Development of a dynamic machine learning algorithm to predict clinical pregnancy and live birth rate with embryo morphokinetics.

Objective: To evaluate the feasibility of generating a center-specific embryo morphokinetic algorithm by time-lapse microscopy to predict clinical pregnancy rates. Design: A retrospective cohort analysis. Setting: Academic fertility clinic in a tertiary hospital setting. Patients: Patients who underwent in vitro fertilization with embryos that underwent EmbryoScope time-lapse microscopy and subsequent transfer between 2014 and 2018. Interventions: None. Main Outcome Measures: Clinical pregnancy.
Results: A supervised, random forest learning algorithm from 367 embryos successfully predicted clinical pregnancy from a training set with overall 65% sensitivity and 74% positive predictive value, with an area under the curve of 0.7 for the test set. Similar results were achieved for live birth outcomes. For the secondary analysis, embryo growth morphokinetics were grouped into five clusters using unsupervised clustering. The clusters that had the fastest morphokinetics (time to blastocyst = 97 hours) had pregnancy rates of 54%, whereas a cluster that had the slowest morphokinetics (time to blastocyst = 122 hours) had a pregnancy rate of 71%, although the differences were not statistically significant (P=.356). Other clusters had pregnancy rates of 51%-60%. Conclusions: This study shows the feasibility of a clinic-specific, noninvasive embryo morphokinetic simple machine learning model to predict clinical pregnancy rates.

Discuss: You can discuss this article with its authors and other readers at https://www.fertstertdialog.com/posts/xfre-d-21-00181

Embryo quality assessment has historically relied on a variety of snapshot morphological parameters to determine the embryos that have the highest probability of resulting in pregnancy (1), including morphology of the pronuclear stage (2), early embryo cleavage (3, 4), the presence of multinucleation (5, 6), and zona pellucida characteristics (7). However, standard grading only provides a brief view of embryonic morphology at a specific time point, while developmental changes over multiple time points may provide a more robust impression of the implantation potential (8). To obtain objective, real-time embryo morphologic data with potential clinical relevance, we focused on time-lapse microscopy as a noninvasive method to characterize embryo quality. Many studies have identified parameters associated with improved clinical outcomes, including time of appearance and fading of two pronuclei (9), the timing of cell cycle events (10, 11), the time between 2- to the 8-cell cleavage stage (12, 13, 14, 15), cleavage events (16), cytokinesis duration (17), time to initiation of blastulation (18, 19, 20), and combined morphokinetic and morphologic or molecular parameters (21, 22, 23, 24, 25, 26, 27). Other studies showed that no differences were found in the time-lapse parameters of euploid embryos that led to miscarriages (28). Given the divergent findings, site-specific algorithms have been proposed (29).

Recently, machine learning has also been employed to predict clinical pregnancy using demographic and clinical data alongside embryo scoring qualities as training data (30, 31, 32). Other machine learning techniques aim to predict embryo morphological grading using inputs of still images or videos with high accuracy (33, 34) and also to predict pregnancy rates with high accuracy (35, 36, 37). Numerous variables, including demographics, clinical data, patterns in imaging, or video data, have been used to train artificial intelligence (AI) algorithms. Many AI approaches begin with a multistep process of first extracting representations of these complex variables using hand-crafted methods such as feature-engineering or using unsupervised learning (38) before training an algorithm. Factors that may confound or affect the outcome include laboratory parameters (intracytoplasmic sperm insemination), embryo quality, and patient diagnoses (age, prior treatment, ovarian response, endometrial quality) (39, 40), and therefore need to be addressed in AI approaches. For instance, an increase in maternal age was associated with faster early morphokinetics and slower morula and blastocyst formation (40). When computational resources are devoted to high quantity, complex data, the resultant complexity of many parameters can possibly negatively affect the accuracy of algorithms (41, 42). Additionally, this requires expensive, time-consuming multiprocessing cores, a resource to which few laboratories have access. To ameliorate the computational burden and simplify the selection of clinically relevant parameters tied to embryo quality, we hypothesized that a simple machine learning algorithm could be trained on characteristic embryo morphokinetic time points to predict pregnancy outcomes with high accuracy. Our secondary outcome was pregnancy rates in clusters generated by unsupervised machine learning as a proof of concept.

Material and methods

Study Criteria

A retrospective cohort study was approved by the Institutional Review Board (protocol no. H39094) at Baylor College of Medicine, Houston, Texas. Inclusion criteria were all autologous and donor cycle embryos that were incubated in the EmbryoScope time-lapse microscope (Vitrolife, Sweden) and were subsequently transferred, with in vitro culture of embryos to the blastocyst stage (day 5 to day 7), resulting in a clinical pregnancy (defined by fetal heartbeat after 6 weeks) or no clinical pregnancy (not excluding biochemical pregnancies) from 2014 to 2018 at the Texas Children’s Hospital Family Fertility Center. Exclusion criteria included embryo culture in other incubators, embryos without morphokinetic data or incomplete annotations, and dual embryo transfers with discordant outcomes. The sample size was chosen based on other studies that have found significant patterns in <200 embryos and validation cohorts <100 embryos (29). There were no ectopic gestations included in the data sets.

Ovarian Stimulation and Oocyte Retrieval

Patients included were women who underwent oocyte retrieval and subsequent embryo transfer at a single site academic fertility center. For each stimulation, an individualized antagonist ovarian stimulation protocol was employed using follicle-stimulating hormone follitropin alfa (Gonal F; EMD Serono, Rockland, MA) or (Follistim; Merck, and Co.), and follicle-stimulating hormone/luteinizing hormone menotropin (Menopur; Ferring Pharmaceuticals, NJ). Antagonist cetrorelix acetate (Cetrotide; Freedom Fertility Pharmacy, Byfield, MA) was initiated when a lead follicle reached 13 mm. Oocyte trigger was performed with 3–10,000 units of chorionic gonadotropin (Novarel, Ferring Pharmaceuticals) or leuprolide 36 hours before oocyte retrieval when two or more follicles measured greater than 18 mm and in conjunction with appropriate estradiol values. Follicular number and size were obtained via transvaginal ultrasound during ovarian stimulation monitoring, and serum samples were collected via peripheral venipuncture taken within 1 hour of the ultrasound measurement. Transvaginal oocyte aspiration was performed under total intravenous sedation 36 hours post-trigger. The demographic data collected included age and the patient’s medical diagnosis.

Fertilization and Embryo Incubation

Oocyte-cumulus complexes were cultured for 3 hours in 60 mm organ culture dishes containing Quinn’s Advantage Fertilization Media (Cooper Surgical, Trumbull, CT) with 10% serum protein supplement (Cooper Surgical) and overlaid with OVOIL (Vitrolife, Englewood, CO) in a Planer BT37 benchtop incubator (Cooper Surgical).

After exactly 3 hours, oocytes were denuded using hyaluronidase (Cooper Surgical) and assessed for nuclear maturity as evidenced by the presence of a polar body. Identified metaphase II oocytes underwent intracytoplasmic sperm injection (ICSI) or conventional insemination. After ICSI, the oocytes were immediately placed in the EmbryoScope time-lapse incubator (Vitrolife, Englewood, CO) and were cultured in Sage 1-Step media at 37 ºC in 5.5% CO2 and 6% O2. Embryos were assessed beginning at 16–18 hours after insemination using the EmbryoScope viewer. For conventional insemination, zygotes were placed in the EmbryoScope after fertilization. Embryo quality was graded on day 3, day 5, and day 6 using established morphological grading criteria (43, 44).

From April 2014 to July 2015, embryos were cultured in sequential media (Quinn’s Advantage Plus media and Quinn’s Advantage Blastocyst Plus media). From September 2015 onwards, embryos were cultured in Sage 1-Step media. Both culture conditions were compared, and no changes in blastocyst development were observed locally (45), similar to findings at other centers (46).

Embryos selected for preimplantation genetic testing (PGT) underwent assisted hatching by laser on day 3 and were biopsied on day 5 or 6. Embryos were vitrified with Vit-Freeze kits (Irvine Scientific, Santa Ana, CA) using the manufacturer’s protocol. Embryos for frozen transfers were warmed using Vit-Thaw Kit (Irvine Scientific) at least 3 hours before the transfer. After warming, assisted hatching was performed by laser (Hamilton-Thorne Beverly, MA) on embryos that had not been previously biopsied for PGT.

Time-Lapse Microscopy and Morphokinetic Parameters

The interval image acquisition for the EmbryoScope was every 10 minutes at seven focal planes; images were obtained starting from the time after fertilization until the embryo was either transferred or cryopreserved. All morphokinetic parameters starting from the fading of pronuclei to blastocyst formation were annotated manually using EmbryoViewer software. The morphokinetic parameters were determined by a team of trained personnel that included two physicians and one embryologist. Morphokinetic annotation, measurement consistency, and internal quality control were verified by the senior embryologist to decrease annotator bias. For ICSI, t0 was the time of insemination, and for conventional in vitro fertilization (IVF), t0 was defined as the time of addition of sperm to oocytes, similar to other reported studies (23, 47, 48). Parameter information collected included: pronuclei fade (PNf), time to 2-cell (t2), t3, t4, t5, t8, t9, time to morula (defined as 50% cells with indistinct membranes), time to start of early blastulation, time to blastocyst (Fig. 1).

Figure 1

Visual abstract of morphokinetic data gathering. Illustration of embryo development following fertilization and the morphokinetic parameters studied, including individual time points.

Embryo Transfer and Clinical Outcomes

All embryo transfers were performed under ultrasound guidance as fresh (n = 30) or frozen (n = 350) transfers. All frozen embryos were transferred on day 5 after progesterone administration of approximately 5 days. Luteal support was provided by vaginal progesterone gel (Crinone, Juniper Pharmaceuticals) and continued if pregnancy occurred. Embryo selection was based on a standard grading system of morphologic characteristics (43, 49) for both PGT for aneuploidy (PGT-A) and non-PGT-A tested embryos. Serum β human chorionic gonadotropin (hCG) levels were obtained 9–10 days after embryo transfer. Once a positive β-hCG was established, a repeat hCG test was performed 48–72 hours later, and then a transvaginal ultrasound was performed between 6 and 7 weeks of gestational age to determine the presence of clinical pregnancy as defined by a detectable fetal heartbeat.

Supervised Classification Training

Embryo morphokinetic data were randomly divided into 70% training sets and 30% test sets (Fig. 2). The ratio of data assignment to training and test was set by the operator (70%/30%), and this parameter was entered into a ranger. Embryo assignment was achieved with the use of a random module in python based on the Mersenne Twister function that assigns a pseudo-random number starting with an original number or seed, such as 1, that changes with each iteration for each embryo (50). Learning curves were developed using scikit-learn (51) to assess the optimal number of training samples, which was at least 150 samples. (Supplemental Fig. 1, available online). Total variables used included all morphokinetic time points. Data were trimmed manually by removing time points that had high collinearity with variable inflation factor >10 which resulted in the use of only six-time points (Supplemental Fig. 2A). Binary outcomes included either positive or negative pregnancy (Supplemental Fig. 2B). Models trained included logistic regression, XGBoost, decision tree, and random forest using Exploratory and ranger (52) (Fig. 2). For each model, the training set was used to develop an algorithm using the training variables and applied to the data on the test set.

Figure 2

Workflow diagram of supervised machine learning.

Model Comparison

For each model, the predicted probability of pregnancy was calculated for each embryo and compared with the actual outcome. A receiver operating characteristic curve was developed by graphing the true positive rate (sensitivity) with the false positive rate (1 − specificity). The area under the curve (AUC) was calculated for each receiver operating characteristic curve.

Unsupervised Clustering

All morphokinetic time points were normalized and used in K-means clustering (Supplemental Fig. 3A). To determine the number of clusters, the average sum of squares (distances between all points per cluster) was graphed against the number of clusters using the Elbow method (53) (Supplemental Fig. 3B). Thus, 5 clusters were selected. The Hartigan-Wong algorithm was used to group the data into 5 clusters with a random seed of 1 and a maximum iteration of 10. The normalized number of hours per cluster and the pregnancy rate per cluster were calculated.

Results

Out of 479 embryos surveyed, a total of 367 embryos met the inclusion criteria for the current study (Supplemental Fig. 4A). Embryos were transferred based on morphology. Embryo development from fertilization to blastocyst formation (Fig. 1) was manually annotated hours postinsemination. No differences were observed between the frequency of donor eggs, patient age, and diagnoses (Supplemental Fig. 4B). Multiple types of algorithms were tested and compared, including logistic regression, XGBoost, decision tree, and random forest (Supplemental Table 1), using six morphokinetic parameters as described in the Methods section. The best performing algorithm was random forest, which was named Yang-Peavey Embryo Enhancement Algorithm, with an AUC of 0.91 and 0.69 in the training and test sets, respectively (Fig. 3A). The calculated sensitivity was 65%, specificity 60%, positive predictive value 74%, and negative predictive value 50% from the validation data set only, not including the training set. To address whether PGT-A or single embryo transfer status affected the quality of the algorithm, we incorporated both statuses as covariates (Supplemental Fig. 5B and C, respectively). Similar AUC was achieved (0.92). However, lower predictive power was noted in the test sets as the sample size reached n = 4 in one of the arms (Supplemental Figs. 6 and 7). When the algorithm was applied to live birth rate data, we found similar predictive power with the AUC of 0.85 and 0.64 in the training and test sets, respectively (Fig. 3B).

Figure 3

Receiver operating characteristic (ROC) curve based on clinical pregnancy and live birth rate outcomes. Two ROC curves developed from the training and test sets based on morphokinetics in orange and blue, respectively, based on clinical pregnancy outcome (A). Two ROC curves developed from the training and test sets were based on morphokinetic parameters of time to pronuclei fade, t4, time to morula, t9, and time to blastocyst (B). The area under the curve (AUC) is noted next to the graph. The x-axis is the true positive rate. y-axis is the false positive rate (1 − specificity).

For our secondary analysis, we sought to understand the pattern of embryo growth in an unbiased manner by unsupervised machine learning. We asked whether the embryos could be grouped based on shared morphokinetic characteristics. Given the large range of the number of hours from 19 hours to 160 hours, the data were normalized or rescaled from −3 to 3. Data were grouped into 5 clusters by the distance between each time point (range, n = 21 to n = 97) (Fig. 4B). The normalized morphokinetic values were compared for each cluster showing high and low normalized values (Fig. 4A). Notably, cluster 4 had higher normalized values or slower time points, and cluster 5 had lower values or faster time points (Fig. 4A). The differences in pregnancy rates among the 5 clusters were not statistically significant (χ2 test P=.356). When we compared the pregnancy rates among the clusters, cluster 4 (71%) and cluster 2 (60%) had a trend of higher rates of pregnancy, whereas clusters 1 (51%) and cluster 3 (51%) had lower rates (Fig. 4B). In this patient population, a small group of embryos (cluster 4, n = 25) with high rates of pregnancy (71%) developed overall more slowly, whereas another cohort of embryos (cluster 2, n = 78) with faster compaction and blastulation had a 60% pregnancy rate (Fig. 4B and Supplemental Table 2) and may represent intrinsic embryo qualities for future hypothesis testing. Overall, cluster 5 represented a group of embryos with faster overall morphokinetics, and cluster 4 represented those with slower morphokinetics (Supplemental Fig. 8).

Figure 4

Unsupervised clustering characteristics. Distribution of values of each normalized morphokinetic parameter grouped by cluster (A). Box plot includes the first and third quartile with minimum and maximum. The center of the box plot presents the median. The x-axis is a cluster. The y-axis is the normalized value. The key for each morphokinetic time point is listed in the figure legend—frequency of embryos with positive outcomes in each cluster (B). Blue indicates no pregnancy. Orange indicates pregnancy.

Discussion

We show that simple machine learning can be used to develop an algorithm to predict clinical pregnancy rates with as few as 200 embryo outcomes with AUCs of 0.91 and 0.69, respectively, in the training and test cohorts. The algorithm also predicted live birth rate outcomes. When PGT-A and single embryo transfer statuses of each embryo were incorporated, the predictive power of the training set declined, likely due to the overfitting of the data because of the small samples size ranging from 4 to 50 patients per arm (Supplemental Figs. 6 and 7). Embryos that had faster blastulation and compaction and a select few with slower development were also associated with favorable pregnancy outcomes. A small, distinct group of embryos with slower development was noted to have higher pregnancy rates, although it was not statistically significant. This is in contrast to other studies that have reported faster development associated with more favorable outcomes (19). Most likely, there was sample size bias, or there was likely an embryo-related biological bias associated with this small group of embryos, or other factors such as uterine or environmental factors unrelated to the embryo.

This algorithm is unique in that it distills innumerable clinical, laboratory, and genetic factors into a few quintessential features to predict pregnancy, assuming that these characteristics are represented by individualized embryo morphokinetics. Many studies use over 32 demographic, laboratory, or grading criteria in the algorithms (1, 30, 47) or complex imaging processing (35, 36), which increases model learning difficulty requiring complex calculations (41). This algorithm can be run on any platform on-site at the laboratory and can be quickly adapted or retrained to the changes in laboratory conditions such as media with input from only 200 embryos. This technology is also noninvasive and may aid in embryo selection after PGT-A. Further tests are needed to determine whether the Yang-Peavey Embryo Enhancement Algorithm can predict pregnancy rates in embryos that did or did not undergo PGT-A.

Whereas previous studies emphasized single parameters or ratios of time points as the best predictors of blastulation (8, 54) or implantation (10, 17) or that parameters cannot be used to predict blastulation (55), our data demonstrate that a prediction model is best developed from multiple morphokinetic time points (56) and can successfully predict pregnancy. Reassuringly, our model is also consistent with other reports of time to morula and time to blastocyst time points being predictive of implantation (17, 19, 20, 57). Further work is needed to explore whether a few key morphokinetic parameters can universally predict implantation using the built-in EmbryoViewer software (58, 59).

A limitation of this study is that all transferred embryos underwent morphologic grading, which influenced embryo selection. Our findings were made after standard morphologic grading. Future randomized studies are needed to compare predictions between the algorithm and morphological grading. Other limitations include multiple transfers per patient and donor embryos that could be affected by maternal factors. Another limitation is that laboratory parameters included many varying treatments, including ICSI and conventional insemination, PGT-A or untested, assisted hatching, fresh and frozen embryo transfer cycles, and fresh and frozen oocytes.

Time of insemination (t = 0) in conventional IVF may affect or skew the embryo morphokinetics compared with that of ICSI. Some studies reported up to a 1.5-hour delay in early morphokinetics and up to 4.1 hours of advancement in development to the blastocyst stage in conventional IVF compared with ICSI (60). Because of conventional insemination, only time points after time to PNf (48) were included in our study. In our data, a 1.5-hour difference is within the SD (1.7 to 3.5 hours) at the time to PNf, and a 4.1-hour advancement is also within the SD at the blastocyst stage (3.6 to 6.5 hours) observed among the morphokinetic groups by unsupervised clustering (Supplemental Table 2). However, more validation studies are needed to determine whether t0 in conventional IVF significantly impacts morphokinetics from that of ICSI. For instance, to control for the exact time of t0, one would need to study embryos that underwent ICSI only for future larger cohorts as part of a continuously updated, dynamic algorithm.

Assisted hatching was performed for all samples undergoing PGT testing and could have been accounted for by studying PGT status. In the analysis, including PGT-A status, we did not observe an improvement in predictive power because of low sample numbers. The addition of confounding factors likely depletes the accuracy of the model given the low representative sample numbers. Therefore, it would be useful to analyze the effect of assisted hatching with a larger data set. In addition, sex selection based on PGT-A results could confound results. However, in our cohort, the few patients who requested sex selection had embryos that were transferred based on morphological grading. These factors could be included in the algorithm for future predictions with larger cohorts or applied to a cohort with similar characteristics as the current cohort. If conditions change, the algorithm would need to be updated to be applied prospectively. One of the challenges of applying this algorithm is the need for manual annotation for each embryo prospectively which requires additional time by a trained embryologist or technical staff. Unfortunately, at this facility, there was no access to software that performed automated annotations. In the future, the algorithm could be combined with automated annotation software to improve efficiency.

Another limitation is a lack of radiological examination to confirm zygosity to distinguish the difference between two embryo transfers that resulted in monozygotic duplication and loss of the other embryo compared with true implantation of both embryos.

Conclusion

We demonstrate that embryo morphokinetics are associated with pregnancy outcomes in a dynamic machine learning algorithm that is specific to a clinic. We also showed that embryos that exhibit slow overall morphokinetics (time to blastocyst = 122 hours vs. 97 hours) had higher rates of pregnancy (71% vs. 54%). As a proof of concept, machine learning can be used at a local IVF center to design a selection algorithm with as few as 200 embryos. As the sample size and the parameters and characteristics change with time, a new model will need to be continually optimized. Rather than using a single universal model, we demonstrate that a model tailored to the local embryo and laboratory characteristics with continuous updates could be used.