机器学习预测足月死产的局限性:意大利北部302192例妊娠人群队列
该研究基于意大利艾米利亚-罗马涅大区2014-2023年出生登记数据,对302192例足月妊娠(322例死产,发生率1.07/1000)进行二次分析,训练三种机器学习模型预测足月死产。测试集(n=60438,66例死产)上,弹性网逻辑回归AUC为0.658,XGBoost为0.617,随机森林为0.534,均低于临床筛查常用的AUC 0.70阈值,决策曲线分析未显示任何模型在临床合理阈值下优于全部干预策略。结论认为静态登记变量不足以实现临床有用的足月死产预测。
要点
- 三种机器学习模型预测足月死产的AUC在0.53至0.66之间,均低于0.70的临床筛查阈值。
- 弹性网逻辑回归显著优于随机森林(DeLong p=0.001),但阳性预测值接近人群基线0.106%。
- 决策曲线分析显示无模型在临床合理阈值下产生优于全部干预策略的净获益。
核心结论
- 三种机器学习模型对足月死产的区分能力有限(AUC 0.53至0.66),静态登记变量信号不足。
- 简单正则化回归显著优于随机森林,但所有模型均未达到临床筛查可用水平。
研究设计与样本
- 基于艾米利亚-罗马涅大区出生登记2014-2023年302192例足月妊娠的二次分析,含322例死产(发生率1.07/1000)。
- 采用分层80%训练集(类重平衡ROSES)和20%测试集(n=60438,66例死产),评估弹性网逻辑回归、随机森林和XGBoost。
- 性能指标包括AUC-ROC、AUC-PR、敏感度、特异度、预测值(Wilson 95%CI)、校准截距与斜率、O:E比和决策曲线分析。
- 敏感性分析包括与Salerno(2023)逻辑回归头对头比较、完整病例与缺失置零分析、特征消融、无类重平衡重新拟合、bootstrap乐观校正和时间验证。
主要发现
- 测试集上弹性网逻辑回归AUC 0.658(0.589-0.726),XGBoost AUC 0.617(0.544-0.690),随机森林AUC 0.534(0.470-0.597)。
- 逻辑回归显著优于随机森林(DeLong p=0.001);Youden最优阈值下阳性预测值接近人群基线0.106%。
- 决策曲线分析显示无模型在临床合理阈值下净获益优于全部干预策略。
- 校准后逻辑回归O:E比1.12(95%CI 0.88-1.43);孕周尤其是早期足月分娩为最有影响力预测因子。
- 时间验证使所有模型AUC降至0.57至0.58;Salerno(2023)逻辑回归在同一测试集AUC为0.637。
关键数值与推荐
- AUC 0.70为临床筛查阈值,所有模型均未达到。
- 弹性网逻辑回归AUC 0.658(0.589-0.726),XGBoost AUC 0.617(0.544-0.690),随机森林AUC 0.534(0.470-0.597)。
- 阳性预测值在Youden最优阈值下与人群基线0.106%的Wilson区间距离内。
- 校准后逻辑回归O:E比1.12(95%CI 0.88-1.43)。
- 时间验证AUC 0.57至0.58;Salerno(2023)逻辑回归AUC 0.637。
临床怎么用
- 静态登记变量构建的机器学习模型尚不能用于足月死产的临床筛查。
- 简单正则化回归可作为基线模型,但需谨慎解读其预测值。
- 未来研究应转向动态、基于生理学的监测数据以提升预测能力。
局限性
- 仅使用静态登记变量,未纳入动态生理学指标。
- 死产事件数较少(测试集66例),可能影响模型稳定性。
- 时间验证显示性能下降,提示模型随时间推移区分能力降低。
- 未评估其他潜在预测因子或外部人群验证。
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
BACKGROUND: Stillbirth at term is a major public-health concern. In high-income settings, the most commonly identifiable causes are unrecognised fetal growth restriction (about 14%) and placental abruption (about 12%), but 20 to 40% of term stillbirths remain unexplained. Whether modern machine learning can substantively improve prediction over established clinical risk-factor models is unresolved. OBJECTIVE: To evaluate the predictive performance of three machine-learning models, elastic-net logistic regression, random forest, and extreme gradient boosting, for stillbirth at term in a population-based cohort, and to quantify the achievable ceiling using static registry variables. STUDY DESIGN: Secondary analysis of 302,192 term pregnancies (322 stillbirths; rate 1.07 per 1,000), recorded in the Emilia-Romagna "Certificato di Assistenza al Parto" birth registry between 2014-2023. Three machine-learning models were trained on a stratified 80% training set with class rebalancing by Random Over-Sampling Examples applied within each cross-validation fold. Logistic-regression predicted probabilities were rescaled to natural prevalence by intercept-only Platt-style recalibration on a held-out training slice. Performance was evaluated on the held-out 20% test set (n=60,438; 66 stillbirths) using area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUC-PR), sensitivity, specificity, predictive values (Wilson 95% CIs), calibration intercept and slope, observed:expected (O:E) ratio, and decision curve analysis. Sensitivity analyses included a head-to-head comparison with the previously published Salerno (2023) logistic regression specification, complete-case and missing-as-zero analyses, feature ablation, refitting without class rebalancing, bootstrap optimism correction, and temporal validation (training 2014-2021; test 2022-2023). RESULTS: On the test set, elastic-net logistic regression achieved AUC 0.658 (0.589-0.726), XGBoost AUC 0.617 (0.544-0.690), and random forest AUC 0.534 (0.470-0.597). All three fell below the AUC of 0.70 threshold for clinical screening. Logistic regression significantly outperformed random forest (DeLong p=0.001). Positive predictive values at Youden-optimal thresholds were within Wilson-interval distance of the population baseline (0.106%). Decision curve analysis showed no model produced net benefit superior to a treat-all strategy at clinically plausible thresholds. After intercept recalibration, the logistic regression was well calibrated in the large (observed:expected ratio 1.12, 95% CI 0.88 to 1.43). Gestational age, particularly early-term delivery, was the most influential predictor across models. Temporal validation degraded all three AUCs to 0.57 to 0.58. The previously published Salerno (2023) logistic regression specification achieved AUC 0.637 on the same test set, only 0.02 below the elastic-net logistic regression. CONCLUSIONS: In a large, decade-spanning population cohort, three machine-learning models achieved similarly limited discrimination for term stillbirth (AUC 0.53 to 0.66), with the simpler regularised regression significantly outperforming the random forest. Static registry variables, even those captured by a high-quality regional surveillance system, do not carry sufficient signal for clinically useful term-stillbirth prediction. The empirical ceiling defined by this study should motivate a shift toward dynamic, physiology-based surveillance for future prediction research.
引用
- Benuzzi M, Kawakita T, Monari F, et al. Limits of Machine Learning for Predicting Stillbirth at Term: A Population-Based Cohort of 302,192 Pregnancies in Northern Italy. Am J Obstet Gynecol MFM. 2026. PMID: 42858874. doi:10.1016/j.ajogmf.2026.102148
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。