文献速递乳房整形与再造烧伤美容整形

整形外科术前风险预测AI/ML模型的方法学质量与性能系统评价

Methodological quality and performance of artificial intelligence and machine learning models for preoperative risk prediction in plastic surgery: A systematic review.
PubMed 文献 · J Plast Reconstr Aesthet Surg · 2026-07-17

该系统评价采用PROBAST+AI框架评估整形外科术前风险预测AI/ML模型的方法学质量、偏倚风险与预测性能,检索5个数据库至2025年10月,纳入10项研究,涵盖自体乳房再造、异体乳房再造、头颈重建、烧伤手术和美容手术。8项报告区分度的研究AUC为0.66至0.82,但仅2项(20%)进行外部验证,7项(70%)未报告模型校准,存在重大方法学局限,尚不支持临床实施。

要点

  • 纳入10项研究,随机森林最常用(n=4),AUC范围0.66至0.82。
  • 仅2项(20%)行外部验证,7项(70%)未报告模型校准。
  • 7项开发研究中5项(71.4%)每变量事件数<10,样本量不足。

核心结论

  • 当前整形外科术前风险预测AI/ML模型性能不一且存在重大方法学局限,尚不支持临床实施。
  • 需多机构前瞻性验证研究后方可临床采用。

研究设计与样本

  • 系统评价,采用PROBAST+AI框架评估方法学质量、偏倚风险与预测性能。
  • 检索5个数据库自建库至2025年10月,纳入10项研究。
  • 研究涵盖自体乳房再造(n=2)、异体乳房再造(n=5)、头颈重建(n=1)、烧伤手术(n=1)和美容手术(n=1)。

主要发现

  • 随机森林最常用(n=4),其次为神经网络(n=2)、deep forest联合RUSBoost(n=1)、支持向量机(n=1)和逻辑回归(n=1)。
  • 8项报告区分度的研究AUC范围为0.66至0.82。
  • 仅2项研究(20%)进行外部验证。
  • 7项开发研究中5项(71.4%)每变量事件数<10,提示样本量不足。
  • 7项研究(70%)未报告模型校准。

关键数值与推荐

  • 预协调评分者间信度基于102对领域级评分,Cohen kappa为0.240, Prevalence-Adjusted Bias-Adjusted Kappa为0.039。
  • 所有分歧通过结构化共识解决。

临床怎么用

  • 现有模型尚不具备临床实施条件,不应直接用于术前风险预测决策。
  • 临床采用前需等待多机构前瞻性验证研究结果。

局限性

  • 仅纳入10项研究,证据基础有限。
  • 多数研究缺乏外部验证和模型校准报告。
  • 开发研究普遍样本量不足(每变量事件数<10)。

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

Artificial intelligence (AI) and machine learning (ML) are increasingly being applied to preoperative risk prediction in plastic surgery; however, the methodological quality and clinical readiness of these models are yet to be systematically evaluated. This systematic review assessed the quality, risk of bias, and predictive performance of AI/ML preoperative risk prediction models in plastic surgery using the PROBAST+AI framework. Five databases were searched from inception through October 2025. Ten studies met the inclusion criteria, encompassing autologous breast reconstruction (n = 2), alloplastic breast reconstruction (n = 5), head and neck reconstruction (n = 1), burn surgery (n = 1), and aesthetic surgery (n = 1). Random forest was the most frequently used algorithm (n = 4), followed by neural networks (n = 2), deep forest with RUSBoost (n = 1), support vector machine (n = 1), and logistic regression (n = 1). AUC ranged from 0.66 to 0.82 among the 8 studies reporting discrimination. Critical methodological limitations were identified: only 2 studies (20%) performed external validation, 5 of 7 development studies (71.4%) had events per variable <10 indicating inadequate sample size, and 7 studies (70%) did not report model calibration. Pre-reconciliation inter-rater reliability across 102 paired domain-level ratings yielded a Cohen's kappa of 0.240 and Prevalence-Adjusted Bias-Adjusted Kappa of 0.039, consistent with published benchmarks for PROBAST-based systematic reviews. All discrepancies were resolved via structured consensus. Current AI/ML models for preoperative risk prediction in plastic surgery demonstrate variable performance and substantial methodological limitations that preclude clinical implementation. Multi-institutional prospective validation studies with rigorous methodology are needed before clinical adoption.

来源:J Plast Reconstr Aesthet Surg,查看原文。版权归原作者所有。

引用

  1. Kumar AR, Singh S, Ishii M, et al. Methodological quality and performance of artificial intelligence and machine learning models for preoperative risk prediction in plastic surgery: A systematic review. J Plast Reconstr Aesthet Surg. 2026. PMID: 42537558. doi:10.1016/j.bjps.2026.07.015

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。