深度学习胚胎排序预测活产时间:一种新的多重插补方法
深度学习算法平均活产时间为1.68次移植(95%CI:1.63-1.72),比人工排序的1.78次(95%CI:1.73-1.83)缩短6.1%。
要点
- 深度学习算法平均活产时间为1.68次移植,较人工排序的1.78次缩短6.1%。
- 算法在人群水平的AUC为0.633(95%CI:0.620-0.645),治疗水平平均性能为0.672(95%CI:0.656-0.687)。
- 该方法依赖对缺失结局的准确预测,估计的活产时间仅能近似实际临床效用。
核心结论
- 研究提示胚胎选择算法可能通过提高活产可能性来改善体外受精治疗的临床结局,从而提高辅助生殖技术的有效性和成功率。
研究设计与样本
- 回顾性队列研究,纳入大学附属私立体外受精中心2015年至2022年进行的治疗。
- 采用链式方程多重插补(MICE)程序估计未移植胚胎的结局。
主要发现
- 深度学习算法的平均活产时间为1.68次移植(95%CI:1.63-1.72),比人工排序的1.78次移植(95%CI:1.73-1.83)缩短6.1%。
- 在人群水平评估算法排序性能的受试者工作特征曲线下面积(AUC)为0.633(95%CI:0.620-0.645)。
- 在治疗水平的平均性能为0.672(95%CI:0.656-0.687)。
临床怎么用
- 原文未据此给出推荐。
局限性
- 所提出的方法依赖于对缺失结局的准确预测。
- 除胚胎质量外,许多其他因素也会影响实际结局,因此估计的活产时间只能近似实际临床效用。
- 本研究估计的活产时间仅适用于特定数据集。
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
STUDY QUESTION: What is the clinical utility of embryo selection algorithms in estimating the time to live birth (TTLB)? SUMMARY ANSWER: Using multiple imputations, the clinical utility of embryo selection algorithms was estimated, showing their potential to improve prediction of TTLB compared to manual ranking. WHAT IS KNOWN ALREADY: Estimation of the clinical utility of embryo selection algorithms is challenging due to the unknown outcomes of non-transferred embryos. Existing studies have tried to mitigate the unknown outcome by using a biased subsample of the observed treatments or constructing treatments based on embryos from different patients. STUDY DESIGN, SIZE, DURATION: Retrospective cohort study in a university-affiliated private IVF center. Treatments performed from 2015 through 2022 were included. PARTICIPANTS/MATERIALS, SETTING, METHODS: Three thousand seven hundred and eighty-three treatments with a total of 17 914 usable embryos, among these, 7571 embryos were transferred in either fresh or vitrified/warmed transfers. The outcome of non-transferred embryos was estimated using the multiple imputation by chained equations (MICE) procedure. MAIN RESULTS AND THE ROLE OF CHANCE: The average TTLB for the deep learning algorithm was 1.68 (95% CI: 1.63-1.72) transfers which was 6.1% shorter than the average TTLB of 1.78 (95% CI: 1.73-1.83) transfers for manual ranking. Measuring the algorithm ranking performance on the population level resulted in an area under the receiver operating characteristic curve (AUC) of 0.633 (95% CI: 0.620-0.645) while the average performance on the treatment level was 0.672 (95% CI: 0.656-0.687). LIMITATIONS, REASONS FOR CAUTION: The proposed methods rely on an accurate prediction of the missing outcomes. As many other factors besides embryo quality will impact the actual outcome, the estimated TTLB can only approximate the actual clinical utility. In addition, the estimated TTLB in this study is only applicable to the specific dataset. WIDER IMPLICATIONS OF THE FINDINGS: This study suggests that embryo selection algorithms could potentially improve clinical outcomes in IVF treatments by increasing the likelihood of live births compared to traditional morphological gradings hence increasing the efficacy and success rates of assisted reproductive technology. STUDY FUNDING/COMPETING INTEREST(S): No funding was received to support this study. L.B. reports Contract in Fundación IVI, Instituto de Investigación Sanitaria, La Fe (Valencia, Spain). M. J. is an employee at Vitrolife A/S. J.B. is an employee and shareholder at Vitrolife A/S. E.T.P. has no financial or personal interests related to the research. M.M. reports funding from ISCIII (PI21/00283), cofounded by ERDF, 'A way to make Europe' awarded to M.M. for the submitted work; Contract in Fundación IVI, Instituto de Investigación Sanitaria, La Fe (Valencia, Spain), and Contract in IVIRMA Valencia; and honoraria from various companies, including Merck, Vitrolife, MSD Ferring, AiVF, Theramex, Gedeon Richter, Genea Biomedx, and Life Whisperer. TRIAL REGISTRATION NUMBER: N/A.
引用
- Bori L, Johansen MN, Berntsen J, et al. Predicting time to live birth with deep learning embryo ranking: a novel multiple imputation approach. Hum Reprod. 2025. PMID: 40514039. doi:10.1093/humrep/deaf102
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。