基于机器学习识别心境障碍住院患者自杀行为的模型开发与内部验证

Abstract

Background:

Suicidal behavior (SB) is a major concern in mood disorders. Models developed from retrospective records are vulnerable to ambiguous targets, data leakage, and inappropriate handling of class imbalance. We developed and internally validated a model to identify the risk of suicidal behavior in patients with mood disorders upon admission.

Methods:

The dataset included 1, 099 inpatients with ICD-10 mood disorders admitted to Beijing Anding Hospital in 2022. Exclusion of 20 records with more than 20% row-level missingness and one record without an outcome left 1, 078 patients, including 126 with SB. Current SB comprised documented suicidal ideation, planning, or attempt within the three months before index admission, including the admission assessment. Previous SB and previous suicide attempt (SA) referred to events more than three months before admission. Elastic Net selected 14 predictors from 65 candidates in the development set. We compared 13 classifiers from six model families. All preprocessing and resampling were confined to training folds, and evaluation used only observed patients. Assessment included repeated fivefold cross-validation, probability calibration, bootstrap confidence intervals, precision-recall and decision-curve analyses, clinical baselines, and sensitivity analyses.

Results:

The analytic cohort comprised 952 SB-negative and 126 SB-positive patients (11.69%). The development set contained 864 patients (97 SB-positive), and the observed test set contained 214 patients (29 SB-positive). Calibrated linear discriminant analysis (LDA) provided the most consistent balance of discrimination and calibration. In repeated cross-validation, mean AUC was 0.721 ± 0.052, mean AUPRC was 0.268 ± 0.060, and mean Brier score was 0.095. In the observed test set, AUC was 0.704 (0.597–0.803), AUPRC was 0.283 (0.198–0.453), and Brier score was 0.111. At the training-derived threshold of 0.111, sensitivity was 0.552 and specificity was 0.659. The leading SHAP contributions were onset-polarity category, previous SA, age, and previous SB, but all attributions were associational.

Conclusions:

Model performance was acceptable. These internally validated findings are exploratory and require prospective assessment, standardized outcome measurement, and external validation before clinical use.

1 Introduction

Suicide is a major global public-health problem. The World Health Organization estimates that approximately 727, 000 people die by suicide each year, and many more attempt suicide (). Patients with mood disorders experience a substantial burden of suicidal ideation, planning, and attempts, although estimates vary across diagnostic and clinical settings (, ).

Suicidal behavior in mood disorders reflects demographic, clinical, psychosocial, biological, and health-service factors (–). Previous suicide attempts and other historical indicators have a suggestive effect on the current suicide risk of patients (–). Identifying patients with a high burden of recent suicidal behavior may support timely, structured assessment and closer monitoring during hospitalization.

Previous studies have used registry, demographic, psychometric, genetic, and laboratory data to identify suicide risk (–). Machine-learning models can capture nonlinear associations and interactions, but they can also overfit, produce poorly calibrated probabilities, and be difficult to interpret when outcome events are uncommon (, , ). ROC AUC alone does not show how well a model identifies the minority class, estimates absolute risk, or supports clinical decisions. Evaluation should therefore include precision-recall performance, calibration, uncertainty, and decision-curve analysis, together with comparison against simple clinical baselines (–).

Shapley additive explanations (SHAP) can describe how predictors contribute to individual model outputs (27). Related work in major depressive disorder has combined conventional statistical characterization with interpretable machine learning (28). However, SHAP values describe model associations rather than causal mechanisms, and importance rankings can be unstable in modest samples. Transparent evaluation therefore requires leakage-free preprocessing, explicit handling of class imbalance, calibration, and repeated assessment of explanation stability.

We used a single-center inpatient cohort to examine whether demographic characteristics, clinical history, and routine endocrine and biochemical measurements available around admission could identify patients with suicidal behavior documented during the preceding three months. The clinical aim was to support prioritization for detailed risk assessment at admission. We compared 13 model families and several imbalance strategies, assessed incremental value beyond remote suicide history, and evaluated discrimination, minority-class performance, calibration, uncertainty, net benefit, and explanation stability.

2 Materials and methods

2.1 Study design and participants

This retrospective single-center study used de-identified inpatient medical-record data from Beijing Anding Hospital, Capital Medical University, for calendar year 2022. Eligible patients received an ICD-10 diagnosis within F30–F39, independently assigned by at least two attending psychiatrists. Exclusion criteria were comorbid schizoaffective disorder, personality disorder, intellectual disability, severe physical illness, or substance abuse. Records with more than 20% missing fields were excluded. The study was approved by the institutional ethics committee (R202409190355) and registered retrospectively (ChiCTR2400090041) in 2024, after data collection. Written authorization for anonymized research use of medical-record information was recorded.

The dataset initially contained 1, 099 records. The row-level missingness rule excluded 20 records, and one further record lacked an ascertainable current-SB outcome. Outcomes were not imputed. The analytic cohort therefore included 1, 078 patients, of whom 864 formed the development set and 214 formed the evaluation set. Event proportions were 11.23% and 13.55%, respectively. The partition was treated as random rather than exactly outcome-stratified.

2.2 Outcome and temporal definition

Current SB was defined as suicidal ideation, a suicide plan, or a suicide attempt documented within the previous three months, including at admission. Patients with any of these phenomena were classified as SB-positive; all others with an observed outcome were classified as SB-negative. Previous SB and previous SA were non-overlapping historical predictors and referred to events more than three months before admission. Routine laboratory results were available within 72 hours after admission. The model therefore identifies recent SB at admission and does not estimate prospective risk of a later attempt or suicide death.

2.3 Candidate predictors and missing data

After removal of the identifier and outcome, current suicide attempt and current non-suicidal self-injury were excluded because they referred to the same three-month window as the outcome and could introduce target leakage. Sixty-five candidate predictors remained, covering demographics, prior self-harm, coded diagnosis and onset polarity, illness and hospitalization characteristics, medical comorbidity, and routine laboratory measurements. The de-identified dataset did not retain a reliable mapping for major depressive disorder, bipolar I disorder, and bipolar II disorder. Diagnosis was therefore excluded from the final 14-predictor set, and subtype analyses were not performed.

Onset polarity was defined as the polarity of the first documented affective episode and was coded as depressive onset (1), manic onset (2), or not recorded/not applicable. Per-variable missingness was calculated before imputation and is reported in Supplementary Table 1. Most candidate variables had 0–10% missingness; ACTH had 18.0% missingness. The onset-polarity field had 33.3% unrecorded values and was treated as an explicit ‘not recorded/not applicable’ category rather than numerically median-imputed, because non-recording may be structurally related to diagnosis. Continuous predictors were median-imputed and standardized; categorical predictors were assigned an explicit missing category and one-hot encoded. Every imputer, encoder, and scaler was fitted only on the relevant training fold and then applied to the held-out fold or test set. The outcome was never imputed. A sensitivity analysis excluded both onset polarity and ACTH.

2.4 Statistical characterization and feature selection

Continuous variables were summarized as median (interquartile range) and compared using the Mann–Whitney U test; rank-biserial correlation and the common-language effect size were reported. Categorical variables were summarized as n (%) and compared using χ² tests with Cramér’s V. Benjamini–Hochberg correction controlled the false-discovery rate across all 65 candidate comparisons.

Feature selection was conducted only in the 864-patient development set and was completed before test-set evaluation. To limit overfitting given the 97 SB events in the development data, Elastic Net reduced the 65 candidates to 14 predictors: ACTH, ALT, AST, complement C4, CK-MB, age, cortisol, creatinine, high-density lipoprotein, onset-polarity category, previous SA, previous SB, progesterone, and testosterone. The Elastic Net used an L1/L2 mixing ratio of 0.25 and an inverse regularization strength of 0.0178. We compared this selection with rankings from mutual information, recursive feature elimination, and cross-fitted permutation importance. Bootstrap resampling assessed rank stability (Supplementary Table 2) (29). The methods did not select exactly the same variables, so the retained set should be regarded as uncertain rather than definitive.

2.5 Model development, imbalance strategies, and internal validation

Thirteen classifiers were compared: Elastic-Net logistic regression, linear discriminant analysis (LDA), Gaussian naïve Bayes, k-nearest neighbors, radial-basis-function support vector machine, multilayer perceptron, decision tree, random forest, AdaBoost, histogram gradient boosting, XGBoost, LightGBM, and CatBoost. Model fitting used fixed settings, with hyperparameters and software versions reported in Supplementary Table 3.

The cohort included 126 SB-positive and 952 SB-negative patients, representing moderate class imbalance. We compared observed-prior modeling, equal-prior or cost-sensitive LDA, class-weighted Elastic-Net logistic regression, and SMOTE-NC. SMOTE-NC was included as a sensitivity analysis because the predictors included both categorical and continuous variables. Resampling, when used, occurred only within training folds; the test set remained unchanged. Strategy selection considered AUC, AUPRC, and Brier score because oversampling and cost weighting can alter calibration.

Stability was assessed by fivefold stratified cross-validation repeated five times in the development set. Within each outer training fold, preprocessing was re-fitted and threefold internal calibration was used where applicable. The final selected model was refitted on the complete development set and sigmoid-calibrated using fivefold internal predictions. The operating threshold (0.111) maximized Youden’s J on out-of-fold development predictions and was locked before test evaluation. The 214 observed test patients were evaluated once without resampling.

2.6 Performance, clinical baselines, calibration, and explanation

We report ROC AUC, AUPRC, Brier score, accuracy, sensitivity, specificity, precision, F1 score, and MCC. Stratified nonparametric bootstrap resampling (2, 000 replicates) provided 95% confidence intervals on the observed test set. Calibration was summarized by a calibration plot, intercept, and slope. Decision-curve analysis used calibrated probabilities at thresholds from 0.02 to 0.50. Paired bootstrap differences compared the selected model with leading alternatives. Accuracy was not used as the model-selection criterion.

Clinical incremental value was examined against (1) a logistic model containing previous SA alone and (2) a parsimonious model containing previous SA, previous SB, age, and onset-polarity category. Sensitivity analyses removed previous SA and previous SB, removed onset polarity, and removed predictors with more than 15% missingness (onset polarity and ACTH).

For the final LDA model, SHAP values were computed on the observed test patients on the log-odds scale. One-hot levels were aggregated to their source predictor. Orange indicates higher feature values and blue indicates lower values; an explicit missing category is coded lower than recorded polarity categories only for visualization. SHAP ranking stability was assessed across 200 bootstrap refits. SHAP values are interpreted as model associations and not causal effects.

2.7 Sample-size adequacy and software

A post hoc sample-size calculation used 14 predictor parameters, a prevalence of 0.117, an anticipated C-statistic of 0.721, and a target shrinkage of 0.90. The calculation estimated a minimum of 1, 850 patients and 217 events. The available 1, 078 patients and 126 events fell below this estimate. We therefore used penalization, limited the number of predictors, repeated resampling, and reported interval estimates.

Analyses used Python 3.12 with pandas 3.0, scipy 1.18, scikit-learn 1.9, imbalanced-learn 0.14, XGBoost 3.4, LightGBM 4.7, CatBoost 1.2, statsmodels 0.14, and SHAP-compatible linear attribution routines. Statistical tests were two-sided. Reproducible analysis scripts and the fitted-model specification are available from the corresponding author subject to institutional data-governance requirements. Reporting was aligned with TRIPOD+AI ().

3 Results

3.1 Cohort and class accounting

Among 1, 078 analyzed patients, 126 (11.69%) had current SB within the three-month window and 952 (88.31%) did not. The development set contained 97 events among 864 patients; the observed test set contained 29 events among 214 patients. No evaluation patient was duplicated or synthesized. Figure 1 shows the exclusions, class counts, and bootstrap stability of the final SHAP ranking. Table 1 describes the 14 retained predictors. Younger age, previous SA, previous SB, onset-polarity category, AST, CK-MB, creatinine, and testosterone remained different after false-discovery-rate correction. Marginal effect sizes were generally small to modest.

Figure 1

Table 1

Distribution and effect size of retained predictors in the analytic cohort.

Values are median (IQR) unless otherwise indicated. CL, common-language effect size (probability that a randomly selected SB patient has a higher value than a randomly selected SB-N patient); r_rb, rank-biserial correlation. Laboratory units are those recorded by the hospital laboratory information system.

3.2 Model comparison and imbalance handling

Gaussian naïve Bayes had the highest mean uncalibrated AUC in repeated development-set cross-validation. Its test-set discrimination and calibration, however, were less consistent. LDA showed the best overall balance of AUC, AUPRC, Brier score, and test-set performance and was selected for calibration and interpretation. Paired bootstrap comparisons between calibrated LDA and Gaussian naïve Bayes, CatBoost, or random forest all produced AUC-difference intervals that included zero. The detail mentioned above shown in Table 2 and Figures 2, 3.

Table 2

Repeated cross-validation and observed-test performance of 13 model families.

Cross-validation values are mean ± SD across fivefold cross-validation repeated five times. Values are for fixed reproducible model settings before final sigmoid calibration.

Figure 2

Figure 3

Models fitted to the observed class distribution achieved cross-validated AUC 0.719, AUPRC 0.281, and Brier score 0.097. Equal-prior LDA, class-weighted logistic regression, and fold-restricted SMOTE-NC yielded Brier scores of 0.215, 0.208, and 0.210, respectively, without consistent gains in discrimination. The final probability model therefore used neither oversampling nor class weighting. Its operating threshold was selected from out-of-fold development predictions after model fitting. Table 3, Panel C reports the full comparison.

Table 3

Final-model performance, clinical baselines, sensitivity analyses, and class-imbalance strategies.

Values in parentheses in Panel A are 95% stratified bootstrap confidence intervals based on 2, 000 resamples. For Panel C, each cell reports repeated-cross-validation mean/observed-test value. SA, suicide attempt; SB, suicidal behavior.

3.3 Final model performance, calibration, and incremental value

After sigmoid calibration, repeated cross-validation yielded mean AUC 0.721 ± 0.052, AUPRC 0.268 ± 0.060, and Brier score 0.095 ± 0.004. On the 214 observed test patients, AUC was 0.704 (0.597–0.803), AUPRC 0.283 (0.198–0.453), and Brier score 0.111. At the locked threshold 0.111, the confusion matrix was TN = 122, FP = 63, FN = 13, and TP = 16. Calibration intercept was 0.572 (95% CI -0.807 to 1.951) and slope 1.164 (95% CI 0.488 to 1.840); intervals were wide because only 29 test events were available.

The previous-SA-only baseline yielded AUC 0.591; the four-variable parsimonious logistic model yielded AUC 0.691. The final model therefore improved AUC by 0.112 over previous SA alone but by only 0.013 over the parsimonious model. After excluding previous SA and previous SB, AUC decreased to 0.661 and AUPRC to 0.209. Excluding onset polarity alone yielded AUC 0.662; excluding both onset polarity and ACTH yielded AUC 0.647. These analyses indicate incremental information beyond suicide history but also substantial reliance on historical and onset-polarity fields. The detail mentioned above shown in Figures 2, 3.

3.4 Model explanation

Mean absolute SHAP values ranked onset-polarity category first, followed by previous SA, age, previous SB, testosterone, and high-density lipoprotein. Across 200 bootstrap refits, median ranks (IQR) were 1 (1–2) for onset polarity, 2 (2–4) for previous SA, 3 (3–4) for age, and 4 (3–7) for previous SB. Lower-ranked laboratory features showed wider rank variation. The beeswarm direction agreed with conventional group comparisons: recorded manic-onset category and positive previous-SA/SB values contributed toward higher modeled log odds, whereas older age and higher testosterone tended to contribute toward lower modeled log odds. These are conditional model associations and do not establish causation. The detail of SHAP shown in Figure 4.

Figure 4

4 Discussion

The calibrated LDA model showed moderate discrimination and the most consistent balance of repeated-validation stability, test-set performance, precision-recall performance, and calibration. Complex boosting models showed no statistically clear advantage.

Previous SA and previous SB, both occurring more than three months before admission, were among the strongest predictors. This is consistent with evidence that a history of suicidal behavior remains clinically informative (–). The previous-SA-only baseline had limited discrimination, whereas adding previous SB, age, and onset polarity brought performance close to that of the full model. A small set of historical and relatively stable characteristics available at admission may therefore help identify patients who need more intensive suicide-risk assessment and monitoring during hospitalization. The additional laboratory variables provided only modest incremental information.

Within this cohort, recorded manic onset contributed more strongly to a higher modeled probability of SB and was more common among SB-positive patients. This direction differs from several studies restricted to BD, which associated depressive onset with suicide attempts (30–35). One plausible explanation is that onset polarity partly acted as a proxy for diagnostic subtype. A documented manic first episode is consistent with bipolar-spectrum illness rather than unipolar major depressive disorder (MDD), whereas a depressive first episode can occur in either MDD or bipolar disorder (BD). In a comparative cohort of 3, 284 adults with major mood disorders, suicidal ideation, suicide attempts, and suicide were all more prevalent in BD than in MDD (36). BD diagnosis also remained independently associated with suicidal acts (36). Thus, greater representation of bipolar-spectrum illness in the manic-onset category could have contributed to the higher modeled probability observed here. This between-diagnosis interpretation complements, rather than replaces, evidence linking depressive onset to suicide attempts within BD. However, diagnostic subtype could not be verified in the de-identified dataset, and 33.3% of onset-polarity values were unrecorded. Referral and coding patterns at this tertiary center may also have contributed. The lower AUC after excluding onset polarity confirms model reliance on this field but does not establish a causal or broadly generalizable association.

Laboratory variables provided less stable information than historical and demographic predictors. Although testosterone, high-density lipoprotein, ACTH, ALT, progesterone, and CK-MB were retained during feature selection, most showed small marginal contributions and unstable bootstrap rankings. Their apparent associations should therefore be interpreted cautiously. Measurements obtained within 72 hours of admission may reflect transient physiological conditions rather than stable risk markers. They may also be influenced by acute illness severity, recent treatment, fasting status, sex, age, and comorbidities. This interpretation is further limited by the absence of medication data, including lithium, antidepressants, antipsychotics, and agents affecting endocrine or metabolic measures, which prevented adjustment for treatment-related confounding. Inflammatory markers emphasized in earlier studies were also only partially represented in this dataset (37–39). Accordingly, these laboratory findings are best viewed as exploratory signals that require confirmation in prospectively standardized cohorts with detailed medication and sampling information.

Because only about 12% of patients were SB-positive, we tested equal-prior learning, class weighting, and SMOTE-NC. None improved discrimination while preserving calibration, and SMOTE-NC lowered test AUC and increased the Brier score. The final model therefore retained the observed class distribution. Its operating threshold was selected from development-set predictions and was not adjusted on the test set. This threshold remains provisional because the clinical consequences of false-positive and false-negative classifications were not assessed.

4.1 Strengths and limitations

A principal strength of this study is the structured development and evaluation of the SB identification model. We compared 13 classifiers and several class-imbalance strategies, then assessed the selected model using repeated cross-validation and an untouched observed test set. Evaluation covered discrimination, minority-class performance, calibration, bootstrap uncertainty, decision-curve analysis, clinical baselines, sensitivity analyses, and SHAP-rank stability. Non-overlapping temporal definitions, exclusion of concurrent target-adjacent variables, and fold-specific preprocessing reduced the risk of information leakage.

Several design and measurement limitations constrain interpretation. The study was retrospective, single-center, and restricted to one year of inpatient data, with retrospective registration and no external validation. The chart-derived outcome was not measured with a validated suicide instrument, and inter-rater reliability was unavailable. Standardized symptom severity, hopelessness, impulsivity, detailed medication exposure, and raw PSP scores were also unavailable. Missingness was high for onset polarity (33.3%) and ACTH (18.0%); fold-wise imputation cannot remove bias when missingness is informative. The absence of reliable diagnostic subtype mapping also prevented subgroup analysis.

Model uncertainty was also considerable. The 126 events fell below the post hoc adequacy estimate of 217, and agreement between feature selectors was incomplete. Residual dependence between predictors may have affected both selection and attribution. The test set contained only 29 events, producing wide confidence intervals and unstable calibration estimates. SHAP values describe conditional model associations and do not imply that changing a laboratory marker would alter suicidal behavior.

5 Conclusion

In this retrospective inpatient cohort, a calibrated LDA model using 14 routinely available predictors showed moderate ability to identify suicidal behavior occurring within the three months preceding index admission, and complex boosting models did not show a clear advantage. More remote suicide history, onset-polarity category, and age carried much of the model signal, with limited but detectable incremental information from the wider predictor set. The model should be regarded as exploratory and internally validated only; prospective standardized outcome ascertainment and independent external validation are required before clinical use.

Statements

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Ethics statement

The studies involving humans were approved by Ethics Committee of Beijing Anding Hospital, Capital Medical University. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

Author contributions

SL: Writing – original draft. QJ: Formal analysis, Writing – original draft. JZ: Resources, Writing – original draft. YD: Formal analysis, Writing – original draft. YXX: Resources, Visualization, Writing – original draft. GZ: Resources, Writing – original draft. XL: Visualization, Writing – original draft. YJX: Data curation, Resources, Visualization, Writing – original draft. HY: Data curation, Resources, Writing – original draft. CH: Visualization, Writing – review & editing. SS: Conceptualization, Methodology, Writing – review & editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This work was supported by Beijing Municipal Administration of Hospitals Incubating Program (grant numbers px2025064). Open Access funding enabled by Beijing Municipal Administration of Hospitals Incubating Program.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fpsyt.2026.1950186/full#supplementary-material

References

• 1

  World Health Organization. Suicide: Key Facts (2025). Available online at: https://www.who.int/news-room/fact-sheets/detail/suicide (Accessed August 26, 2026).

  • Google Scholar
  • View reference in article

• 2

  BrådvikL. Suicide risk and mental disorders. Int J Environ Res Public Health. (2018) 15:2028. doi: 10.3390/ijerph15092028

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  SherLOquendoMA. Suicide: an overview for clinicians. Med Clinics North America. (2023) 107:119–30. doi: 10.1016/j.mcna.2022.03.008

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  ChengATChenTHChenCCJenkinsR. Psychosocial and psychiatric risk factors for suicide. Case-control psychological autopsy study. Br J Psychiatry: J Ment Sci. (2000) 177:360–5. doi: 10.1192/bjp.177.4.360

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  TureckiGBrentDAGunnellDO'ConnorRCOquendoMAPirkisJet al. Suicide and suicide risk. Nat Rev Dis Primers. (2019) 5:74. doi: 10.1038/mp.2013.153

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  LangeSRehmJTranABaggeCLJasilionisDKaplanMSet al. Comparing gender-specific suicide mortality rate trends in the United States and Lithuania, 1990-2019: putting one of the "deaths of despair" into perspective. BMC Psychiatry. (2022) 22:127. doi: 10.1186/s12888-022-03766-w

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  O'ConnorDBGartlandNO'ConnorRC. Stress, cortisol and suicide risk. Int Rev Neurobiol. (2020) 152:101–30. doi: 10.1016/bs.irn.2019.11.006

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  KwakYKimYKwonSJChungH. Mental health status of adults with cardiovascular or metabolic diseases by gender. Int J Environ Res Public Health. (2021) 18:514. doi: 10.3390/ijerph18020514

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  BallardEDVande VoortJLLuckenbaughDAMachado-VieiraRTohenMZarateCA. Acute risk factors for suicide attempts and death: prospective findings from the STEP-BD study. Bipolar Disord. (2016) 18:363–72. doi: 10.1111/bdi.12397

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  ThompsonAH. The suicidal process and self-esteem. Crisis. (2010) 31:311–6. doi: 10.1027/0227-5910/a000045

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  Bille-BraheUKerkhofADe LeoDSchmidtkeACrepetPLonnqvistJet al. A repetition-prediction study of European parasuicide populations: a summary of the first report from part II of the WHO/EURO Multicentre Study on Parasuicide in co-operation with the EC concerted action on attempted suicide. Acta Psychiatrica Scandinavica. (1997) 95:81–6. doi: 10.1111/j.1600-0447.1997.tb00378.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  TietQQIlgenMAByrnesHFMoosRH. Suicide attempts among substance use disorder patients: an initial step toward a decision tree for suicide management. Alcoholism Clin Exp Res. (2006) 30:998–1005. doi: 10.1111/j.1530-0277.2006.00114.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  Porras-SegoviaAMorenoMBarrigónMLLópez CastromanJCourtetPBerrouiguetSet al. Six-month clinical and ecological momentary assessment follow-up of patients at high risk of suicide: a survival analysis. J Clin Psychiatry. (2023) 84:22m14411. doi: 10.4088/jcp.22m14411

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  FavrilLYuRGeddesJRFazelS. Individual-level risk factors for suicide mortality in the general population: an umbrella review. Lancet Public Health. (2023) 8:e868–77. doi: 10.1016/s2468-2667(23)00207-4

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  AminiPAhmadiniaHPoorolajalJMoqaddasi AmiriM. Evaluating the high risk groups for suicide: a comparison of logistic regression, support vector machine, decision tree and artificial neural network. Iranian J Public Health. (2016) 45:1179–87.

  • Google Scholar
  • View reference in article

• 16

  SuRJohnJRLinPI. Machine learning-based prediction for self-harm and suicide attempts in adolescents. Psychiatry Res. (2023) 328:115446. doi: 10.1016/j.psychres.2023.115446

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  RozekDCAndresWCSmithNBLeifkerFRArneKJenningsGet al. Using machine learning to predict suicide attempts in military personnel. Psychiatry Res. (2020) 294:113515. doi: 10.1016/j.psychres.2020.113515

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  GradusJLRoselliniAJHorváth-PuhóEStreetAEGalatzer-LevyIJiangTet al. Prediction of sex-specific suicide risk using machine learning and single-payer health care registry data from Denmark. JAMA Psychiatry. (2020) 77:25–34. doi: 10.1001/jamapsychiatry.2019.2905

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  ZhengSZengWWuQLiWHeZLiEet al. Predictive models for suicide attempts in major depressive disorder and the contribution of EPHX2: a pilot integrative machine learning study. Depression Anxiety. (2024) 2024:5538257. doi: 10.1155/2024/5538257

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  McHughCMLargeMM. Can machine-learning methods really help predict suicide? Curr Opin Psychiatry. (2020) 33:369–74. doi: 10.1097/yco.0000000000000609

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  BelsherBESmolenskiDJPruittLDBushNEBeechEHWorkmanDEet al. Prediction models for suicide attempts and deaths: a systematic review and simulation. JAMA Psychiatry. (2019) 76:642–51. doi: 10.1001/jamapsychiatry.2019.0174

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  van den GoorberghRvan SmedenMTimmermanDVan CalsterB. The harm of class imbalance corrections for risk prediction models: illustration and simulation using logistic regression. J Am Med Inform Assoc. (2022) 29:1525–34. doi: 10.1093/jamia/ocac093

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  CarrieroALuijkenKde HondAMoonsKGMvan CalsterBvan SmedenM. The harms of class imbalance corrections for machine learning based prediction models: a simulation study. Stat Med. (2025) 44:e10320. doi: 10.1002/sim.10320

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  SteyerbergEWHarrellFEBorsboomGJEijkemansMJVergouweYHabbemaJD. Internal validation of predictive models: efficiency of some procedures for logistic regression analysis. J Clin Epidemiol. (2001) 54:774–81. doi: 10.1016/s0895-4356(01)00341-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  SteyerbergEWHarrellFE. Prediction models need appropriate internal, internal-external, and external validation. J Clin Epidemiol. (2016) 69:245–7. doi: 10.1016/j.jclinepi.2015.04.005

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBet al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. (2024) 385:e078378. doi: 10.1136/bmj-2023-078378

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  LundbergSMNairBVavilalaMSHoribeMEissesMJAdamsTet al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat BioMed Eng. (2018) 2:749–60. doi: 10.1038/s41551-018-0304-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28

  NtakoliaCYotsidiVRannouIGournellisR. Interpretable machine learning approach for predicting clinically significant suicide risk: a case study of patients with major depressive disorder in Greece. Psychiatry Res. (2025) 351:116607. doi: 10.1016/j.psychres.2025.116607

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29

  NtakoliaCKokkotisCMoustakidisSTsaopoulosD. Identification of most important features based on a fuzzy ensemble technique. Int J Med Inform. (2021) 156:104614. doi: 10.1016/j.ijmedinf.2021.104614

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  CarvalhoAFMcIntyreRSDimelisDGondaXBerkMNunes-NetoPRet al. Predominant polarity as a course specifier for bipolar disorder: a systematic review. J Affect Disord. (2014) 163:56–64. doi: 10.1016/j.jad.2014.03.035

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  ChaudhurySRGrunebaumMFGalfalvyHCBurkeAKSherLParseyRVet al. Does first episode polarity predict risk for suicide attempt in bipolar disorder? J Affect Disord. (2007) 104:245–50. doi: 10.1016/j.jad.2007.02.022

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32

  CremaschiLDell'OssoBVismaraMDobreaCBuoliMKetterTAet al. Onset polarity in bipolar disorder: a strong association between first depressive episode and suicide attempts. J Affect Disord. (2017) 209:182–7. doi: 10.1016/j.jad.2016.11.043

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  KalmanJLOlde LoohuisLMVreekerAMcQuillinAStahlEARuderferDet al. Characterisation of age and polarity at onset in bipolar disorder. Br J Psychiatry: J Ment Sci. (2021) 219:659–69. doi: 10.1192/bjp.2021.102

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  Roy-ByrnePPostRMUhdeTWPorcuTDavisD. The longitudinal course of recurrent affective illness: life chart data from research patients at the NIMH. Acta Psychiatrica Scandinavica Supplementum. (1985) 317:1–34. doi: 10.1111/j.1600-0447.1985.tb10510.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  PerugiGMicheliCAkiskalHSMadaroDSocciCQuiliciCet al. Polarity of the first episode, clinical characteristics, and course of manic depressive illness: a systematic retrospective investigation of 320 bipolar I patients. Compr Psychiatry. (2000) 41:13–8. doi: 10.1016/s0010-440x(00)90125-1

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  BaldessariniRJTondoLPinnaMNuñezNVázquezGH. Suicidal risk factors in major affective disorders. Br J Psychiatry. (2019) 215:621–6. doi: 10.1192/bjp.2019.167

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  CanselNYaginFHAkanMIlkay AygulB. Interpretable estimation of suicide risk and severity from complete blood count parameters with explainable artificial intelligence methods. Psychiatria Danubina. (2023) 35:62–72. doi: 10.24869/psyd.2023.62

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  WangHKanWJFengYFengLYangYChenPet al. Nuclear receptors modulate inflammasomes in the pathophysiology and treatment of major depressive disorder. World J Psychiatry. (2021) 11:1191–205. doi: 10.5498/wjp.v11.i12.1191

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  KeatonSAMadajZBHeilmanPSmartLGritJGibbonsRet al. An inflammatory profile linked to increased suicide risk. J Affect Disord. (2019) 247:57–65. doi: 10.1016/j.jad.2018.12.100

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

Keywords

calibration, internal validation, machine learning, mood disorders, SHAP, suicidal behavior

Citation

Liang S, Jiang Q, Zhang J, Deng Y, Xu Y, Zhao G, Liu X, Xing Y, Yu H, Hu C and Sha S (2026) Development and internal validation of machine learning models for identifying suicidal behavior in inpatients with mood disorders. Front. Psychiatry 17:1950186. doi: 10.3389/fpsyt.2026.1950186

Received

27 July 2026

Revised

04 September 2026

Accepted

08 September 2026

Published

05 October 2026

Volume

17 - 2026

Edited by

Gábor Gazdag, Jahn Ferenc Dél-Pesti Kórház és Rendelőintézet, Hungary

Updates

Copyright

© 2026 Liang, Jiang, Zhang, Deng, Xu, Zhao, Liu, Xing, Yu, Hu and Sha.

This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Sha Sha, shashaanding@ccmu.edu.cn

Disclaimer

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.