Chen Shanshan, Liu Lingjuan, Hu Dingji, Zhu Yike, Wu Changde, Fu Haoya, Wu Jing, Zhang Kexin, Chang Maoxiao, Hao Tong, Wang Wan, Liu Songqiao
Jiangsu Provincial Key Laboratory of Critical Care Medicine, Department of Critical Care Medicine, Trauma Center, Zhongda Hospital, School of Medicine, Southeast University, Nanjing 210009 Jiangsu, People's Republic of China; Department of Respiratory and Critical Care Medicine, Second Affiliated Hospital of Xuzhou Medical University, Xuzhou Mining Group General Hospital, Xuzhou 221006 Jiangsu, People's Republic of China.
Jiangsu Provincial Key Laboratory of Critical Care Medicine, Department of Critical Care Medicine, Trauma Center, Zhongda Hospital, School of Medicine, Southeast University, Nanjing 210009 Jiangsu, People's Republic of China.
Intensive Crit Care Nurs. 2026 Feb;92:104165. doi: 10.1016/j.iccn.2025.104165. Epub 2025 Jul 28.
To evaluate the effectiveness of Artificial Intelligence (AI) in improving clinical outcomes in Extracorporeal Membrane Oxygenation (ECMO) management, focusing on ECMO initiation, prognosis, and complications.
A meta-analysis following PRISMA guidelines were conducted, with literature searches in PubMed, Embase, and Cochrane Library for studies on AI-based ECMO prediction models between January 1980 and June 2024. Data extraction included AI methodologies, performance metrics, and key findings, with risk of bias was assessed using PROBAST. Meta-analysis used a random-effects model to account for anticipated heterogeneity, with pooled AUCs calculated for ECMO initiation and prognosis prediction.
Of 212 initial records, 36 studies met criteria for inclusion. For ECMO initiation, the pooled AUC was 0.838 (95% CI: 0.804-0.873), and for prognosis prediction, the pooled AUC was 0.776 (95% CI: 0.755-0.797). Significant heterogeneity was observed (I=96.5 % for ECMO initiation, I=98.6 % for prognosis). Subgroup analysis revealed single-center studies exhibited higher AUCs for both initiation (AUC=0.888, 95% CI: 0.865-0.910) and prognosis prediction (AUC=0.803, 95% CI: 0.688-0.918) compared to multi-center studies in initiation (AUC=0.823, 95% CI: 0.782-0.864) and prognosis prediction (AUC=0.772, 95% CI: 0.752-0.792). Key AI applications included patient identification, mortality prediction, enhancing resource allocation and decision-making. However, due to data variability and limited external validation, the pooled findings should be interpreted in light of the limitations identified.
AI has shown promise, albeit with significant heterogeneity, in improving ECMO management by providing predictions for initiation timing and patient outcomes. Future research should focus on enhancing model generalizability through multi-center validation, and standardizing data to reduce heterogeneity.
AI models could enhance ECMO management by identifying high-risk patients, predicting adverse events, and guiding timely interventions. Healthcare providers should consider integrating AI tools to ECMO management, while considering its limitations in data and validation.
评估人工智能（AI）在改善体外膜肺氧合（ECMO）管理临床结局方面的有效性，重点关注ECMO启动、预后和并发症。
按照PRISMA指南进行荟萃分析，在PubMed、Embase和Cochrane图书馆中检索1980年1月至2024年6月间关于基于AI的ECMO预测模型的研究。数据提取包括AI方法、性能指标和主要发现，并使用PROBAST评估偏倚风险。荟萃分析采用随机效应模型来考虑预期的异质性，计算ECMO启动和预后预测的合并AUC。
在212条初始记录中，36项研究符合纳入标准。对于ECMO启动，合并AUC为0.838（95%CI：0.804 - 0.873）；对于预后预测，合并AUC为0.776（95%CI：0.755 - 0.797）。观察到显著的异质性（ECMO启动时I = 96.5%，预后时I = 98.6%）。亚组分析显示，与多中心研究相比，单中心研究在启动（AUC = 0.888，95%CI：0.865 - 0.910）和预后预测（AUC = 0.803，95%CI：0.688 - 0.918）方面的AUC更高，多中心研究在启动（AUC = 0.823，95%CI：0.782 - 0.864）和预后预测（AUC = 0.772，95%CI：0.752 - 0.792）方面的AUC较低。关键的AI应用包括患者识别、死亡率预测、优化资源分配和决策制定。然而，由于数据变异性和外部验证有限，应根据已确定的局限性来解释汇总结果。
尽管存在显著异质性，但AI在通过提供启动时机和患者结局预测来改善ECMO管理方面已显示出前景。未来的研究应专注于通过多中心验证提高模型的通用性，并标准化数据以减少异质性。
AI模型可通过识别高危患者、预测不良事件和指导及时干预来改善ECMO管理。医疗保健提供者应考虑将AI工具整合到ECMO管理中，同时考虑其在数据和验证方面的局限性。
