机器学习分析冷冻复苏单整倍体胚胎移植临床妊娠因素
识别出7个影响临床妊娠的因素,模型AUC在0.760至0.778之间。
要点
- 既往周期数、AMH、内膜厚度等7个因素与临床妊娠相关
- XGBoost模型AUC最高(0.778),AdaBoost最低(0.760)
- XGBoost与随机森林的AUC差异显著(P=0.048)
核心结论
- 机器学习模型识别了整倍体胚胎移植周期中临床妊娠的关键决定因素,可能有助于风险分层和优化IVF治疗策略。
研究设计与样本
- 回顾性队列研究,纳入2011年10月至2023年2月在土耳其Sisli Memorial医院生殖遗传中心进行的4300个冷冻复苏单整倍体胚胎移植周期,均来自ICSI或IMSI。
- 采用五折交叉验证、F1分数和AUC评估模型性能,使用SHAP解释模型输出。
主要发现
- 识别出7个影响临床妊娠的临床相关因素:既往周期数、抗苗勒管激素浓度、内膜厚度、复苏后胚胎质量、母亲年龄、冷冻胚胎数和内膜准备方案。
- 模型区分性能大多相当,AUC值范围从0.760(AdaBoost)到0.778(XGBoost)。
- 校准分析显示,在全特征设置下随机森林和LightGBM表现最佳(Brier分数分别为0.178和0.179),而在精选特征设置下XGBoost校准最优(Brier分数0.201)。
- 配对bootstrap分析(1000次迭代)发现XGBoost与随机森林之间AUC差异显著(P=0.048),其他配对比较无显著差异。
临床怎么用
- 可基于识别的7个因素进行风险分层,辅助个体化咨询和IVF策略优化。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
RESEARCH QUESTION: Which factors influence clinical pregnancy outcomes in vitrified-warmed single euploid embryo transfer cycles using machine learning models? DESIGN: This retrospective cohort study included 4300 vitrified-warmed single euploid embryo transfer cycles derived exclusively from intracytoplasmic sperm injection or intracytoplasmic morphologically selected sperm injection performed at the Assisted Reproductive Technology and Reproductive Genetics Centre of Sisli Memorial Hospital, Istanbul, Turkey between October 2011 and February 2023. Twenty-six clinical, demographic and embryological variables were analysed using multiple machine learning algorithms, namely Adaptive Boosting (AdaBoost), Random Forest, Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine, and Extremely Randomized Trees. Model performance was evaluated using five-fold cross-validation, F1-score and area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations were used to interpret model outputs. RESULTS: Seven clinically relevant factors influencing clinical pregnancy were identified: number of previous cycles, anti-Müllerian hormone concentration, endometrial thickness, post-warming embryo quality, maternal age, number of vitrified embryos, and endometrial preparation method. Discriminatory performance was mostly comparable across models, with AUC values ranging from 0.760 (AdaBoost) to 0.778 (XGBoost). Calibration analysis demonstrated that Random Forest and LGBM achieved the best performance in the full feature setting (Brier scores 0.178 and 0.179, respectively), whereas XGBoost showed optimal calibration in the selected feature setting (Brier score 0.201). Pairwise bootstrap analysis (1000 iterations) identified a significant AUC difference between XGBoost and Random Forest (P = 0.048), with no other significant pairwise differences. CONCLUSIONS: Machine learning models identified key determinants of clinical pregnancy in euploid embryo transfer cycles. These insights may facilitate risk stratification and optimize IVF treatment strategies.
引用
- Ozer G, Pehlivanli AC, Deniz E, et al. Machine learning-based analysis of factors associated with clinical pregnancy in vitrified-warmed single euploid embryo transfer cycles. Reprod Biomed Online. 2026. PMID: 42566879. doi:10.1016/j.rbmo.2026.105864
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。