摘要
目的:为使用经济、简单、便捷的方法早期识别妊娠期糖尿病(GDM)高风险孕妇,开发并评估多种机器学习模型,筛选出最优疾病预测模型并构建此模型的临床决策支持系统(CDSS)。方法:纳入2023年1月1日至2024年12月30日在大连医科大学附属第二医院就诊的464例孕产妇,其中386例用于建立预测模型(训练集231例、测试集155例),另外78例作为验证集。采用点双列相关和卡方检验的方法,筛选出特征变量后构建4种机器学习模型:逻辑回归、随机森林、支持向量机和极限梯度提升(XGBoost)。初步判断最大权重模型,再进一步比较各模型的区分能力、校准能力、临床实用性以评估选择出最优模型,开发其CDSS,并验证模型的准确性。结果:①相关性分析发现,年龄、孕前体质量指数、收缩压、舒张压、白细胞计数、血红蛋白、淋巴细胞比率、空腹血糖、尿酸、直接胆红素、慢性高血压合并妊娠、辅助生殖受孕为GDM的预测因子。②XGBoost在集成模型中占主导地位;其区分能力、校准能力和临床实用性在4种模型中均表现最佳,AUC值为0.931(95%CI 0.910~0.967)。③通过CDSS验证XGBoost模型的准确度78.2%,敏感度64.7%,特异度82.0%。结论:XGBoost模型早期预测GDM能力最高,开发其CDSS不仅便于医生快速评估GDM风险,还适宜推广至偏远地区,通过远程数据实现高风险人群筛查。
Objective:To develop an economical,simple,and accessible method for early identification of high-risk pregnant women with gestational diabetes mellitus(GDM),this study developed and evaluated multiple machine learning models,identified the optimal prediction model,and constructed a clinical decision support system(CDSS)based on this model.Methods:A total of 464 pregnant women who visited the Second Affiliated Hospital of Dalian Medical University from January 1,2023 to December 30,2024 were included,of which 386 were used to establish a prediction model(231 in the training set and 155 in the testing set),and the remaining 78 were used as a validation.Adopting the methods of double-point sequence correlation and chi-square test,four machine learning models were constructed after selecting feature variables:Logistic Regression,Random Forest,Support Vector Machine,and eXtreme Gradient Boosting(XGBoost).Preliminary judgment of the maximum weight model,further comparison of the discriminative ability,calibration ability,and clinical practicality of each model to evaluate and select the optimal model,develop its CDSS,and verify the accuracy of the model.Results:①Correlation analysis identified predictors of GDM:age,pre-pregnancy body mass index(BMI),systolic/diastolic blood pressure,white blood cell count,hemoglobin,lymphocyte ratio,fasting plasma glucose,uric acid,direct bilirubin,chronic hypertension complicating pregnancy,and assisted reproductive technology conception.②XGBoost dominated the ensemble model and demonstrated the best performance in discrimination(AUC 0.931,95%CI 0.910-0.967),calibration,and clinical utility among the four models.③The CDSS achieved an accuracy of 78.2%,sensitivity of 64.7%,and specificity of 82.0%in the XGBoost model.Conclusions:The XGBoost model has the highest ability to predict GDM in the early stage.Developing its CDSS not only facilitates doctors to quickly assess GDM risk,but also is suitable for promotion to remote areas,where high-risk population screening can be achieved through remote data.
作者
刘家琦
高家震
孟燕妮
王畅
郑东颖
王丽霞
LIU Jiaqi;GAO Jiazhen;MENG Yanni(Obstetrics Department,Second Affiliated Hospital of Dalian Medical University,Dalian Liaoning 116021,China;First Operating Room,Second Affiliated Hospital of Dalian Medical University,Dalian Liaoning 116021,China;Department of Gynaecology and Obstetrics,Dongguan Huangjiang Hospital,Dongguan Guangdong 523750,China)
出处
《实用妇产科杂志》
北大核心
2025年第11期915-921,共7页
Journal of Practical Obstetrics and Gynecology
基金
国家自然科学基金(编号:82401990)。
关键词
妊娠期糖尿病
极限梯度提升
早期预测
临床决策支持系统
Gestational diabetes mellitus
Extreme gradient boosting
Early prediction
Clinical decision support system
