机器学习算法优化扳机日以改善体外受精结局
该研究基于2008至2019年7866例ICSI周期,用T-learner因果推断模型决定扳机当日或延后一日。遵循模型建议平均可多获3.015个2PN和1.515个可用囊胚,但医生与模型一致率仅52.57%和61.89%。
要点
- 模型建议较医生决策平均多获3.015个2PN(95%CI 2.626-3.371)。
- 模型建议平均多获1.515个可用囊胚(95%CI 1.134-1.871)。
- 医生与模型一致率低,算法辅助决策每周期净增1.430个2PN。
核心结论
- 使用该机器学习算法优化扳机时机,可能较医生决策显著增加IVF周期获得的2PN数和可用囊胚总数。
研究设计与样本
- 描述性与比较性新技术研究,纳入2008至2019年接受IVF联合ICSI的患者共7866例。
- 采用T-learner进行因果推断,以bagged决策树执行推断,决策框架为当日扳机或再等一日。
主要发现
- 遵循模型建议对2PN的平均获益为多3.015个(95%CI 2.626, 3.371)。
- 遵循模型建议对可用囊胚总数的平均获益为多1.515个(95%CI 1.134, 1.871)。
- 医生与模型的一致率分别为52.57%和61.89%,算法辅助扳机决策平均每刺激周期多获1.430个2PN和0.577个可用囊胚。
- 模型决策权重最高的特征依次为直径16-20 mm卵泡数、直径11-15 mm卵泡数和雌二醇水平。
关键数值与推荐
- 模型推荐当日扳机或再等一日,依据当日全部患者特征与刺激参数。
- 主要结局为2PN总数和可用囊胚总数较医生决策的平均改善。
临床怎么用
- 该算法可作为扳机日决策的辅助工具,帮助识别可能从延后扳机中获益的患者。
- 应用时需注意模型与医生决策存在不一致,且研究为回顾性因果推断,需前瞻性验证。
局限性
- 未来需前瞻性研究确认这些发现。
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
OBJECTIVE: To determine whether a machine learning causal inference model can optimize trigger injection timing to maximize the yield of fertilized oocytes (2PNs) and total usable blastocysts for a given cohort of stimulated follicles. DESIGN: Descriptive and comparative study of new technology. SETTING: Tertiary academic medical center. PATIENT(S): Patients undergoing IVF with intracytoplasmic sperm injection from 2008 to 2019 (n = 7,866). INTERVENTION(S): Causal inference was performed with the use of a T-learner. Bagged decision trees were used to perform inference. The decision was framed as either triggering on that day or waiting another day. All patient characteristics and stimulation parameters on a given day were used to determine the recommendation. MAIN OUTCOME MEASURE(S): Average outcome improvement in total 2PNs and usable blastocysts compared with the physician's decision. RESULT(S): For evaluation of average outcome improvement on 2PNs, the benefit of following the model's recommendation was 3.015 (95% CI 2.626, 3.371) more 2PNs. For total usable blastocysts, the benefit was 1.515 (95% CI 1.134, 1.871) more usable blastocysts. Given that the physicians-model agreement was 52.57% and 61.89%, respectively, algorithm-assisted trigger decisions yield, on average, 1.430 more 2PNs and 0.577 more total usable blastocysts per stimulation. The most important features weighted in the model's decision were the number of follicles 16-20 mm in diameter, the number of follicles 11-15 mm in diameter, and estradiol level, in that order. CONCLUSION(S): The use of this machine learning algorithm to optimize trigger injection timing may lead to a significant increase in the number of 2PNs and total usable blastocysts obtained from an IVF stimulation cycle when compared with physician decisions. Future research is required to confirm these findings prospectively.
引用
- Hariton E, Chi EA, Chi G, et al. A machine learning algorithm can optimize the day of trigger to improve in vitro fertilization outcomes. Fertil Steril. 2021. PMID: 34256948. doi:10.1016/j.fertnstert.2021.06.018
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。