Xu Fangfang, Ma Qianqing, Lai Penghao, Hu Lili, Gao Chuanfen, Xu Qianhua, Fang Youyan, Guo Yixuan, Yao Wen, Zhang Chaoxue
Department of Ultrasound, The First Affiliated Hospital of Anhui Medical University, Anhui, P. R. China.
Department of Ultrasound, Yijishan Hospital, The First Affiliated Hospital of Wannan Medical College, Wannan, P. R. China.
Reprod Biomed Online. 2025 May;50(5):104743. doi: 10.1016/j.rbmo.2024.104743. Epub 2024 Nov 28.
Can an optimal machine learning model be developed to predict reproductive outcomes following frozen embryo transfer (FET)?
This prospective study included 787 infertile females who underwent FET. The participants were split into a training cohort (n = 550) and a test cohort (n = 237) at a ratio of seven to three. Radiomics features were extracted from ultrasound images of the endometrium and junctional zone. A radiomics model was developed to generate the radiomics score (rad score). Logistic regression was applied to process the clinical data and create a clinical model. A fusion machine learning model was developed by integrating the rad score with independent clinical data using the XGboost algorithm. The performance of the models was compared using the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) method was used to interpret and visualize the contributions of features to the outcomes of FET.
The fusion model demonstrated superior performance, as indicated by an AUC of 0.861 (95% CI 0.829-0.890), in the training cohort, surpassing both the clinical model (AUC 0.680, 95% CI 0.635-0.722; P < 0.001) and the radiomics model (AUC 0.814, 95% CI 0.777-0.848; P < 0.001). The SHAP summary plot reveals the impacts of each feature on the predictive model, and the rad score was found to be the main feature. SHAP force plots provided explanations at the individual level.
An explainable machine learning model was established utilizing clinical data and ultrasound images to forecast the outcomes of FET. By utilizing the SHAP method, clinicians may better comprehend the contributors to the outcomes of FET in individual patients, and make better decisions before FET.
能否开发出一种最优的机器学习模型来预测冷冻胚胎移植（FET）后的生殖结局？
这项前瞻性研究纳入了787名接受FET的不孕女性。参与者按七比三的比例分为训练队列（n = 550）和测试队列（n = 237）。从子宫内膜和结合带的超声图像中提取放射组学特征。开发了一种放射组学模型以生成放射组学评分（rad评分）。应用逻辑回归处理临床数据并创建临床模型。通过使用XGboost算法将rad评分与独立临床数据整合，开发了一种融合机器学习模型。使用受试者操作特征曲线下面积（AUC）比较模型的性能。采用SHapley加性解释（SHAP）方法来解释和可视化特征对FET结局的贡献。
在训练队列中，融合模型表现出卓越的性能，AUC为0.861（95%CI 0.829 - 0.890），超过了临床模型（AUC 0.680，95%CI 0.635 - 0.722；P < 0.001）和放射组学模型（AUC 0.814，95%CI 0.777 - 0.848；P < 0.001）。SHAP汇总图揭示了每个特征对预测模型的影响，并且发现rad评分是主要特征。SHAP力的图在个体水平上提供了解释。
利用临床数据和超声图像建立了一种可解释的机器学习模型来预测FET的结局。通过使用SHAP方法，临床医生可以更好地理解个体患者FET结局的影响因素，并在FET前做出更好的决策。
