Scholarlink

Objective <p>This diagnostic test accuracy meta-analysis aimed to provide clinically interpretable estimates (sensitivity, specificity, likelihood ratios (LR), predictive value (PV) and post‑test probabilities) for machine‑learning (ML) models predicting suicide‑related outcomes.</p> Methods <p>A systematic search of PubMed, Embase, PsycINFO, and Web of Science identified studies published between January 2010 and December 2024. Eligible studies used a single-gate design (cross-sectional or longitudinal), included at least 100 participants, and reported diagnostic performance metrics (e.g., sensitivity, specificity, area under the receiver operating curve (AUC)) for ML algorithms. Models examined included Support Vector Machines (SVM), Logistic Regression, Random Forest, XGBoost, Artificial Neural Networks (ANN), Gradient Boosting, and Ensemble approaches. Two reviewers independently extracted data. Pooled estimates of sensitivity, specificity, PPV, NPV, and AUC were calculated using a bivariate random-effects model. Risk of bias was assessed using QUADAS-2.</p> Results <p>Of 500 screened records, 22 studies met inclusion criteria. Ensemble models demonstrated the highest pooled AUC (0.95; 95% CI, 0.92–0.96), with specificity of 0.97 (95% CI, 0.95–0.98) and sensitivity of 0.50 (95% CI, 0.29–0.71). Gradient-boosting and ensemble approaches showed strong discriminative performance overall; model-specific estimates are reported in the Results. At a pre-test probability of 25%, post-test probabilities for a positive result ranged from 64% (Logistic Regression) to 88% (Ensemble Models). </p> Conclusion <p>Machine-learning approaches demonstrated promising diagnostic accuracy for suicide-related outcomes in heterogeneous clinical populations. However, because primary studies rarely reported diagnosis- or outcome-specific performance, these findings should not be assumed to generalize across specific disorders or to suicide mortality. Future research should incorporate diagnosis-stratified and outcome-resolved validation to clarify clinical applicability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diagnostic accuracy of machine learning approaches for suicide‑related outcomes: a meta‑analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月9日

查看全文
  • Parisa Kohnepoushi,
  • Maryam Afraie,
  • Hamza Rahmani,
  • Seyedeh Asrin Seyedoshohadaei,
  • Babak Ghadirzadeh,
  • Yousef Moradi

摘要

Objective

This diagnostic test accuracy meta-analysis aimed to provide clinically interpretable estimates (sensitivity, specificity, likelihood ratios (LR), predictive value (PV) and post‑test probabilities) for machine‑learning (ML) models predicting suicide‑related outcomes.

Methods

A systematic search of PubMed, Embase, PsycINFO, and Web of Science identified studies published between January 2010 and December 2024. Eligible studies used a single-gate design (cross-sectional or longitudinal), included at least 100 participants, and reported diagnostic performance metrics (e.g., sensitivity, specificity, area under the receiver operating curve (AUC)) for ML algorithms. Models examined included Support Vector Machines (SVM), Logistic Regression, Random Forest, XGBoost, Artificial Neural Networks (ANN), Gradient Boosting, and Ensemble approaches. Two reviewers independently extracted data. Pooled estimates of sensitivity, specificity, PPV, NPV, and AUC were calculated using a bivariate random-effects model. Risk of bias was assessed using QUADAS-2.

Results

Of 500 screened records, 22 studies met inclusion criteria. Ensemble models demonstrated the highest pooled AUC (0.95; 95% CI, 0.92–0.96), with specificity of 0.97 (95% CI, 0.95–0.98) and sensitivity of 0.50 (95% CI, 0.29–0.71). Gradient-boosting and ensemble approaches showed strong discriminative performance overall; model-specific estimates are reported in the Results. At a pre-test probability of 25%, post-test probabilities for a positive result ranged from 64% (Logistic Regression) to 88% (Ensemble Models).

Conclusion

Machine-learning approaches demonstrated promising diagnostic accuracy for suicide-related outcomes in heterogeneous clinical populations. However, because primary studies rarely reported diagnosis- or outcome-specific performance, these findings should not be assumed to generalize across specific disorders or to suicide mortality. Future research should incorporate diagnosis-stratified and outcome-resolved validation to clarify clinical applicability.