目的 应用机器学习法构建新生儿中等长度导管（midline catheter，MC）并发症风险预测模型，验证并对比各模型性能，为新生儿MC并发症预测及早期防控提供参考。 方法 回顾性收集2020年8月至2023年12月在安徽省蚌埠市某三级甲等综合医院新生儿重症监护室接受MC置入的243例患儿的临床资料，包括新生儿一般资料、基础疾病、置管相关信息等。将研究对象按照7∶3的比例随机分为训练集170例和测试集73例。借助递归特征消除算法实现特征的初步筛选，采用逻辑回归（logistic regression，LR）、随机森林（random forest，RF）、支持向量机（support vector machine，SVM）、最小绝对收缩与选择算子（LASSO）和极端梯度提升（extreme gradient boosting，XGBoost）5种机器学习算法构建预测模型，通过交叉验证优化模型参数，比较不同模型的性能，选择最佳模型进行内部验证，并构建列线图。 结果 243例MC置入患儿中有37例（15.23%）发生并发症。在训练集中，5种模型均具有较好的拟合度，其中RF模型受试者操作特征曲线下面积（area under curve，AUC）最大（0.843），特异度（0.916）、阳性预测值（0.613）、准确率（0.882）、F1值（0.655）优于其他模型；在测试集中，RF模型在AUC（0.881）、特异度（0.937）、阳性预测值（0.636）、准确率（0.904）、F1值（0.666）方面也均优于其他模型。RF模型中，前5位最重要的临床特征依次是出生体重、导管尖端位置、胎龄、送管是否顺利、穿刺部位。 结论 基于机器学习构建的新生儿MC并发症风险预测RF模型性能较优，列线图有助于临床护理人员及早识别MC并发症高风险患儿，进而采取针对性的预防措施。

Objective To construct a risk prediction model for complications of midline catheters（MC） in neonates using machine learning methods，verify and compare the performance of various models，and provide a reference for the prediction and early prevention of MC complications in neonates. Methods The clinical data of 243 neonates who received MC insertion in the neonatal intensive care unit of a provincial tertiary general hospital in Bengbu City，Anhui Province from August 2020 to December 2023 and met the inclusion and exclusion criteria were retrospectively collected，including general neonatal information，underlying diseases，catheterization-related information，etc. The research subjects were randomly divided into a training set of 170 cases and a testing set of 73 cases in a 7：3 ratio. Recursive feature elimination algorithm was used to initially screen the features，and five machine learning algorithms，including logistic regression，random forest，support vector machine，least absolute shrinkage and selection operator，and extreme gradient boosting，were used to construct the prediction models. Model parameters were optimized through cross-validation，and the performance of different models was compared to select the best model for internal validation and to construct a nomogram. Results Among the 243 neonates with MC insertion，37（15.23%） developed complications. In the training set，all five models had good fitting degrees，among which the RF model had the largest AUC（0.843），specificity（0.916），PPV（0.613），accuracy（0.882），and F1 value（0.655），outperforming the other models. In the test set，the RF model also outperformed the other models in terms of AUC（0.881），specificity（0.937），PPV（0.636），accuracy（0.904），and F1 value（0.666）. In the RF model，the top five most important clinical features were birth weight，catheter tip position，gestational age，smoothness of catheter insertion，and puncture site. Conclusion The RF model for predicting the risk of MC complications in neonates based on machine learning has certain accuracy and practicality. The nomogram is helpful for clinical medical staff to identify high-risk neonates in advance and take targeted preventive measures.