文献速递烧伤创面修复

人工智能在烧伤结局预测中的应用:范围综述

Artificial intelligence for outcome prediction in burn care: A scoping review.
PubMed 文献 · Burns · 2026-06-20

机器学习在死亡率、脓毒症、吸入性损伤和急性肾损伤预测中表现较好,但多数研究为回顾性单中心且缺乏外部验证。

要点

  • 机器学习模型在烧伤死亡率预测中持续优于传统评分,仅用少量常规预测因子。
  • 脓毒症预测模型准确度优于美国烧伤协会脓毒症标准和Sepsis-3定义。
  • 基于中性粒细胞明胶酶相关脂质运载蛋白的急性肾损伤模型AUROC最高达0.97。

核心结论

  • 机器学习在烧伤关键结局预测中展现较大潜力,尤其死亡率、吸入性损伤和急性肾损伤。
  • 多数研究为回顾性、单中心且缺乏外部验证,临床部署前需多中心前瞻性评估。

研究设计与样本

  • 遵循PRISMA-ScR指南的范围综述,检索两个数据库2020至2025年文献。

主要发现

  • 死亡率预测最成熟,机器学习模型用年龄、烧伤总面积、深度和吸入性损伤等少量常规预测因子持续优于传统评分。
  • 脓毒症预测模型准确度优于美国烧伤协会脓毒症标准和Sepsis-3定义,尤其基于聚焦特征集时。
  • 住院时长模型增益有限,吸入性损伤分级和气管切开预测表现较强。
  • 谵妄和术中失血预测模型准确度中至高,基于影像的手术候选评估结果不一。
  • 基于中性粒细胞明胶酶相关脂质运载蛋白的急性肾损伤模型AUROC最高达0.97。

临床怎么用

  • 机器学习模型可用于烧伤死亡率、脓毒症、吸入性损伤和急性肾损伤的风险分层辅助。
  • 临床部署前应关注模型的外部验证情况,避免直接套用单中心回顾性研究结果。

局限性

  • 多数研究为回顾性设计。
  • 多数研究为单中心。
  • 缺乏外部验证。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

原文

INTRODUCTION: Machine learning (ML) is increasingly applied across burn care, but its clinical utility remains unclear. This scoping review mapped current evidence on ML models developed to predict key outcomes in burn patients, including mortality, sepsis, inhalation injury, delirium, surgical decision making, intraoperative blood loss and acute kidney injury (AKI). METHODS: Following the Preferred Reporting Items for Systematic Reviews and Meta Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines, two databases were searched from 2020 to 2025. Studies using ML to predict diagnostic, prognostic, or treatment-related outcomes in burn patients were included. RESULTS: Twenty-three studies met inclusion criteria. Mortality prediction was the most advanced domain, with ML models consistently outperforming traditional scores using a small set of routinely collected predictors (age, total body surface area burned, depth, inhalation injury). Sepsis prediction models demonstrated superior accuracy to American Burn Association (ABA) sepsis criteria and Sepsis-3 definitions, particularly when based on focused feature sets. Length of stay models showed modest gains, whilst inhalation injury grading and tracheostomy prediction achieved strong performance. Delirium and intraoperative blood loss prediction models showed moderate to high accuracy, while imaging-based approaches for surgical candidacy produced variable results. AKI prediction using neutrophil gelatinase-associated lipocalin based models achieved area under the receiver operating characteristic curve (AUROC) values of up to 0.97. CONCLUSION: ML shows substantial promise for predicting key outcomes in burn care, particularly mortality, inhalation injury and AKI. However, most studies were retrospective, single-centre, and lacked external validation, highlighting the need for robust multicentre and prospective evaluation before clinical deployment.

来源:Burns,查看原文。版权归原作者所有。

引用

  1. Ahn J, Ahn J, Herron J, et al. Artificial intelligence for outcome prediction in burn care: A scoping review. Burns. 2026. PMID: 42407287. doi:10.1016/j.burns.2026.108121

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。