The paper introduced focuses on the application of artificial intelligence in studying the association between embryo development and early pregnancy loss. It conducts a multicenter retrospective analysis of 37,717 embryos from 13 centers in France and Spain between 2017 and 2022 through AI annotation of 11 biological kinetic events, exploring the relationship between embryo kinetics and pregnancy outcomes. This research provides a new basis for identifying embryos prone to early pregnancy loss and is of great significance for improving the efficiency and accuracy of assisted reproductive technology.
Research Background
In assisted reproductive technology, embryo quality assessment is crucial for successful pregnancy, but traditional methods have obvious limitations. Embryologists rely on manual observation of the development process recorded by time-lapse systems (TLS), focusing on a few known kinetic events to select embryos. This approach is not only limited by sample size (most studies involve fewer than 500 embryos) but also subject to subjective biases due to inter-observer variations. Although deep learning algorithms have been applied in embryo assessment, their complex architecture makes them “black-box” models, which are difficult to interpret and sensitive to slight visual changes, resulting in insufficient stability. More importantly, previous studies mostly only compared non-pregnant and pregnant embryos, failing to identify embryos that can achieve early pregnancy but are prone to loss. This may lead to ineffective transfers in clinical practice, wasting resources and imposing physical and mental burdens on patients. Therefore, there is an urgent need in the field of assisted reproduction to develop an automated, accurate, and transparent model that can clearly reveal the association between embryo development and pregnancy outcomes.
Research Methods
To address the above issues, the research team conducted a multicenter retrospective analysis, whose research methods cover data collection, model construction and validation, statistical analysis, and other aspects. In terms of data collection, the study included 37,717 embryos from 13 clinical centers in France and Spain between 2017 and 2022. These embryos were derived from 7,028 oocyte retrieval cycles and recorded by three mainstream TLS devices, including Embryoscope or Embryoscope+, GERI, and MIRI. To ensure the rigor of the study, the team divided the data into two core datasets: Dataset A contains 9,606 embryos with known implantation data (KID), used to analyze the association between embryo kinetics and pregnancy outcomes; Dataset B contains 11,361 embryos (including discarded, frozen, and transferred embryos), used to validate the clinically practical value of the subsequently developed artificial intelligence model.
In model construction, the research team developed a computer vision model called Biological Events Extraction (BEE). This model was trained on 1,909 embryo videos (containing 14,696 manually annotated biological events) and can automatically identify 11 key kinetic events in embryo development, including cleavage times from 2-cell to 8-cell stages (t2-t8), morula formation time (tM), start of blastulation (tSB), and blastocyst formation time (tB). The F1 scores for the identification of each event range from 46% (t8) to 93% (t2), with a weighted average of 66%. It can be adapted to different TLS devices, effectively solving the standardization problem of manual annotation.
In terms of statistical analysis, the study used univariate and multivariate logistic regressions to analyze the association between 45 embryo development stages (such as t5-t2, tB-tSB) and pregnancy outcomes (non-early pregnancy, early pregnancy, early pregnancy loss, clinical pregnancy). Multivariate analysis also included clinical parameters such as oocyte age and body mass index (BMI) to exclude the interference of maternal factors, with P ≤ 0.05 as the criterion for significant results.
Research Results
Through the analysis of a large amount of data, the study revealed a clear association between embryo development kinetics and different pregnancy outcomes. Firstly, embryos with an overall faster development rate are more likely to achieve early pregnancy. Data show that the average time from 2-cell to blastocyst formation (tB-t2) for early pregnancy embryos is 77.18 ± 8.33 hours, which is significantly shorter than that of non-early pregnancy embryos (80.34 ± 8.76 hours, P < 0.001).
Secondly, embryos that can develop into clinical pregnancy exhibit a balanced development pattern of “deceleration during cleavage and acceleration during blastulation”: the average time for the cleavage stage (from 5-cell to the start of blastulation, tSB-t5) is 46.05 ± 13.60 hours, relatively slow; while the average time for the blastulation stage (from the start of blastulation to complete blastocyst formation, tB-tSB) is only 9.81 ± 5.04 hours, relatively fast.
In contrast, embryos with early pregnancy loss show an imbalanced pattern of “excessive acceleration during cleavage and excessive deceleration during blastulation”: the average time for their cleavage stage (tSB-t5) is 41.60 ± 17.08 hours, significantly faster than that of embryos with clinical pregnancy; the average time for their blastulation stage (tB-tSB) is 12.73 ± 5.69 hours, significantly slower than that of embryos with clinical pregnancy (both P < 0.001).
In addition, the AI alert system developed based on the above findings showed practical value in clinical tests. In Dataset B, the system issued alerts for 30% of the embryos, 71% of which had been discarded by embryologists in advance; among the transferred embryos, 78% of the embryos with alerts did not achieve clinical pregnancy, significantly higher than 65% of the embryos without alerts.
Innovation
This study demonstrates multiple innovations in the field of assisted reproduction. Firstly, it is the first to focus on the key issue of early pregnancy loss. Previous studies mostly only focused on whether embryos can result in pregnancy, while this study deeply distinguishes between embryos with early pregnancy and those with early pregnancy loss, revealing more refined development rules and providing a new perspective for precise embryo selection. Secondly, it realizes large-scale automatic annotation of embryo development stages. Through the BEE model, 9,606 KID embryos are processed, covering 45 development stages, which far exceeds the sample size and analysis dimensions of traditional studies, and reduces the subjectivity and differences in manual annotation. Thirdly, it constructs a transparent prediction model. Unlike “black-box” deep learning models, this model makes judgments based on clear kinetic events (such as tSB-t5, tB-tSB), with clear and interpretable decision logic, making it easier to gain clinical trust.
Research Limitations
Despite the significant achievements, the study still has some limitations. Firstly, due to the insufficient sample size of late pregnancy loss (loss after fetal heartbeat confirmation), the study cannot analyze the direct association between embryo kinetics and live birth, making it difficult to clarify the differences in embryo development characteristics between clinical pregnancy and live birth. Secondly, the imaging characteristics of different TLS devices may affect the results. When analyzing the data of a single device separately, due to the reduced sample size, the significance of some results disappears, requiring a larger sample size to verify the stability of the model among multiple devices. In addition, the performance of the AI alert system still has room for improvement. Its AUC value in transferred embryos is 0.60, which, although in line with the average level in the field, is still far from the ideal precise prediction.
Clinical Significance and Prospects
The artificial intelligence embryo kinetics analysis system constructed in this study provides a breakthrough tool for clinical practice in assisted reproduction: the automatic annotation of 11 development events based on the BEE model significantly improves the assessment efficiency. At the same time, by identifying the high-risk combination characteristics of “accelerated cleavage (tSB-t5 < 42 hours) and delayed blastulation (tB-tSB > 12 hours)”, it can accurately screen 78% of embryos at risk of early miscarriage, effectively reducing the rate of ineffective transfers. The revealed balanced development rule of “deceleration during cleavage (tSB-t5 ≈ 46 hours) and acceleration during blastulation (tB-tSB ≈ 10 hours)” establishes an objective and quantitative standard for embryo selection. In the future, it is necessary to further verify the universality of the algorithm in different culture devices, explore the direct association between kinetic characteristics and live birth rate, and promote the in-depth integration of the AI early warning system with clinical decision-making processes, thereby improving the practice path of precise reproductive medicine.
reference
Gidel-Dissler N, Roque T, Canat G, Angelard B, Vandame J, Boussommier-Calleja A. Association between embryo development and early pregnancy loss revealed by artificial-intelligence-annotated kinetic events. Reprod Biomed Online. 2024 Oct 20;51(3):104493. doi: 10.1016/j.rbmo.2024.104493. Epub ahead of print.