Nonsuicidal self-injury, often abbreviated NSSI, refers to deliberate destruction of body tissue—cutting, burning, or hitting oneself—in the absence of suicidal intent. It is far from rare in adolescence and early adulthood, and decades of research show that people who self-injure carry a markedly elevated risk of eventually attempting suicide. The difficulty for clinicians is that this elevated risk is not uniform. Self-injury interacts with depression, poor emotion regulation, low distress tolerance, weakened resilience, and a host of demographic and social factors in ways that simple additive checklists struggle to capture. The transition from self-harm to suicidal behavior is thought to unfold gradually, as repeated injury erodes the fear of pain and death and builds what theorists call the acquired capability for suicide.

To untangle this web, the researchers recruited 601 South Korean adults aged 19 to 29 who reported at least five lifetime episodes of nonsuicidal self-injury, gathered through an online survey. Five or more episodes is a threshold that clinical frameworks treat as clinically significant, ensuring the sample represented a genuinely high-risk population rather than occasional self-harm. Participants provided information spanning sociodemographics, psychosocial circumstances, psychiatric diagnoses, psychological traits such as emotion regulation, distress tolerance, and resilience, and the characteristics of their self-injury itself—including how often they had harmed themselves and how many different methods they had used, a property known as versatility. The outcome the models were trained to classify was simple and stark: whether the participant had ever made a suicide attempt.

The team compared three widely used machine learning classifiers: a support vector machine, a random forest, and eXtreme Gradient Boosting, commonly called XGBoost. Support vector machines work by finding the optimal boundary that separates two classes of data points in a high-dimensional space; random forests pool the votes of hundreds of decision trees, each trained on a random slice of the data; and gradient boosting builds trees sequentially, with each new tree correcting the errors of the last. In the baseline comparison, the support vector machine came out ahead, achieving a mean area under the receiver operating characteristic curve—an AUROC—of 0.811. An AUROC of 0.5 would reflect chance-level guessing, while 1.0 reflects perfect discrimination, so a value above 0.8 indicates the model separated attempters from non-attempters substantially better than chance.

The researchers then refined their model using recursive feature elimination with cross-validation, a technique that repeatedly trains the model while shaving off the least informative variables, keeping only the subset that maximizes predictive performance. This process serves a dual purpose: it guards against overfitting, where a model memorizes quirks of the training data rather than learning generalizable patterns, and it distills a long list of candidate variables into a compact, interpretable set. The optimized support vector machine, evaluated on a held-out test set the model had never seen during training, achieved an AUROC of 0.865—a robust result for a psychiatric classification task and one that suggests the model captured genuine signal rather than statistical noise.

Accuracy alone, however, is of limited clinical value if a model operates as an inscrutable black box. To open it up, the team applied Shapley Additive Explanations, or SHAP values, a method borrowed from cooperative game theory that assigns each feature a quantified contribution to each individual prediction. By averaging these contributions across the sample, the researchers could rank every variable by its overall importance in driving the model’s classifications. This transparency step is increasingly regarded as essential in medical machine learning, where a prediction that cannot be explained is a prediction clinicians cannot act on—or trust.

The feature importance analysis delivered a clear message. The single most salient indicator of suicide attempt history was the frequency of nonsuicidal self-injury: the more often participants had harmed themselves, the more strongly the model leaned toward classifying them as having attempted suicide. This finding aligns closely with the interpersonal theory of suicide, which holds that repeated painful and provocative experiences progressively raise a person’s tolerance for physical pain and fear of death, thereby building the capability required for a lethal act. In practical terms, it suggests that the sheer repetition of self-injury is not merely a symptom of distress but an active pathway toward suicidal behavior—one that interventions could potentially interrupt.

Right behind NSSI frequency, two other features rose to the top: the presence of a mood disorder diagnosis, such as major depression or bipolar disorder, and resilience, the psychological capacity to recover from adversity. The pairing is instructive because the two factors point in opposite directions. Active mood pathology amplifies risk, while resilience buffers against it, and their prominence in the model implies that the transition from self-injury to suicide attempt is driven jointly by pathological burden and by the absence of protective resources. The authors argue that this combination carries a direct clinical implication: prevention efforts in high-risk self-injuring populations should simultaneously treat mood symptoms and actively cultivate resilience, rather than addressing either in isolation.

The study’s design carries real strengths. The sample size of 601 is large for research on self-injury, the restriction to participants with five or more episodes ensured a clinically meaningful population, and the broad feature set—spanning demographics, diagnoses, psychological traits, and injury characteristics—gave the algorithms rich material from which to learn. The use of three competing algorithms, cross-validated feature selection, and a held-out test set follows best practices in predictive modeling, and the SHAP-based interpretability analysis addresses the black-box critique head-on. The work was funded by Korea’s Ministry of Health and Welfare and the National Research Foundation of Korea, with the funders playing no role in the design or interpretation of the research.

Certain caveats remain. The data were collected through an online survey and reflect a single time point, so the models classify lifetime suicide attempt history rather than predicting future attempts prospectively—a distinction that matters greatly for clinical deployment. Self-reported histories may be subject to recall bias, and the sample, drawn from young adults in South Korea, may not generalize to other ages or cultural contexts. Even so, the study demonstrates that within one of the highest-risk populations in psychiatry, a handful of measurable features—chief among them how often someone self-injures, whether a mood disorder is present, and how resilient they are—carry most of the predictive weight. If future prospective studies confirm these signals, the approach could give clinicians a data-driven way to identify, among the many young people who self-harm, those who most urgently need preemptive intervention before self-injury hardens into suicidal behavior.

Subject of Research: Machine learning classification of suicide attempt history among young adults with nonsuicidal self-injury

Article Title: Identifying key features of suicide attempts in young adults with nonsuicidal self-injury: a machine learning approach

Article References: Identifying key features of suicide attempts in young adults with nonsuicidal self-injury: a machine learning approach. (n.d.). https://doi.org/10.1186/s12888-026-08668-9

Image Credits: AI Generated

DOI: 10.1186/s12888-026-08668-9

Keywords: nonsuicidal self-injury, suicide attempt, machine learning, support vector machine, SHAP values, mood disorder, resilience, risk factors, young adults, predictive modeling, BMC Psychiatry, suicide prevention