IVF Unit, Department of Obstetrics and Gynecology, Shamir Medical Center, Zerifin, Israel; Sackler School of Medicine, Tel Aviv University, Tel-Aviv, Israel.
FertilAi, Ramat Gan, Israel.
Fertil Steril. 2023 Nov;120(5):1004-1012. doi: 10.1016/j.fertnstert.2023.07.008. Epub 2023 Jul 23.
To develop a machine learning model designed to predict the time of ovulation and optimal fertilization window for performing intrauterine insemination or timed intercourse (TI) in natural cycles.
A retrospective cohort study.
A large in vitro fertilization unit.
PATIENT(S): Patients who underwent 2,467 natural cycle-frozen embryo transfer cycles between 2018 and 2022.
INTERVENTION(S): None.
MAIN OUTCOME MEASURE(S): Prediction accuracy of the optimal day for performing insemination or TI.
RESULT(S): The data set was split into a training set including 1,864 cycles and 2 test sets. In the test sets, ovulation was determined according to either expert opinion, with 2 independent fertility experts determining ovulation day ("expert") (496 cycles), or according to the disappearance of the leading follicle between 2 consecutive days' ultrasound examinations ("certain ovulation") (107 cycles). Two algorithms were trained: an NGBoost machine learning model estimating the probability of ovulation occurring on each cycle day and a treatment management algorithm using the learning model to determine an optimal insemination day or whether another blood test should be performed. The estradiol progesterone and luteinizing hormone levels on the last test performed were the most influential features used by the model. The mean numbers of tests were 2.78 and 2.85 for the "certain ovulation" and "expert" test sets, respectively. In the "expert" set, the algorithm correctly predicted ovulation and suggested day 1 or 2 for performing insemination in 92.9% of the cases. In 2.9%, the algorithm predicted a "miss," meaning that the last test day was already ovulation day or beyond, suggesting avoiding performing insemination. In 4.2%, the algorithm predicted an "error," suggesting performing insemination when in fact it would have been performed on a nonoptimal day (0 or -3). The "certain ovulation" set had similar results.
CONCLUSION(S): To our knowledge, this is the first study to implement a machine learning model, on the basis of the blood tests only, for scheduling insemination or TI with high accuracy, attributed to the capability of the algorithm to integrate multiple factors and not rely solely on the luteinizing hormone surge. Introducing the capabilities of the model may improve the accuracy and efficiency of ovulation prediction and increase the chance of conception.
HMC-0008-21.
开发一种机器学习模型，旨在预测自然周期中进行宫腔内授精或定时性交（TI）的排卵时间和最佳受精窗口。
回顾性队列研究。
一家大型体外受精单位。
2018 年至 2022 年间接受了 2467 个自然周期冷冻胚胎移植周期的患者。
无。
进行授精或 TI 的最佳日期的预测准确性。
数据集分为训练集（包括 1864 个周期）和 2 个测试集。在测试集中，排卵是根据两位独立的生育专家确定排卵日的专家意见（496 个周期）或根据连续两天的超声检查中主导卵泡的消失情况确定的（“确定排卵”）（107 个周期）。训练了两种算法：一种 NGBoost 机器学习模型，用于估计每个周期排卵发生的概率，以及一种使用学习模型确定最佳授精日期或是否应进行另一项血液检查的治疗管理算法。模型使用的最具影响力的特征是最后一次检测中的雌二醇、孕酮和黄体生成素水平。对于“确定排卵”和“专家”测试集，最后一次测试的平均测试次数分别为 2.78 和 2.85。在“专家”组中，该算法正确预测排卵，并建议在 92.9%的情况下在第 1 天或第 2 天进行授精。在 2.9%的情况下，该算法预测“错过”，这意味着最后一次测试日已经是排卵日或之后，建议避免进行授精。在 4.2%的情况下，该算法预测“错误”，建议在实际上应该在非最佳日（0 或-3）进行授精时进行授精。“确定排卵”组也有类似的结果。
据我们所知，这是第一项基于仅血液检测实施机器学习模型以高精度安排授精或 TI 的研究，这归因于算法整合多种因素的能力，而不是仅依赖于黄体生成素激增。引入模型的功能可能会提高排卵预测的准确性和效率，并增加受孕的机会。
HMC-0008-21。
