1. This cross-sectional study developed and compared nine machine learning models to predict suicidal ideation in 292 patients with bipolar disorder.

2. The random forest model demonstrated excellent predictive performance (area under the curve (AUC): 0.915) for identifying high-suicide-risk bipolar patients.

3. Quality of life emerged as the most significant predictor, followed by the number of depressive episodes, a history of suicide attempts, cognitive functioning (both subjective complaints and objective performance), and emotional abuse in childhood trauma.

Suicide, defined as death caused by intentional self-directed harm, remains a leading cause of mortality worldwide, accounting for more than 700,000 deaths annually [1]. Suicidal ideation (SI), a precursor to suicidal behavior, involves self-reported inclinations toward self-harm without preparatory actions [2]. Recognized as the nascent phase in the progression toward suicide, SI is deemed a prerequisite for suicidal behavior and serves as a key predictor of future suicidal attempts (SAs), necessitating early intervention [3]. The risk of suicide is extraordinarily high among patients diagnosed with bipolar disorder (BD). Indeed, among psychiatric disorders, BD is linked to the highest suicide rates [4, 5, 6]. This risk is estimated to be 30–60 times greater than that observed in the general population [4, 5, 6]. Studies indicate that approximately 30%–50% of adult patients with BD attempt suicide, and approximately 15%–20% ultimately die by suicide [4, 5, 6]. SI is prevalent in patients with BD and appears to be a risk factor for completed suicide. Approximately 59% of patients with BD report SI, a proportion that is 20–30 times greater than that in the general population [7]. However, some studies conflate SI with SAs, limiting its clinical utility [8]. This conflation obscures distinct intervention pathways. Therefore, precisely identifying and addressing the risk factors for SI specifically among patients with BD are critical public health priorities for targeted prevention and intervention before suicidal behaviors escalate.

Prior studies have identified multiple risk factors for SI in patients with BD, such as a history of SAs, younger age, younger age at onset of illness, more severe depressive or manic symptoms, BD type II/nos, higher levels of insight and impulsiveness, concurrent hopelessness, a history of childhood trauma, neurocognitive dysfunction, functional impairment, a poor or dysfunctional family environment, and low psychological resilience, social support and quality of life (QoL) [9, 10, 11, 12, 13, 14, 15]. Crucially, for patients with BD, a previous SAs can powerfully predict future SI [13, 15], with 40% of suicide deaths occurring subsequent to such an attempt [15]. The risk of SI is particularly pronounced among younger patients with BD, especially within the first few years following the onset of the illness [15]. Aaltonen et al. [12] reported that younger age, severe depressive symptoms, BD type II/nos, hopelessness, and childhood physical abuse can independently predict SI in individuals with BD. Moreover, other studies have demonstrated that compared with manic and hypomanic phases, mixed and depressive episodes of BD are associated with a higher likelihood of SI [16]. Our previous research revealed that depressive symptoms and cognitive functioning are predictors of SI in patients with BD [14]. Furthermore, adverse family dynamics, such as excessively expressed emotions, reduced cohesion, and increased conflict, are correlated with an elevated risk of SI and SAs [10]. Notably, SI can mediate the relationship between family and psychosocial functioning [9]. Given the complex interplay among these risk factors for SI in patients with BD, the development of clinically actionable prediction tools is critical for enabling early interventions to reduce preventable suicide deaths.

Although existing suicide prediction models have achieved acceptable classification accuracy, their predictive utility remains constrained by methodological challenges. Traditional approaches often rely on univariate screening followed by logistic regression, which risk oversimplifying the complex, interactive nature of psychopathological data. These methods may introduce multicollinearity or overfitting while failing to capture the nonlinear dynamics inherent in high-dimensional clinical datasets [17]. Furthermore, they typically lack the ability to quantify the individual contribution of each risk factor to the outcome. Consequently, advanced analytical techniques that can more effectively integrate these multifaceted predictors are needed to develop robust clinical prediction tools. Machine learning (ML) is expected to address these constraints by algorithmically detecting multidimensional patterns within high-dimensional data [18]. By circumventing linearity assumptions and automatic feature weighting, ML techniques identify clinically significant variables while enhancing predictive accuracy, offering a methodological advance for SI risk modeling [18].

ML has emerged as a highly promising tool in psychiatric research because of its robust ability to integrate multidimensional predictors and classify outcomes with heightened precision [19]. Ensemble methods and deep learning architectures model complex nonlinear interactions and adapt to subtle data variability within high-dimensional clinical-behavioral data by capturing patterns eluding conventional approaches [20]. In addition, automated feature selection and scalable pattern recognition can efficiently address the complex risk factors associated with suicidality [17]. Furthermore, ML natively handles datasets comprising mixed feature types (continuous and categorical), which aligns with our predictor set comprising scales, counts, and binary indicators. The ensemble structure of methods such as Random Forest (RF), built on bagging and random feature selection, further confers robustness to noise and missing values. These techniques not only increase the precision of prediction models but also generate clinically actionable insights to guide targeted interventions and personalized health care.

ML models have been applied to predict SI or SAs in diverse populations. For example, Deng et al. (2025) [21] constructed risk prediction models for SAs in patients with mood disorders using the RF method, which exhibited good discriminant ability, stability, and calibration. In addition, ML algorithms have been utilized to predict SI among university students in Bangladesh (with more than 90% accuracy) [22] and among youth with autism spectrum disorder [23], demonstrating the potential of data-driven approaches for informing precision prevention strategies and guiding tailored support for high-risk subgroups.

To the best of our knowledge, research using ML models to predict SI among patients with BD is nearly nonexistent. Thus, this study aimed to develop ML-based prediction models for SI through multidimensional assessments of psychiatric, psychological, and sociodemographic variables in patients with BD.

In this study, nine ML models were developed and compared to predict the risk of SI in patients with BD. The RF model demonstrated superior performance, achieving an AUC of 0.915 with a classification accuracy of 81.8%. Through interpretable SHAP analysis, QoL emerged as the most salient predictor, followed by the number of depressive episodes, history of SAs, cognitive functioning, and emotional abuse. These findings underscore the multifaceted nature of SI risk in BD.

Our results align with existing evidence on SI risk factors in BD [15, 44]. Machine learning approaches have also been successfully applied to identify suicide risk in other populations, such as veterans, by leveraging cross-modal interactions among psychosocial factors [45]. The multimodal ML model developed by Pigoni et al. [44] focused on predicting SAs within a 12-month period and found that combining clinical features with gray matter volume significantly improved predictive sensitivity (reaching 80%). This highlights the important supplementary value in predicting suicidal behavior of neurostructural features, particularly in the prefrontal-temporal-cerebellar regions. Although the present study did not include neuroimaging data, it also achieved high predictive accuracy for SI by incorporating multidimensional features such as psychosocial factors, cognitive functioning, and biomarkers, further supporting the necessity of integrating multidimensional information in suicide risk prediction. In addition, the theoretical model of suicide proposed by Malhi et al. [15] emphasizes that the transition from ideation to action involves multi-stage mechanisms such as motivation and volition. The importance of factors such as previous SAs and cognitive functioning in our findings corresponds to elements like acquired capability and cognitive appraisal in Malhi’s [15] model, while childhood trauma and QoL reflect the foundational roles of early stress and current psychosocial functioning in the formation of SI. Furthermore, Deng et al. [21] also demonstrated the utility of ML and their model achieved an AUC of 0.741 in testing and 0.788 in external validation. Notably, their SHAP analysis identified polarity (i.e., depressive vs. manic episode) and previous SAs as the most influential predictors, which resonates with the “acquired capability” component in Malhi’s model [15] and underscores the transdiagnostic relevance of these factors. Although Deng et al. [21] did not incorporate neuroimaging biomarkers, their model’s strong performance further validates the importance of integrating multidimensional clinical data for suicide risk prediction. In summary, by focusing on the earlier risk stage of SI, this study expands the application of ML in suicide risk prediction in BD and establishes strong theoretical and empirical connections with existing research.

Through feature importance analysis, QoL was determined to be a paramount predictor of SI. These findings are consistent with those of a study by Mazaheri et al. [46] involving 140 inpatients with BD type I, which demonstrated an inverse association between QoL and suicide risk. This suggests that the subjective burden of illness, encompassing functional impairment and psychosocial well-being, may be a more immediate driver of suicidal thinking than some traditional clinical metrics are. It is widely recognized that the direct disease burden of BD significantly compromises QoL through symptomatic manifestations and functional limitations. In reality, diminished QoL also serves as a clinical indicator of compromised psychosocial resources, such as inadequate social support networks and reduced cognitive reserves, all of which collectively contribute to the development of SI [47].

The number of depressive episodes was another critical predictor of SI. These findings concur with prior research demonstrating that increased depressive episodes correlate with increased suicide risk in patients with BD [48]. Recurrent depressive episodes can exacerbate hopelessness, helplessness, and perceived burdensomeness, which are well-established psychological precursors to suicide. In addition to illness characteristics, a history of SAs was also confirmed as a pivotal predictor, which is consistent with the literature [49]. For example, Jiang et al. [50] reported that a history of SAs was a significant predictor of suicide risk in both AdaBoost and binary logistic regression algorithms. This heightened vulnerability may arise from acquired familiarity with suicidal means and lowered aversion to death among individuals with prior SAs. With respect to developmental trajectories, younger age correlated significantly with SI, which is consistent with the literature reporting elevated suicide risk specifically among BD patients aged 35 years [51]. This association likely reflects, in part, the confluence of earlier illness onset and underdeveloped coping mechanisms coupled with limited psychosocial resources in younger individuals. Furthermore, the initial period following diagnosis represents a high-risk window for suicidal behaviors [51], offering a plausible explanation for the increased SI prevalence observed in the younger patients in this study.

Our results reinforce the evidence linking both subjective and objective cognitive deficits to SI. Neurobiologically, cognitive impairments are correlated with functional alterations in the ventral and dorsolateral prefrontal cortices, with cognitive inhibition impairment specifically recognized as a neurocognitive marker for suicide prediction [52]. With respect to the relationship between subjective and objective measures, the current evidence indicates limited convergence. Some studies propose subjective complaints as harbingers of subsequent objective decline [53], whereas others report no significant association between the two [54]. Cognitive deficits may exacerbate depressive symptoms, and depressive episodes may amplify cognitive dysfunction, collectively increasing suicide risk [14, 55]. For instance, executive dysfunction can lead to deficient inhibitory control, impulsive decision-making, emotional dysregulation, and compromised problem-solving abilities, fostering helplessness and intensifying depressive effects. Conversely, recurrent depressive episodes may further decrease cognitive capacity and the ability to suppress intrusive suicidal thoughts.

The significant contribution of emotional abuse from childhood trauma is consistent with the established literature [56]. Exposure to such trauma may foster maladaptive cognitive schemas, heightening an individual’s susceptibility to hopelessness during stressful periods and subsequently increasing their vulnerability to SI. Evidence indicates a potential dose‒response association between the extent of emotional abuse exposure and the severity of BD [57], further contextualizing its clinical relevance. Against this backdrop of vulnerability, resilience was recognized as a protective predictor against SI in our analysis, corroborating the established literature [58]. Individuals with greater resilience may preferentially adopt adaptive coping strategies (such as problem solving), thereby disrupting the progression from hopelessness to SI. An intriguing phenomenon was observed from the methodological perspective. Variables that were not statistically significant in the univariate analysis (e.g., self-esteem) exhibited relatively high $|$SHAP$|$ values in the RF model. This divergence likely reflects the inherent constraints of univariate approaches, which evaluate variables linearly and independently, overlooking the potential intricate interactions among variables. In contrast, the ensemble decision trees of the RF model can capture nonlinear associations and higher-order variable interactions, potentially elucidating the nuanced relationship between self-esteem and SI.

In the present study, social support was identified as a significant predictor of SI. Existing evidence indicates divergent mechanistic pathways. Tamizi et al. [59] demonstrated a direct association, in which social support independently accounted for 33.3% of the variance in SI among Malaysian university students. Conversely, Owen et al.’s [60] four-month longitudinal study of 80 patients with BD revealed a more complex chain mediation pathway in which low social support leads to feelings of defeat, fosters entrapment cognition, and subsequently amplifies hopelessness, ultimately increasing the risk of SI.

Biochemical markers, such as TG levels, Ca²⁺ levels, and RBC counts, demonstrated the predictive potential of the SI for BD. However, mechanistic interpretations necessitate nuanced contextualization. With respect to TG levels, our analysis revealed significantly lower concentrations in SI patients, potentially reflecting disease-associated appetite suppression and decreased nutritional intake. Conversely, Liu et al. [61] reported that elevated TG levels ($>$1.7 mmol/L) were associated with an increased risk of SI (odds ratio = 2.58). This disparity can be attributed to population differences. Specifically, the latter study concentrated on children and adolescents with metabolic syndrome, which may heighten cardiovascular risks and, subsequently, vulnerability to SI. In addition to lipid metabolism, patients with SI exhibited reduced Ca²⁺ levels. This phenomenon may be attributed to lithium-induced inositol depletion and calcium channel blockade, both of which lead to a reduction in the intracellular Ca²⁺ concentration. This can potentially indicate illness chronicity and treatment duration. Concurrently, the decreased RBC counts in patients with SI align with previous findings [62], which could be ascribed to chronic low-grade inflammation in BD that suppresses erythropoiesis and shortens the lifespan of erythrocytes. Future research should explore additional biomarkers, such as neurotransmitters and inflammatory factors. For example, Li et al. [55] developed an optimized prediction model (AUC = 0.91) that integrated serum brain-derived neurotrophic factor (BDNF) levels, cognitive functioning, depressive symptoms, and sleep quality to predict SI in patients with bipolar depressive disorder.

In this study, nine ML models for predicting SI in patients with BD were developed and compared, with the RF model demonstrating the optimal performance. Interpretability analyses revealed that QoL, the number of depressive episodes, history of SAs, cognitive functioning, and childhood emotional abuse were the most influential predictors. This study, which combined hybrid feature selection with explainable ML, provides a validated framework for assessing suicide risk and highlights specific targets for clinical intervention. However, the model requires further external validation in prospective, real-world cohorts to confirm its efficacy and generalizability before it can be recommended for routine clinical use. Future research should focus on longitudinal designs to establish causal inference and on integrating these models into collaborative care settings to test their impact on patient outcomes.

SI, suicidal ideation; BD, bipolar disorder; SAs, suicide attempts; ML, machine learning; HDRS-17, 17-item Hamilton depression rating scale; YMRS, Young mania rating scale; BMI, body mass index; WHR, waist-hip ratio; RBC, red blood cell; WBC, white blood cell; HGB, hemoglobin; AST, glutamic oxaloacetic transaminase; ALT, glutamic-pyruvic transaminase; ALB, albumin; TP, total protein; GLO, globulin; TBIL, total bilirubin; BG, blood glucose; CHO, cholesterol; TG, triglycerides; HDL, high-density lipoprotein cholesterol; LDL, low-density lipoprotein cholesterol; APOA1, apolipoprotein A1; APOB, apolipoprotein B; COBRA, cognitive complaints in bipolar disorder rating assessment; RBD, resilience questionnaire for bipolar disorder; RSES, rosenberg self-esteem scale; CTQ, childhood trauma questionnaire; SSRS, social support rating scale; FAD, family assessment device; QoL, quality of life; SF-12, 12-item short form health survey; LR, logistic regression; XGBoost, extreme gradient boosting; LightGBM, light gradient boosting machine; CatBoost, categorical boosting; GBM, gradient boosting machine; AdaBoost, adaptive boosting; ET, extremely randomized trees; SVC, support vector classification; ROC, receiver operating characteristic; AUC-ROC, area under the receiver operating characteristic curve; MCC, Matthews correlation coefficient; PR, precision-recall; TRIPOD, transparent reporting of a multivariable prediction model for individual prognosis or diagnosis.

The data and materials used in this study are available from the corresponding author upon reasonable request.

Conception–XLuo, XLin; Design–XLuo, XLin, QZ; Supervision–XLin, QZ; Materials—XY, TZ, DL, QS; Fundings–XLin; Data Collection and/or Processing–XLuo, SM, XY, TZ, CZ; Processing –XLuo, TZ, CZ; Analysis and/or Interpretation–XLuo, QS, DL; Literature Review–XLuo, XY, CZ; Writing–XLuo, XLin; Critical Review–XLuo, QZ, CZ, DL. All authors have read and approved the final manuscript. All authors contributed to editorial changes in the manuscript. All authors participated sufficiently in the work and agreed to be accountable for all aspects of the work.

This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Sun Yat-Sen University (ethics approval number: 2019ZSLYEC-021). All participants or their legal guardians provided signed informed consent.

We gratefully acknowledge Ms. Shizhu Li for her assistance and instructions regarding the statistics used in this study.

This research was funded by the Postdoctoral Fellowship Program of CPSF (grant number: GZC20242135), the Fujian Province Science and Technology Project (grant number: 2025Y9768), the Special Funding for Postdoctoral Research Projects of Chongqing (grant number: 2023CQBSHTB3107), and Nursing Collaborative Innovation, Chongqing Medical University (grant number: 20240202).

The authors declare no conflict of interest.

During the preparation of this work, the authors used DeepSeek-R1 to check the spell and grammar. After using this tool, the authors reviewed and edited the content as needed, and took full responsibility for the content of the publication.

Supplementary material associated with this article can be found, in the online version, at https://doi.org/10.31083/AP45589.