ISSN 2097-6046（网络）
ISSN 2096-7446（印刷）
CN 10-1655/R
主管：中国科学技术协会
主办：中华护理学会
中华急危重症护理杂志

中华急危重症护理杂志 ›› 2025, Vol. 6 ›› Issue (5): 536-542.doi: 10.3761/j.issn.2096-7446.2025.05.004

• 儿科重症风险管理与实践 • 上一篇    下一篇

应用机器学习法构建及验证新生儿压力性损伤风险预测模型的研究

施卸丽(), 祝柯鑫, 胡小黎, 王芳()   

  1. 310006 杭州市 浙江大学医学院附属妇产科医院新生儿科（施卸丽，祝柯鑫），产五科（胡小黎），护理部（王芳）
  • 收稿日期:2024-09-13 出版日期:2025-05-10 发布日期:2025-04-28
  • 通讯作者: 王芳，E-mail：fangw@zju.edu.cn
  • 作者简介:施卸丽：女，硕士，主管护师，护士长，E-mail：selly112@zju.edu.cn
  • 基金资助:
    浙江省医药卫生科技计划项目(2022KY853);浙江省医药卫生科技计划项目(2022RC041)

Research on the construction and validation of a risk prediction model for neonatal pressure injury using machine learning methods

SHI Xieli(), ZHU Kexin, HU Xiaoli, WANG Fang()   

  1. Department of Neonatal Intensive Care Unit，Hospital of Obstetrics and Gynecology，Zhejiang University School of Medicine，Hangzhou，310006，China
  • Received:2024-09-13 Online:2025-05-10 Published:2025-04-28

RichHTML

PDF下载

摘要：

目的 应用机器学习法构建新生儿压力性损伤风险预测模型，验证并比较各模型的性能，以期为新生儿压力性损伤预测及早期防治提供参考。方法 2020年1月—2022年5月，选取浙江省某三级甲等专科医院新生儿重症监护病房的所有新生儿的临床资料形成数据集，将数据集按照8：2随机分成训练集和验证集，主要结局指标为是否发生压力性损伤。通过文献回顾和危险因素分析选取14个临床中常见的特征为预测因子，分别基于Logistic回归、随机森林（RF）、支持向量机（SVM）和决策树（DT）4种机器学习算法构建新生儿压力性损伤风险预测模型。模型通过五折交叉验证进行测试校准，通过准确率、精准率、召回率、F分数和受试者工作特征曲线下面积（AUC）评价和比较模型性能；采用决策曲线分析（DCA）验证模型的实用性；采用SHAP图解释最优机器学习模型的输出。结果 共纳入1 642例新生儿，其中发生压力性损伤251例，发生率为15.3%。4种机器学习构建的模型均有良好的预测性能，Logistic回归、随机森林、支持向量机和决策树模型的AUC值分别为0.85、0.88、0.90、0.84，准确率分别为0.78、0.80、0.83、0.79，精准率分别为0.72、0.75、0.77、0.73，召回率分别为0.91、0.93、0.94、0.91，F分数分别0.80、0.83、0.85、0.81。DCA曲线显示，在0.1~0.4的阈值概率下，随机森林模型的净收益最高。支持向量机和随机森林模型中前3位最重要的临床特征依次为经鼻持续气道正压通气的使用、出生孕周和出生体重。结论 机器学习可以用于构建新生儿压力性损伤风险预测模型，基于随机森林构建的预测模型性能与临床实际应用价值最高。

关键词: 机器学习, 新生儿, 压力性损伤, 预测模型

Abstract:

Objective To construct prediction models for the risk of neonatal pressure injury using machine learning，and compare the performance of the models，in order to provide insights for predicting，early prevention and treatment of neonatal pressure injury. Methods Clinical data of neonates from a tertiary class A hospital in Zhejiang Province were collected and randomly divided into a training set and a validation set according to a ratio of 8：2. The main outcome was whether pressure injury occurred. Through literature review and risk factor analysis，14 common clinical features were selected as predictors，and four machine learning algorithms，logistic regression，random forest，support vector machine and decision tree，were used to construct prediction models for the risk of neonatal pressure injury. The model was tested and calibrated by five-fold cross validation，and the model performance was evaluated and compared by accuracy，precision，recall，F score and area under the receiver operating characteristic curve（ROC AUC）. Decision curve analysis（DCA） was used to evaluate the clinical utility of the models，and the importance of the features was displayed by SHAP. Results A total of 1 642 neonates were included，251 of which developed pressure injury（15.3%）. All four prediction models had good predictive performance，and the AUC values of logistic regression，random forest，support vector machine and decision tree models were 0.85，0.88，0.90 and 0.84，respectively. The accuracy rates were 0.78，0.80，0.83 and 0.79，respectively. Accuracy were 0.72，0.75，0.77，0.73. The recall rates were 0.91，0.93，0.94，0.91，respectively. The F scores were 0.80，0.83，0.85 and 0.81，respectively. DCA showed that under the threshold probability of 0.1-0.4，the random forest model had the highest net benefit. The use of non-invasive respiratory support，gestational age at birth，and birth weight were the top three most important clinical characteristic variables for the SVM and random forest models. Conclusion Machine learning can be used to construct the risk prediction models of neonatal pressure injury. The model based on random forest is the best considering both on the performance and clinical utility.

Key words: Machine Learning, Neonatal, Pressure Injury, Predictive Model