Scholarlink

<p>Timely activation of massive hemorrhage protocols (MHP) is critical to prevent exsanguination and improve survival in trauma patients. Current clinical tools for predicting MHP need have limited sensitivity and moderate specificity. Machine learning (ML) models may enhance early recognition of life-threatening hemorrhage. We conducted a systematic review and meta-analysis of studies that developed or validated ML models to predict MHP need in patients with traumatic injuries (PROSPERO CRD42024574891). Databases searched included MEDLINE, EMBASE, Web of Science, the Cochrane Central Register of Controlled Trials, and LILACS. Risk of bias was assessed with PROBAST, a funnel plot and Deeks’ statistical test, and the GRADE checklist. Of the 4542 abstracts screened, 21 studies (50 ML models) were included. The ML model types with the highest pooled AUROC were random forest (0.89, 95% CI 0.85–0.93), neural networks (0.88, 95% CI 0.83–0.93), and XGBoost Models (0.86, 95% CI 0.80–0.93), all of which commonly outperforming their non-ML comparator. Despite superior discrimination, most ML models lacked robust subgroup performance evaluation, external validation, or feasibility for clinical deployment. The median APPRAISE-AI methodologic quality score was 51 out of 100 (Interquartile range [IQR]: 46, 55). Further, model reporting quality (TRIPOD + AI) was moderate; no study achieved full points across all 27 reporting items. Risk of bias was frequently high or unclear. Thus, pooled performance metrics should be interpreted with caution, as they may be influenced by methodological weaknesses in the underlying studies. These findings underscore the need for more rigorously designed and transparently reported ML studies in trauma care. Realizing the potential of ML for supporting MHP delivery in trauma care will require further collaborative, clinically grounded model development with robust validation, transparent reporting, and a focus on real-world deployment to translate predictive performance into improved patient outcomes.</p><p><i>PROSPERO Registration Number</i>: CRD42024574891.</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

Systematic review and meta-analysis of machine learning models predicting massive hemorrhage protocol in trauma

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Gemma Postill,
  • Anglin Dent,
  • Richard Cheng,
  • Anton Nikouline,
  • Teruko Kishibe,
  • Melissa McGowan,
  • Jethro C. C. Kwong,
  • Brodie Nolan

摘要

Timely activation of massive hemorrhage protocols (MHP) is critical to prevent exsanguination and improve survival in trauma patients. Current clinical tools for predicting MHP need have limited sensitivity and moderate specificity. Machine learning (ML) models may enhance early recognition of life-threatening hemorrhage. We conducted a systematic review and meta-analysis of studies that developed or validated ML models to predict MHP need in patients with traumatic injuries (PROSPERO CRD42024574891). Databases searched included MEDLINE, EMBASE, Web of Science, the Cochrane Central Register of Controlled Trials, and LILACS. Risk of bias was assessed with PROBAST, a funnel plot and Deeks’ statistical test, and the GRADE checklist. Of the 4542 abstracts screened, 21 studies (50 ML models) were included. The ML model types with the highest pooled AUROC were random forest (0.89, 95% CI 0.85–0.93), neural networks (0.88, 95% CI 0.83–0.93), and XGBoost Models (0.86, 95% CI 0.80–0.93), all of which commonly outperforming their non-ML comparator. Despite superior discrimination, most ML models lacked robust subgroup performance evaluation, external validation, or feasibility for clinical deployment. The median APPRAISE-AI methodologic quality score was 51 out of 100 (Interquartile range [IQR]: 46, 55). Further, model reporting quality (TRIPOD + AI) was moderate; no study achieved full points across all 27 reporting items. Risk of bias was frequently high or unclear. Thus, pooled performance metrics should be interpreted with caution, as they may be influenced by methodological weaknesses in the underlying studies. These findings underscore the need for more rigorously designed and transparently reported ML studies in trauma care. Realizing the potential of ML for supporting MHP delivery in trauma care will require further collaborative, clinically grounded model development with robust validation, transparent reporting, and a focus on real-world deployment to translate predictive performance into improved patient outcomes.

PROSPERO Registration Number: CRD42024574891.