Bipolar disorder is a severe psychiatric condition with a low diagnostic rate, characterized by recurrent episodes of depression and mania (1). The depressive state is marked by persistent sadness and anhedonia, and the associated negative emotions can lead to suicidal ideation and even suicide attempts. Consequently, bipolar disorder is associated with a high suicide rate. In a study of 150 patients with bipolar disorder from Southeast Asia, Biazus et al. reported that 35 individuals (23.3%) had attempted suicide (2). A meta-analysis examining the prevalence of suicide attempts in patients with bipolar disorder found that the lifetime prevalence was 33.9%. The high frequency of suicide attempts among individuals with bipolar disorder has significant implications for public health.

In recent decades, numerous studies have focused on predicting suicidal behavior in bipolar disorder, highlighting suicidal ideation as a prominent predictive indicator. Valtonen et al. observed, in a cross-sectional study involving 191 patients diagnosed with bipolar disorder, that all individuals who had attempted suicide reported experiencing suicidal ideation (3). Johnson et al. demonstrated that depression, mixed episodes, and hopelessness are predictive factors for both suicidal ideation and suicide attempts; all these factors contribute to an increased risk of suicide in bipolar disorder (4). Extensive literature has documented a strong association between a history of suicide attempts and subsequent suicidal behavior. Oquendo et al. reported that individuals with a prior suicide attempt were four times more likely to engage in future attempts (5). Factors such as aggression, hostility, and despair have also been identified as predictors of suicidal behavior. Despair is an independent risk factor for attempted suicide (6). Furthermore, studies have indicated that gender, age, family history, and childhood abuse experiences can serve as predictive factors for suicide attempts among patients with bipolar disorder (7, 8).

In the field of machine learning—a branch of computer science—computers analyze given data to predict the next task at hand. Machine learning algorithms are specifically designed to learn from data and make accurate predictions (9). Ryu et al. used random forests modeling techniques to achieve a high prediction accuracy of 88.9% in predicting suicide attempts among patients with suicidal ideation; their model yielded an area under the receiver operating characteristic curve (AUC) of 0.947 (10). Similarly, Oh et al. used a customized artificial neural network classifier to predict suicide attempts in patients with depression and anxiety disorders, achieving overall accuracies of 93.7%, 90.8%, and 87.4% for detecting one-month, one-year, and lifetime suicide attempts, respectively (11). A machine learning study investigating cognition aimed to predict suicide risk in patients with major depressive disorder, achieving a prediction accuracy of 70.6% using XGBoost-2 (12).

Although machine learning has been widely applied to predict suicide behavior in patients with depression, studies utilizing this approach specifically for bipolar disorder remain relatively scarce. Notably, Pigoni et al. recently employed multimodal machine learning (clinical + neuroimaging features) to predict 12-month suicide attempts in bipolar disorder (13); however, biological markers—including thyroid function and biochemical indicators—were not incorporated into their predictive model. Other existing studies have primarily focused on using machine learning to assist in the diagnosis of bipolar disorder or to predict other disease outcomes, rather than to explicitly model suicide risk (14–16).

Contemporary machine learning has advanced beyond traditional algorithms, with newer approaches demonstrating exceptional performance in modeling complex, nonlinear relationships in biomedical data. Neural networks (including deep learning) excel at capturing intricate patterns in high-dimensional clinical data, while Gaussian process regression offers advantages in uncertainty quantification, particularly for small-sample settings. Graphical techniques (e.g., Bayesian networks) can elucidate conditional dependencies among predictors, and ensemble or composite methods that hybridize multiple base learners often outperform single-algorithm approaches. For this initial study, we selected three established algorithms (gradient boosting, random forests, support vector machines) due to their robustness, interpretability, and widespread validation in clinical prediction studies—critical considerations for translating findings into preliminary clinical insights. A pilot comparison with a shallow neural network confirmed that random forests achieved competitive performance (AUC = 0.958 vs. 0.946) while maintaining superior interpretability, justifying our model choice.

In this study, therefore, we build upon the emerging literature by employing machine learning methods to predict suicide attempts in patients with bipolar disorder, incorporating demographic characteristics, physiological indicators (including thyroid function), and psychometric scale scores as key predictive features. Our contribution is threefold: (1) we expand the feature space by integrating routinely available biochemical and endocrine markers; (2) we apply systematic hyper-parameter tuning and class-imbalance handling (SMOTE) to optimize model performance; and (3) we externally validate the best-performing algorithm (random forests) through rigorous cross-validation and calibration assessment.

This study used machine learning to predict suicide-related risk in patients with bipolar disorder. In the study, we included demographic characteristics, physiological indicators, and psychometric scale scores. We tested the predictive performance of three learners: GBM, RF, and SVM. Although there were differences in the machine learning techniques used, it was confirmed that each machine learning technique could effectively predict suicide attempts. Five important factors were found to predict suicide in patients with bipolar disorder, including suicidal ideation, education level, two dimensions of hopelessness and retardation in HAMD, and the immune index TSH.

By comparing the three learners, we confirmed that RF, an integration-based machine learning learner as described above, has the best predictive accuracy. In previous studies, RF was found to be an excellent learner for predicting suicide risk. In a study on the detection of suicide risk by MPI-2 based on machine learning, the k-nearest neighbor algorithm and the RF learner were selected. RF was reported to predict suicidal ideation and suicide attempts with 92.9% and 95% accuracy, respectively (24). In a study using machine learning to predict suicide attempts in people with suicidal ideation, the RF showed good performance (AUC of 0.947), with an accuracy of 88.9% (10). In summary, RF has been proven to have excellent accuracy in predicting suicide behavior. In this study, RF also served as a good learner with high accuracy in predicting suicide attempts in patients with bipolar disorder, as well as preadolescent children with a history of suicidal ideation reporting a history of suicide attempts (25). In an eight-year follow-up study of individuals with affective disorders, including bipolar disorder, depression, and schizoaffective disorder, the researchers found that suicidal ideation was not a significant predictor of suicidal behavior within one year, but was significantly associated with suicide attempts over the long term (26). O’Connor and Kirtley proposed the integrated motivational–volitional (IMV) theory for suicidal behavior, which indicates that defeat and entrapment drive the emergence of suicidal ideation and that a group of factors, entitled volitional moderators (VMs), govern the transition from suicidal ideation to suicidal behavior (27).

Consistent with previous studies, suicidal ideation appears as a high predictor of suicide attempts in the RF model.

Education in the RF model is another predictor for suicide attempts. Previous studies have highlighted the significant relationship between low education levels and suicide attempts (28). This may be due to the lack of social and family support in patients with low education, which has been systematically reviewed as a risk factor for suicide attempts in bipolar disorder (29). The low level of education in patients with bipolar disorder may mean that the overall level of education of the family is not high, and that therefore the family lacks understanding and cannot provide support for these patients. The absence of family members may tend to lead to negative emotions and suicide attempts.

Hopelessness is an important factor that triggers suicide attempts and has a predictive effect on suicidal behavior (5). A 10-year prospective study highlighted that the level of hopelessness was the only score that distinguished participants with suicidal ideation from suicide attempts (30). Feelings of hopelessness were closely related to suicidal behavior one year later in patients with affective disorders (26). We believe that individuals may experience negative emotions and thoughts after experiencing various kinds of pain, but when they still feel hopeful about life, they will not make the final suicide attempt. Once a sense of despair arises, they will feel that any measures are meaningless and choose suicide as the only way forward. According to the three-step theory of suicide, there are four influential factors in the process from the generation of suicidal ideation to suicide attempts, which are pain, despair, connection, and suicidal ability. Hopelessness is required for the development of suicidal ideation, and pain and hopelessness are the two most common motivations for suicide attempts (31). However, other studies have found that between hopelessness and suicidal ideation, individuals’ attitudes towards suicide play a potential moderating role. In other words, at the level of despair, people who are more likely to commit suicide are those who think suicide is an acceptable behavior. Moreover, there seems to be a difference in gender, and the attitude towards suicide seems to have a greater impact on suicidal ideation among women (32). Retardation here refers to psychomotor retardation, which is specifically manifested as poor and dull brain thinking, slow body movement, and slow thinking processes. Previous studies have shown that psychomotor retardation is associated with a greater number of depressive episodes (33). Psychomotor delay is associated with more depressive episodes, producing negative emotions, feelings of pain, suicidal ideation, and finally stimulating individuals to commit suicide attempts.

TSH is also a predictor for suicide attempts in patients with bipolar disorder. The thyroid gland is the largest endocrine organ in the human body, and thyroid hormone secreted by this gland has an impact on individuals’ emotions and cognitive processes. Excessive secretion of thyroid hormone will cause excessive mood swings and may lead to emotional problems such as anxiety, irritability, and depression. In the hypothalamic–pituitary–thyroid axis, TSH has a regulatory effect on thyroid hormone level, which is manifested in that an increase in TSH level will promote secretion of thyroid hormone. Therefore, abnormal TSH values may be accompanied by a series of mental disorders. At the same time, the relationship between the TSH index and suicide attempts has been widely studied in other mental diseases, especially in depression. Many previous studies have shown a correlation between suicide attempts and abnormal TSH level in depressed patients. One study found that TSH in depressed patients is independently correlated with suicide attempts (34). Another study explored the association between TSH and suicide in people with comorbid depression and anxiety and found that patients with comorbid anxiety (MDA) who had attempted suicide had higher TSH levels (35). However, few studies have explored the relationship between TSH level and suicide attempts in patients with bipolar disorder. Our study found that TSH can be identified as a predictor of suicide risk in patients with bipolar disorder through machine learning.

While machine learning models offer powerful tools for pattern recognition, their “black-box” nature can limit adoption in real-world clinical practice. To bridge this gap, we derived a simplified, interpretable risk scoring system based on the top predictive features—suicidal ideation, hopelessness, psychomotor retardation, lower educational attainment, and abnormal TSH levels—that demonstrated excellent discrimination and calibration in internal validation. This score categorizes patients into low, moderate, and high-risk groups, providing clinicians with actionable thresholds for monitoring and early intervention.

Notably, the inclusion of TSH—a readily available biological marker—adds objective, quantifiable value to traditionally subjective psychometric assessments. In combination with psychological factors such as hopelessness and suicidal ideation, this integrated approach may help identify individuals who appear stable but harbor elevated biological risk. The accompanying nomogram allows for rapid visual estimation of suicide attempt risk at the point of care, even without computational tools.

Although this scoring system was derived from a cross-sectional sample and requires external prospective validation before routine implementation, it represents a practical step toward translating complex machine learning findings into clinically usable instruments. Future studies should evaluate whether using such a tool improves risk assessment accuracy, clinician decision-making, and ultimately, patient outcomes.

It is important to acknowledge the limitations of this study. The cross-sectional design means that the model captures associations rather than true predictive relationships, and the high performance metrics may be influenced by the single-center retrospective nature of the data. External validation in diverse populations is essential to confirm the generalizability of the findings.

We developed and internally validated a parsimonious, yet robust, machine-learning model that quantifies lifetime suicide-attempt risk in adult patients with bipolar disorder by using routinely available demographic, endocrine and psychometric variables. Random forests achieved excellent discrimination (AUC 0.962) and calibration (slope 1.04, intercept 0.02), outperforming gradient boosting and support vector machine algorithms. Suicidal ideation, education level, hopelessness, psychomotor retardation and thyroid-stimulating hormone emerged as the top predictors, all of which are clinically actionable and can be captured within a single outpatient visit.

The model is ready for prospective implementation studies in psychiatric hospitals and community clinics to test its ability to flag high-risk individuals who might benefit from enhanced monitoring, safety-planning and early intervention. Integration into electronic medical record systems could facilitate automated risk alerts without additional staff burden.

While our model achieved high discrimination (AUC = 0.962), clinicians should be cautioned that this performance reflects cross-sectional associations rather than true predictive ability. The retrospective assessment of lifetime suicide attempts means that biomarker measurements may reflect consequences rather than causes of suicidal behavior. This model should not be used for individual risk prediction in clinical practice until validated in prospective cohorts with temporally well-defined assessments.” Moreover, although the high performance of the present research model partially reflects the lifetime risk of suicide attempts among local adult patients with bipolar disorder, the limitations inherent to a single-center sample may restrict its generalizability to different environments, ethnicities, nationalities, and healthcare systems. Therefore, it is recommended that future research conduct external validation in multi-center, cross-cohort settings as a prerequisite for clinical implementation.

Cross-sectional design: lifetime suicide attempts were ascertained retrospectively, precluding determination of temporal directionality between predictors and outcome; Single-center, retrospective sample may limit generalizability; external validation in multi-site, ethnically diverse cohorts is required; The lack of ecological effects and the small number of events observed in the class-unbalanced data may be another limitation of the study (36), despite SMOTE, the event rate (31%) remains modest; rare-event correction techniques should be explored in future work; We did not capture potentially relevant factors such as childhood trauma granularity, medication adherence or recent life events; The model predicts lifetime rather than prospective suicide attempts; future studies should define a clear forward-looking time window (e.g., 6- or 12-month risk) and link predictions to subsequent clinical contacts.