利用机器学习模型中的低胎儿分数预测子痫前期
本研究纳入2019年6月13日至2023年1月6日在温州医科大学附属第二医院接受无创产前检测(NIPT)的孕妇,因低胎儿分数(FF)初始NIPT失败的61例为失败组(A组),其中重采血后获得有效结果37例(A1组)、重采血后仍失败24例(A2组),另随机选取119例NIPT成功者为对照组(C组)。A组子痫前期发生率显著高于C组;纳入FF的逻辑回归模型预测子痫前期的AUC为0.750(95%CI:0.639-0.860),纳入10个因素的XGBoost模型AUC为0.956(95%CI:0.868-1.000)、准确率0.944。结论认为低FF可能提示子痫前期风险增加,XGBoost模型预测效能良好,有潜力作为辅助产前筛查工具。
要点
- 因低FF导致NIPT初始失败的孕妇子痫前期发生率显著高于NIPT成功者。
- 重采血后是否成功获得有效结果,两组间子痫前期发生率无显著差异。
- XGBoost模型(含10个因素)预测子痫前期AUC为0.956(95%CI:0.868-1.000)。
核心结论
- NIPT中低FF可能提示子痫前期风险增加;无论重采血是否成功,因低FF初始NIPT失败的孕妇均需警惕子痫前期发生;XGBoost机器学习模型预测子痫前期效能良好,有潜力作为辅助产前筛查工具用于早期诊断。
研究设计与样本
- 回顾性研究,纳入2019年6月13日至2023年1月6日在温州医科大学附属第二医院接受NIPT的孕妇,因低FF初始NIPT失败者64例,失访3例后失败组(A组)61例,其中重采血后获得有效结果37例(A1组)、重采血后仍失败24例(A2组);同期随机选取NIPT成功(FF正常、无胎儿染色体异常)的119例孕妇为对照组(C组)。
主要发现
- A组子痫前期发生率显著高于C组(p < 0.05)。
- A1组与A2组子痫前期发生率无显著差异。
- 纳入FF的逻辑回归模型预测子痫前期AUC为0.750(95%CI:0.639-0.860),灵敏度0.875,特异度0.727。
- 纳入10个因素(FF、年龄、体重、BMI、采样时孕周、采样时收缩压/舒张压、辅助生殖技术史、分娩史、肝素使用史)的XGBoost模型表现更优(AUC = 0.956,95%CI:0.868-1.000;准确率 = 0.944)。
- 最重要的前三个因素为收缩压、舒张压和FF。
临床怎么用
- 作者认为,NIPT中低FF可能提示子痫前期风险增加,无论重采血是否成功,因低FF初始NIPT失败的孕妇均需警惕子痫前期发生;XGBoost机器学习模型有潜力作为辅助产前筛查工具用于早期诊断。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVE: To investigate the association between low fetal fraction (FF) in non-invasive prenatal testing (NIPT) and pregnancy complications or adverse pregnancy outcomes. METHODS: Sixty-four pregnant women undergoing NIPT at the Second Affiliated Hospital of Wenzhou Medical University between 13 June 2019 and 6 January 2023 had an initial NIPT failure due to low FF. Three cases were lost to follow-up, leaving 61 cases in the failure group (Group A). Group A was subdivided into 37 cases with a valid result after redraw (Group A1) and 24 cases remaining unsuccessful after redraw (Group A2). Concurrently, 119 pregnancies with successful NIPT (normal FF, no fetal chromosomal abnormalities) were randomly selected as controls (Group C). Logistic regression and XGBoost models were established, and their area under the curve (AUC), sensitivity, and specificity were calculated and compared. RESULTS: The incidence of pre-eclampsia was significantly higher in Group A than in Group C (p < 0.05). No significant difference in the incidence of pre-eclampsia was found between Groups A1 and A2. A logistic regression model incorporating FF predicted pre-eclampsia with an AUC of 0.750 (95% CI: 0.639-0.860), sensitivity of 0.875, and specificity of 0.727. An XGBoost model incorporating 10 factors (FF, age, weight, BMI, gestational age, systolic/diastolic blood pressure at sampling, ART history, delivery history, heparin use history) demonstrated superior performance (AUC = 0.956, 95% CI: 0.868-1.000; accuracy = 0.944). The top three important factors were systolic blood pressure, diastolic blood pressure, and FF. CONCLUSIONS: Low FF in NIPT may indicate an increased risk of pre-eclampsia. Regardless of the success of redraw, pregnancies with initial NIPT failure due to low FF warrant vigilance for pre-eclampsia development. The XGBoost machine learning model demonstrates good efficacy for predicting pre-eclampsia and has potential as an adjunctive prenatal screening tool for early diagnosis.
引用
- Wang J, Bai Y, Huang S, et al. The Prediction of Pre-Eclampsia Using Low Fetal Fraction in a Machine Learning Model. Prenat Diagn. 2025. PMID: 41309525. doi:10.1002/pd.70033
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。