目的 探讨基于临床特征联合整合炎症生物标志物构建的5种机器学习模型对早产风险的预测价值，比较各模型的预测效能，为其临床应用提供依据。方法 采用回顾性队列研究方法，纳入2022年1月1日至2022年12月31日期间在中山大学附属第五医院妇产科分娩的1335例孕妇，包含全部早产类型（自发性早产+医源性早产）。通过文献回顾和临床观察，筛选出20项临床特征、10项实验室检查指标以及4项整合炎症生物标志物（中性粒细胞与淋巴细胞比值、血小板与淋巴细胞比值、全身炎症综合指数、白蛋白与淋巴细胞比值），并构建了5种机器学习模型（随机森林、XGBoost、支持向量机、Lasso回归和逻辑回归）。采用重复5折交叉验证优化超参数，从区分度、校准度和临床实用性3个维度进行模型评估。结果 5种机器学习模型的区分度，XGBoost模型的曲线下面积（AUC）最高（0.763），显著优于逻辑回归模型（AUC=0.683）（P<0.05）；随机森林（AUC=0.730）、支持向量机（AUC=0.707）、Lasso回归（AUC=0.696）依次递减。校准度方面，XGBoost模型的Brier分数最低（0.182），校准性能最优。决策曲线分析显示，XGBoost模型在20%风险阈值附近的标准化净获益最高，临床实用价值突出。加入整合炎症生物标志物后，模型AUC由0.763降至0.730（ΔAUC=-0.034，P=0.0033），未提升模型预测效能。结论 所构建的5种机器学习模型均可用于早产风险预测，其中XGBoost模型在区分度、校准度和临床实用性3个维度均表现最优，因其特异度极高（97.2%），可作为早产高危人群的二次确认与精细化风险分层工具，而非一线普遍筛查工具。整合炎症生物标志物在纳入的临床全早产人群中未能提供增量预测价值。

Objective To explore the predictive value of five machine learning models based on clinical characteristics and integrated inflammatory biomarkers for preterm birth，and to compare the predictive performance of different models in order to provide evidence for their clinical application. Methods A retrospective cohort study was conducted，including 1335 pregnant women who delivered in the Department of Obstetrics and Gynecology at the Fifth Affiliated Hospital of Sun Yat-sen University between January 1， 2022， and December 31， 2022，with all types of preterm birth included （spontaneous preterm birth and indicated preterm birth）. Totally 20 clinical characteristics，10 laboratory testing indexes，and 4 integrated inflammatory biomarkers（Neutrophil-to-Lymphocyte Ratio，Platelet-to-Lymphocyte Ratio，Aggregate Index of Systemic Inflammation，and Albumin-to-Lymphocyte Ratio）we screened out via literature review and clinical observation，and five machine learning models（Random Forest，XGBoost，Support Vector Machine，Lasso Regression，and Logistic Regression）were constructed. Hyperparameters were optimized via repeated 5-fold cross-validation，and the models were evaluated from 3 dimensions of discrimination，calibration，and clinical utility. Results The discrimination among the five models，the XGBoost model achieved the highest AUC（0.763），significantly higher than that of Logistic Regression（AUC=0.683）（P<0.05），followed by Random Forest（AUC=0.730），Support Vector Machine（AUC=0.707），and Lasso Regression（AUC=0.696）. In terms of calibration，the XGBoost model had the lowest Brier score（0.182），indicating the best calibration performance. Decision curve analysis showed that the XGBoost model had the highest standardized net benefit around the 20% risk threshold，indicating significant clinical value in practice.After adding integrated inflammatory biomarkers，the AUC of the model decreased from 0.763 to 0.730（ΔAUC=-0.034，P=0.0033），which did not improve predictive performance of the model. Conclusions All five machine learning models constructed in this study can be applied to preterm birth risk prediction， among which the XGBoost model demonstrates the best performance across the three dimensions of discrimination， calibration， and clinical utility. Owing to its extremely high specificity （97.2%）， it can serve as a tool for secondary confirmation and refined risk stratification in high-risk populations， rather than as a first-line universal screening tool. Integrated inflammatory biomarkers failed to provide incremental predictive value in the population included in this study.