基于随机森林算法的宫腔粘连术后妊娠结局 预测模型

Prediction Model for Pregnancy Outcomes after Hysteroscopic Adhesiolysis Based on Random Forest Algorithm

摘要: 目的：本研究旨在通过随机森林算法构建“影像–分子”双模态模型预测宫腔粘连术后的妊娠结局。方法：纳入35例符合标准且行TCRA治疗的IUA患者，收集四维超声参数(内膜厚度和血流分级)和免疫组化参数(ER和PR表达)，采用随机森林算法构建预测模型，按照7.5:2.5的比例划分训练集(26例)与测试集(9例)评估模型的预测性能与特征贡献度。结果：训练集的灵敏度为100%，特异度为40%，准确度为77%，AUC = 0.81；测试集的灵敏度为86%，特异度为50%，准确度为78%，AUC = 0.79。根据特征贡献度分析显示，从高到低贡献度依次为血流分级1，PR受体低，内膜厚度B、内膜厚度C、ER受体低、ER受体高、PR受体高、血流分级3、血流分级2。结论：随机森林算法在IUA术后妊娠结局预测中具备良好潜力，训练集AUC达0.81，测试集AUC达0.79，验证了“影像–分子”双模态整合的有效性，当前模型泛化能力受限于小样本与特征稀疏性，扩大样本量、优化特征工程后有望成为临床个体化决策的有效工具。

Abstract: Objective: This study aims to construct a “imaging-molecular” dual-modal model using the random forest algorithm to predict pregnancy outcomes after hysteroscopic adhesiolysis. Methods: Thirty-five patients with intrauterine adhesions (IUA) who met the criteria and underwent transcervical resection of adhesions (TCRA) were enrolled. Four-dimensional ultrasound parameters (endometrial thickness and blood flow grading) and immunohistochemical parameters (estrogen receptor and progesterone receptor expression) were collected. A predictive model was developed using the random forest algorithm, with the data divided into a training set (26 cases) and a test set (9 cases) at a 7.5:2.5 ratio to evaluate the model’s predictive performance and feature contribution. Results: The training set showed sensitivity of 100%, specificity of 40%, accuracy of 77%, and AUC = 0.81; the test set demonstrated sensitivity of 86%, specificity of 50%, accuracy of 78%, and AUC = 0.79. Feature contribution analysis revealed the following descending order of importance: blood flow grading 1, low progesterone receptor, endometrial thickness B, endometrial thickness C, low estrogen receptor, high estrogen receptor, high progesterone receptor, blood flow grading 3, and blood flow grading 2. Conclusion: The random forest algorithm demonstrates promising potential in predicting pregnancy outcomes after IUA surgery, with an AUC of 0.81 in the training set and 0.79 in the test set, validating the effectiveness of “imaging-molecular” dual-modal integration. However, the current model’s generalization capability is limited by small sample size and feature sparsity. Expanding sample size and optimizing feature engineering may enable it to become an effective tool for clinical individualized decision-making.