摘要
目的全面概括机器学习技术预测儿童脓毒性休克的模型表现和预测效果,以便有针对性地提升未来研究的质量和模型的预测能力。方法计算机检索PubMed、Embase、Web of Science、ScienceDirect、CNKI、WanFang Data数据库,检索时限均为建库至2024年4月1日,搜集有关机器学习预测儿童脓毒性休克的研究。由两名研究者独立筛选文献,提取资料并评价纳入研究的偏倚风险后,对于基本信息、研究数据、研究设计及预测模型等进行系统评价。使用随机效应模型合并模型曲线下面积(AUC)进行Meta分析。根据样本量、机器学习模型、预测变量类型、预测变量个数等进行亚组分析,并且对于纳入文献进行发表偏倚与敏感性分析等。结果最终纳入11项研究,其中包含2项低偏倚风险,7项未知偏倚风险,2项高偏倚风险的研究。纳入研究使用到的数据包括公开与非公开的电子病历数据库,用到的机器学习模型包括逻辑回归、随机森林、支持向量机和XGBoost等,且基于不同数据库构建的预测模型出现了不同的特征变量结果,因此确定预测模型的关键变量还需要在其他数据集上进行进一步的验证。Meta分析显示总AUC为0.812[95%CI(0.763,0.860),P<0.001]。进一步的亚组分析显示,较大的样本量(≥1000例)和预测变量类型可以显著提高模型的预测效果(95%CI无重叠)。漏斗图显示纳入研究存在发表偏倚,当剔除极端AUC值后,总AUC为0.815[95%CI(0.769,0.861),P<0.001],表明极端AUC值不敏感。结论机器学习技术在预测儿童脓毒性休克方面展示出一定的潜力,但现有研究在质量上还有待加强,未来的研究工作应提升研究质量并且通过扩大样本量提高模型的预测效果。
Objective To provide a comprehensive overview of model performance and predictive efficacy of machine learning techniques to predict septic shock in children,in order to target and improve the quality and predictive power of models for future studies.Methods To systematically review all studies in four databases(PubMed,Embase,Web of Science,ScienceDirect,CNKI,WanFang Data)on machine learning prediction of septic shock in children before April 1,2024.Two investigators independently conducted literature screening,literature data extraction and bias assessment,and conducted a systematic review of basic information,research data,study design and prediction models.Model discrimination,which area under the curve(AUC),was pooled using a random-effects model and meta-analysis was performed.Subgroup analyses were performed according to sample sizes,machine learning models,types of predictors,number of predictors,etc.And publication bias and sensitivity analyses were performed for the included literature.Results A total of 11 studies were included,of which 2 were at low risk of bias,7 were at unknown risk of bias,and 2 were at high risk of bias.The data used in the included studies included both public and non-public electronic medical record databases,and the machine learning models used included logistic regression,random forest,support vector machine,and XGBoost,etc.The predictive models constructed based on different databases appeared to have different results in terms of the characteristic variables,so identifying the key variables of the predictive models requires further validation on other datasets.Meta-analysis showed the pooled AUC of 0.812(95%CI 0.763 to 0.860,P<0.001),and further subgroup analyses showed that larger sample sizes(≥1000)and predictor variable types significantly improved the predictive effect of the model,and the difference in AUC was statistically significant(95%CI not overlapping).The funnel plot showed that there was publication bias in the study,and when the extreme AUC values were excluded,the meta-analysis yielded a total AUC of 0.815(95%CI 0.769 to 0.861,P<0.001),indicating that the extreme AUC values were insensitive.Conclusion Machine learning technology has shown some potential in predicting septic shock in children,but the quality of existing research needs to be strengthened,and future research work should improve the quality of research and improve the prediction effect of the model by expanding the sample size.
作者
甄凯旋
闫浩伦
陈龙彪
杨晓庆
吴谨准
ZHEN Kaixuan;YAN Haolun;CHEN Longbiao;YANG Xiaoqing;WU Jinzhun(Women and Children’s Hospital Affiliated to Xiamen University,Xiamen 361003,P.R.China;School of Pharmacy,Xiamen University,Xiamen 361102,P.R.China;School of Medicine,Xiamen University,Xiamen 361102,P.R.China;School of Information,Xiamen University,Xiamen 361102,P.R.China)
出处
《中国循证医学杂志》
北大核心
2025年第2期200-205,共6页
Chinese Journal of Evidence-based Medicine
基金
厦门市科技局重大项目(编号:3502Z20221021)。
