基于机器学习预测模型探究薄型子宫内膜患者接受IVF/ICSI-ET治疗的早期流产风险因素

Risk factors for early miscarriage in patients with thin endometrium receiving in vitro fertilization/intracytoplasmic sperm injection-embryo transfer：a study based on machine learning-based predictive modeling

摘要:

目的 基于多种机器学习方法，探讨薄子宫内膜患者在新鲜胚胎移植中发生早期流产的影响因素，并建立预测模型，为预防薄子宫内膜患者在进行新鲜胚胎移植中发生早期流产提供合理的指导思路。方法 纳入了首次进行新鲜胚胎移植的薄子宫内膜患者1153例，通过LASSO回归和随机森林递归特征消除（recursive feature elimination，RFE）筛选特征，建立6种机器学习模型，通过交叉验证、准确度、敏感性、召回率、f1值、ROC曲线下面积及校准曲线比较不同模型的性能。SHAP图用于解释影响早期流产的因素。结果 通过LASSO回归和随机森林RFE筛选出29个特征变量纳入六种机器学习模型，其中多层感知机模型对早期流产的区分度最佳，ROC曲线下面积为0.803（95%CI=0.772~0.834）。随机森林、XGBoost和AdaBoost模型的ROC曲线下面积都高于0.7。结论 开发了薄子宫内膜患者在新鲜胚胎移植中是否发生早期流产的机器学习预测模型，各种评价指标的验证表明该模型的性能良好，有助于临床医生对该人群患者的早期诊断，为未来改善早期流产高危患者的妊娠结局提供指导思路。

Abstract:

Objective To investigate the influencing factors for early miscarriage in in patients with thin endometrium during fresh embryo transfer based on multiple machine learning methods，to establish a predictive model，and to provide reasonable ideas for preventing early miscarriage in patients with thin endometrium undergoing fresh embryo transfer.Methods A total of 1153 patients with thin endometrium who underwent fresh embryo transfer for the first time were enrolled in this study，and LASSO regression and random forest recursive feature elimination（RFE） were used for feature selection. Six machine learning models were developed and compared in terms of cross validation，accuracy，sensitivity，recall rate，f1 value，area under the ROC curve，and calibration curve. SHAP plots were used to elucidate the influencing factors for early miscarriage.Results A total of 29 feature variables were identified by LASSO regression and random forest RFE and were included in the six machine learning models，among which the multilayer perceptron model showed the best discriminatory ability for early miscarriage，with an area under the ROC curve of 0.803（95%CI=0.772-0.834）. The random forest，XGBoost，and AdaBoost models had an area under the ROC curve of >0.7.Conclusion This study establishes a machine learning-based predictive model for early miscarriage in patients with thin endometrium during fresh embryo transfer，and validation of various evaluation metrics shows that the model has good performance and can help clinicians to achieve the early diagnosis of patients，thereby providing ideas for improving the pregnancy outcome of patients at high risk of early miscarriage in the future.