consideration. Curr Opin Psychol 2018;22:18−22.
  • 16. Turner RJ, Noh S. Physical disability and depression: a longitudinal analysis. J Health Soc Behav 1988;29:23−37.
  • 17. Lee JH. The analysis of influencing factors on suicidal ideation cluster of households with disabled members in different types of disabilities. J Disabil Welf 2023;60:5−36. Korean.
  • 18. Lee MK. Care burdens, depression and suicidal ideation among family caregivers of persons with disabilities: a focus on the moderated mediation effects of care services. J Disabil Welf 2019;44:121−48. Korean.
  • 19. Tibshirani R. Regression shrinkage and selection via the LASSO. J R Stat Soc Series B Stat Methodol 1996;58:267−88.
  • 20. Thongpeth W, Lim A, Wongpairin A, et al. Comparison of linear, penalized linear and machine learning models predicting hospital visit costs from chronic disease in Thailand. Inf Med Unlocked 2021;26:100769.
  • 21. van Vuuren CL, van Mens K, de Beurs D, et al. Comparing machine learning to a rule-based approach for predicting suicidal behavior among adolescents: results from a longitudinal population-based survey. J Affect Disord 2021;295:1415−20.
  • 22. Pearlin LI, Lieberman MA, Menaghan EG, et al. The stress process. J Health Soc Behav 1981;22:337−56.
  • 23. Bronfenbrenner U. The ecology of human development: experiments by nature and design. Harvard University Press; 1979.
  • 24. Lee JH, Huber JC. Evaluation of multiple imputation with large proportions of missing data: how much is too much? Iran J Public Health 2021;50:1372−80.
  • 25. Kim JS, Yoon TK, Lim HS. Eco-systemic variables affecting on suicidal ideation of mothers of children with disability. J Educ Welf Disabil 2021;7:51−74. Korean.
  • 26. Seo WS, Seo WY, Sung MS. A study on institutional improvements for suicide prevention among people with disabilities and their families [Internet]. Korea Disabled People’s Development Institute; 2017 [cited 2025 Jun 24]. Available from: https://www.koddi.or.kr/data/research01_view.jsp?brdNum=7404008&brdTp=&searchParamUrl=. Korean.
  • 27. da Silva Gdel G, Jansen K, Barbosa LP, et al. Burden and related factors in caregivers of young adults presenting bipolar and unipolar mood disorder. Int J Soc Psychiatry 2014;60:396−402.
  • 28. Hosmer Jr DW, Lemeshow S, Sturdivant RX. Applied logistic regression. John Wiley & Sons; 2013.
  • 29. Lee J, Pak TY. Machine learning prediction of suicidal ideation, planning, and attempt among Korean adults: a population-based study. SSM Popul Health 2022;19:101231.
  • 30. Gould MS, Wallenstein S, Davidson L. Suicide clusters: a critical review. Suicide Life Threat Behav 1989;19:17−29.
  • 31. Joiner Jr TE. The clustering and contagion of suicide. Curr Dir Psychol Sci 1999;8:89−92.
  • 32. Hawton K, Lascelles K, Stewart A, et al. Identifying and responding to suicide clusters and contagion [Internet]. Public Health England; 2015 [cited 2025 Jun 22]. Available from: https://cronfa.swan.ac.uk/Record/cronfa23339.
  • 33. Rossom RC, Coleman KJ, Ahmedani BK, et al. Suicidal ideation reported on the PHQ9 and risk of suicidal behavior across age groups. J Affect Disord 2017;215:77−84.
  • 34. Navarro MC, Ouellet-Morin I, Geoffroy MC, et al. Machine learning assessment of early life factors predicting suicide attempt in adolescence or young adulthood. JAMA Netw Open 2021;4:e211450.
  • 35. Park BG, Song IS. The relationship between stress and suicidal ideation for old adult living alone: multiple mediator effects of self-criticism and feelings of loneliness, and the moderating effects of social support. J Soc Welf 2014;66:51−74. Korean.

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

Include:

A machine learning approach for predicting suicidal ideation among family members of persons with disabilities: a cross-sectional study in the Republic of Korea
Osong Public Health Res Perspect. 2025;16(6):560-574.   Published online December 11, 2025
DOI: https://doi.org/10.24171/j.phrp.2025.0261
Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.
Format:
Include:
A machine learning approach for predicting suicidal ideation among family members of persons with disabilities: a cross-sectional study in the Republic of Korea
Osong Public Health Res Perspect. 2025;16(6):560-574.   Published online December 11, 2025
DOI: https://doi.org/10.24171/j.phrp.2025.0261
Close

Figure

Related articles

PHRP : Osong Public Health and Research Perspectives

About

Browse articles 

Editorial policies

For authors  

phrp 사이트 QRCODE(URL : https://ophrp.org)
© Korea Disease Control and Prevention Agency.