文献速递呼吸衰竭与机械通气重症综合管理与镇静镇痛

机器学习在机械通气患者呼吸机不同步中的应用进展:系统综述

Application progress of machine learning in patient-ventilator asynchrony during mechanical ventilation: a systematic review.
PubMed 文献 · Crit Care · 2025-07-10
一句话结论机器学习模型在识别呼吸机不同步方面报告性能高(F1 0.731–0.988),但仅2项研究进行外部验证,临床广泛应用受限。

该系统综述遵循PROSPERO注册方案(CRD42024626163),检索PubMed、Embase、Web of Science和Cochrane数据库至2024年11月,纳入14篇应用机器学习评估成人机械通气患者呼吸机不同步(PVA)的原始研究。共采用11种机器学习方法,针对19种PVA类型,其中13项研究聚焦PVA识别,仅1项预测PVA发生。机器学习模型识别PVA表现出高F1分数(0.731–0.988),但仅2项研究进行外部验证,类别不平衡处理(5项研究)和超参数调优(4项研究)等技术细节报告有限。作者认为机器学习在PVA识别和预测方面有显著潜力,但方法学异质性、PVA定义和结局指标差异、外部验证不足及技术细节报告不一致限制了其广泛临床应用。

要点

  • 机器学习模型识别PVA的F1分数范围为0.731–0.988,在报告的其他稳健指标(如准确率、敏感度、特异度、MCC、AUROC)上也表现良好。
  • 14项纳入研究中仅2项进行了外部验证,外部验证不足是临床广泛应用的主要限制之一。
  • 13项研究聚焦PVA识别,仅1项研究预测PVA发生。
  • 技术细节报告有限:仅5项研究处理类别不平衡,仅4项研究描述超参数调优。
  • 作者建议未来研究优先开发可在多样临床环境中外部验证的通用模型,标准化PVA定义和结局指标,并提供详细方法学报告。

适用人群与场景

  • 接受机械通气的成人患者(纳入标准:应用机器学习技术评估PVA的原始研究)。

研究设计与关键结果

  • 研究类型:系统综述,遵循PROSPERO注册方案(CRD42024626163),检索PubMed、Embase、Web of Science和Cochrane数据库自建库至2024年11月。
  • 纳入14篇文献,采用11种不同机器学习方法,针对19种目标PVA类型。
  • 13项研究聚焦PVA识别,仅1项预测PVA发生。
  • 机器学习模型识别PVA的F1分数范围为0.731–0.988,在报告的其他稳健指标(如准确率、敏感度、特异度、MCC、AUROC)上也表现良好。
  • 仅2项研究进行了外部验证。
  • 类别不平衡处理在5项研究中涉及,超参数调优在4项研究中描述。

证据类型、等级与局限

  • 证据类型:系统综述。
  • 局限:方法学异质性大,PVA定义和结局指标存在变异性,外部验证不足,关键技术细节报告不一致。

对床旁实践的改变

  • 作者认为机器学习模型在识别和预测PVA方面显示出显著潜力,但广泛临床应用目前受限于方法学异质性、PVA定义和结局指标差异、外部验证不足及关键技术细节报告不一致。
  • 作者建议未来研究优先开发可在多样临床环境中外部验证的通用模型,标准化PVA定义和结局指标,并提供详细方法学报告以提高可重复性和临床适用性。

提炼说明

  • 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开

INTRODUCTION: Patient-ventilator asynchrony (PVA) is a common and harmful complication during mechanical ventilation, often requiring labor-intensive manual assessment. Machine learning (ML) offers a promising approach for automated and accurate PVA detection and prediction. We conducted a systematic review to evaluate the methodologies and performance of ML models applied to PVA.

METHODS: Following a pre-established PROSPERO-registered protocol (CRD42024626163), we systematically searched PubMed, Embase, Web of Science, and Cochrane databases from inception to November 2024. Original studies applying ML techniques to assess PVA in adult patients receiving mechanical ventilation were included. A refined quality assessment method was developed to evaluate methodological and technical features of included studies.

RESULTS: Among the 74 articles, fourteen were included, employing 11 distinct ML methods to address 19 targeted types of PVA. Thirteen studies focused on PVA identification, while only one predicted its occurrence. ML models demonstrated promising performance for PVA identification, with high F1 scores (ranging from 0.731 to 0.988) and strong results across other robust metrics (e.g., accuracy, sensitivity, specificity, MCC, AUROC) where reported. However, only two studies conducted external validation. Furthermore, reporting of crucial technical details, such as class imbalance handling (addressed in 5 studies) and hyperparameter tuning (described in 4 studies), was often limited.

CONCLUSIONS: ML models show significant potential for identifying and predicting PVA, as evidenced by their high reported performance metrics. However, their widespread clinical application is currently limited by substantial methodological heterogeneity, variability in PVA definitions and outcome measures, insufficient external validation, and inconsistent reporting of crucial technical details. Future research should prioritize developing generalizable models validated externally in diverse clinical settings, standardizing PVA definitions and outcome measures, and providing detailed methodological reporting to enhance reproducibility and clinical applicability.

来源:Crit Care,查看原文。版权归原作者所有。

参考来源

  1. Jiang G, Ma J, Xu H, et al. Application progress of machine learning in patient-ventilator asynchrony during mechanical ventilation: a systematic review. Crit Care. 2025. PMID: 40640842. doi:10.1186/s13054-025-05523-3

供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。