A hybrid artificial intelligence model leverages multi-centric clinical data to improve fetal heart rate pregnancy prediction across time-lapse systems

STUDY QUESTION

Can artificial intelligence (AI) algorithms developed to assist embryologists in evaluating embryo morphokinetics be enriched with multi-centric clinical data to better predict clinical pregnancy outcome?

SUMMARY ANSWER

Training algorithms on multi-centric clinical data significantly increased AUC compared to algorithms that only analyzed the time-lapse system (TLS) videos.

WHAT IS KNOWN ALREADY

Several AI-based algorithms have been developed to predict pregnancy, most of them based only on analysis of the time-lapse recording of embryo development. It remains unclear, however, whether considering numerous clinical features can improve the predictive performances of time-lapse based embryo evaluation.

STUDY DESIGN, SIZE, DURATION

A dataset of 9986 embryos (95.60% known clinical pregnancy outcome, 32.47% frozen transfers) from 5226 patients from 14 European fertility centers (in two countries) recorded with three different TLS was used to train and validate the algorithms. A total of 31 clinical factors were collected. A separate test set (447 videos) was used to compare performances between embryologists and the algorithm.

PARTICIPANTS/MATERIALS, SETTING, METHODS

Clinical pregnancy (defined as a pregnancy leading to a fetal heartbeat) outcome was first predicted using a 3D convolutional neural network that analyzed videos of the embryonic development up to 2 or 3 days of development (33% of the database) or up to 5 or 6 days of development (67% of the database). The output video score was then fed as input alongside clinical features to a gradient boosting algorithm that generated a second score corresponding to the hybrid model. AUC was computed across 7-fold of the validation dataset for both models. These predictions were compared to those of 13 senior embryologists made on the test dataset.

MAIN RESULTS AND THE ROLE OF CHANCE

The average AUC of the hybrid model across all 7-fold was significantly higher than that of the video model (0.727 versus 0.684, respectively, P = 0.015; Wilcoxon test). A SHapley Additive exPlanations (SHAP) analysis of the hybrid model showed that the six first most important features to predict pregnancy were morphokinetics of the embryo (video score), oocyte age, total gonadotrophin dose intake, number of embryos generated, number of oocytes retrieved, and endometrium thickness. The hybrid model was shown to be superior to embryologists with respect to different metrics, including the balanced accuracy (P ≤ 0.003; Wilcoxon test). The likelihood of pregnancy was linearly linked to the hybrid score, with increasing odds ratio (maximum P-value = 0.001), demonstrating the ranking capacity of the model. Training individual hybrid models did not improve predictive performance. A clinic hold-out experiment was conducted and resulted in AUCs ranging between 0.63 and 0.73. Performance of the hybrid model did not vary between TLS or between subgroups of embryos transferred at different days of embryonic development. The hybrid model did fare better for patients older than 35 years (P < 0.001; Mann-Whitney test), and for fresh transfers (P < 0.001; Mann-Whitney test).

LIMITATIONS, REASONS FOR CAUTION

Participant centers were located in two countries, thus limiting the generalization of our conclusion to wider subpopulations of patients. Not all clinical features were available for all embryos, thus limiting the performances of the hybrid model in some instances.

WIDER IMPLICATIONS OF THE FINDINGS

Our study suggests that considering clinical data improves pregnancy predictive performances and that there is no need to retrain algorithms at the clinic level unless they follow strikingly different practices. This study characterizes a versatile AI algorithm with similar performance on different time-lapse microscopes and on embryos transferred at different development stages. It can also help with patients of different ages and protocols used but with varying performances, presumably because the task of predicting fetal heartbeat becomes more or less hard depending on the clinical context. This AI model can be made widely available and can help embryologists in a wide range of clinical scenarios to standardize their practices.

STUDY FUNDING/COMPETING INTEREST(S)

Funding for the study was provided by ImVitro with grant funding received in part from BPIFrance (Bourse French Tech Emergence (DOS0106572/00), Paris Innovation Amorçage (DOS0132841/00), and Aide au Développement DeepTech (DOS0152872/00)). A.B.-C. is a co-owner of, and holds stocks in, ImVitro SAS. A.B.-C. and F.D.M. hold a patent for 'Devices and processes for machine learning prediction of in vitro fertilization' (EP20305914.2). A.D., N.D., M.M.F., and F.D.M. are or have been employees of ImVitro and have been granted stock options. X.P.-V. has been paid as a consultant to ImVitro and has been granted stocks options of ImVitro. L.C.-D. and C.G.-S. have undertaken paid consultancy for ImVitro SAS. The remaining authors have no conflicts to declare.

TRIAL REGISTRATION NUMBER

N/A.

O-124 A new artificial intelligence (AI) system in the block: impact of clinical data on embryo selection using four different time-lapse incubators

Study question

Can AI algorithms assist embryologists in evaluating embryos from any time-lapse system (TLS) along with clinical data to better predict pregnancy outcomes and reduce time-to-pregnancy?

Summary answer

Our algorithm (Embryoly) significantly increases accuracy in predicting clinical pregnancy by 26.9% amongst embryos deemed of fair and good quality when clinical data is included.

What is known already

Embryologists routinely use defined morpho-kinetic criteria to decide which embryo to transfer, and yet, many embryos deemed of good quality fail to lead to a pregnancy. Thus, AI algorithms to assist embryologists in objectively selecting the most promising embryos are in demand. To date, several reports indicate that AI algorithms are capable of predicting pregnancy clinical outcomes but to the best of our knowledge they only consider visual data (or together with a small set of clinical features) from individual TLI systems to generate their predictions. 

Study design, size, duration

A dataset of 6790 embryos (97.82% known clinical pregnancy outcome, 31.47% frozen transfers) from 2519 patients from 11 European fertility centers recorded with 4 different TLS (GERI-Merck, Embryoscope & EmbryoscopePlus-Vitrolife and MIRI-Esco) was used to train and validate Embryoly. Nine out of 93 clinical factors were identified as being the most predictive, including woman age, woman and man BMI and AMH levels. Performances were evaluated on a separate test dataset (393 videos).

Participants/materials, setting, methods

Clinical pregnancy outcome was predicted using a 3D convolutional neural network that analyzed up to 5 days of embryo development. The output score was further analyzed considering the clinical features to generate a second clinical score. Both predictions were compared to those of 10 senior embryologists made on the same test dataset (with and without clinical features). Embryo quality was assessed as: poor, fair, good. Unless specified otherwise, McNemar test was used for statistical tests.

Main results and the role of chance

Overall accuracy of embryologists in predicting clinical pregnancy  based on videos alone was 57.25% (CI 95% : 52.34% - 62.16%) compared to 60.56% (CI 95% : 55.71% - 65.41%) for Embryoly (p = 0.35).

When videos were analyzed together with the clinical factors, overall accuracy of embryologists was significantly lower than Embryoly (60.05% [CI 95% : 55.19% - 64.91%] vs 68.19% [CI 95% : 63.57% - 72.82%], p-value=0.015, respectively). Clinical factors significantly increased our accuracy by 7.63% (p-value=0.030). More specifically, Embryoly algorithms fared better in terms of detecting false positives (31.30% vs 19.34%) compared to embryologists, with a specificity of 74.4% vs. 58.6%, respectively.

If we consider only embryos of fair and good quality (71.50% of our test dataset) Embryoly’s accuracy was 13.52% higher than that of embryologists. This translates into AI having an even better ability to detect false positives for embryos that could be seen as good candidates for transfer (20.28% false positives against 42.70% for the embryologists). Embryoly performs differently across selected TLS when analyzing videos alone, but not when clinical data was also considered (chi2 test, p &lt; 0.001 and 0.5, respectively). Further work will investigate these discrepancies across TLS.

Limitations, reasons for caution

As of today, Embryoly’s accuracy in predicting the outcome of poor-quality embryos is not different to that of embryologists (79.46% vs 84.96%; p-value=0.19). We are improving this by exposing Embryoly to more “poor quality” embryos, so as to also identify poor quality embryos with unexpected potential for implantation.

Wider implications of the findings

Our pioneering findings support the use of AI for a standardized and couple-centered care in clinical embryology, integrating male and female factors with embryo development analyses from multiple TLS. Our approach has the potential to cost-effectively reduce time to pregnancy and is another step toward a personalized embryo transfer strategy.

Trial registration number

Not applicable

Comparative study of preimplantation development following distinct assisted oocyte activation protocols in a PLC-zeta knockout mouse model

Mammalian fertilization encompasses a series of Ca2+ oscillations initiated by the sperm factor phospholipase C zeta (PLCζ). Some studies have shown that altering the Ca2+ oscillatory regime at fertilization affects preimplantation blastocyst development. However, assisted oocyte activation (AOA) protocols can induce oocyte activation in a manner that diverges profoundly from the physiological Ca2+ profiling. In our study, we used the newly developed PLCζ-null sperm to investigate the independent effect of AOA on mouse preimplantation embryogenesis. Based on previous findings, we hypothesized that AOA protocols with Ca2+ oscillatory responses might improve blastocyst formation rates and differing Ca2+ profiles might alter blastocyst transcriptomes. A total of 326 MII B6D2F1-oocytes were used to describe Ca2+ profiles and to compare embryonic development and individual blastocyst transcriptomes between four control conditions: C1 (in-vivo fertilization), C2 (ICSI control sperm), C3 (parthenogenesis) and C4 (ICSI-PLCζ-KO sperm) and four AOA groups: AOA1 (human recombinant PLCζ), AOA2 (Sr2+), AOA3 (ionomycin) and AOA4 (TPEN). All groups revealed remarkable variations in their Ca2+ profiles; however, oocyte activation rates were comparable between the controls (91.1% ± 13.8%) and AOA (86.9% ± 11.1%) groups. AOA methods which enable Ca2+ oscillatory responses (AOA1: 41% and AOA2: 75%) or single Ca2+ transients (AOA3: 50%) showed no significantly different blastocyst rates compared to ICSI control group (C2: 70%). In contrast, we observed a significant decrease in compaction (53% vs. 83%) and blastocyst rates (41% vs. 70%) in the absence of an initial Ca2+ trigger (AOA4) compared with the C2 group. Transcription profiles did not identify significant differences in gene expression levels between the ICSI control group (C2) and the four AOA groups.

Comparative analysis of different nuclear transfer techniques to prevent the transmission of mitochondrial DNA variants

Prevention of mitochondrial DNA (mtDNA) diseases may currently be possible using germline nuclear transfer (NT). However, scientific evidence to compare efficiency of different NT techniques to overcome mtDNA diseases is lacking. Here, we performed four types of NT, including first or second polar body transfer (PB1/2T), maternal spindle transfer (ST) and pronuclear transfer (PNT), using NZB/OlaHsd and B6D2F1 mouse models. Embryo development was assessed following NT, and mtDNA carry-over levels were measured by next generation sequencing (NGS). Moreover, we explored two novel protocols (PB2T-a and PB2T-b) to optimize PB2T using mouse and human oocytes. Chromosomal profiles of NT-generated blastocysts were evaluated using NGS. In mouse, our findings reveal that only PB2T-b successfully leads to blastocysts. There were comparable blastocyst rates among PB1T, PB2T-b, ST and PNT embryos. Furthermore, PB1T and PB2T-b had lower mtDNA carry-over levels than ST and PNT. After extrapolation of novel PB2T-b to human in vitro matured (IVM) oocytes and in vivo matured oocytes with smooth endoplasmic reticulum aggregate (SERa) oocytes, the reconstituted embryos successfully developed to blastocysts at a comparable rate to ICSI controls. PB2T-b embryos generated from IVM oocytes showed a similar euploidy rate to ICSI controls. Nevertheless, our mouse model with non-mutated mtDNAs is different from a mixture of pathogenic and non-pathogenic mtDNAs in a human scenario. Novel PB2T-b requires further optimization to improve blastocyst rates in human. Although more work is required to elucidate efficiency and safety of NT, our study suggests that PBT may have the potential to prevent mtDNA disease transmission.

Platelet-activating factor acetylhydrolase 1B3 (PAFAH1B3) is required for the formation of the meiotic spindle during in vitro oocyte maturation

Platelet-activating factor (PAF) is a well-described autocrine growth factor involved in several reproductive processes and is tightly regulated by its hydrolysing enzyme, PAF acetylhydrolase 1B (PAFAH1B). This intracellular enzyme consists of three subunits: one regulatory, 1B1, and two catalytic, 1B2 and 1B3. PAFAH1B3 has remained uncharacterised until now. Here, we report that PAFAH1B3 is present during the different stages of the first meiotic division in bovine, murine and human oocytes. In these species, the PAFAH1B3 subunit was clearly present in the germinal vesicle, while at metaphase I and II, it localised primarily at the meiotic spindle structure. In cattle, manipulation of the microtubules of the spindle by nocodazole, taxol or cryopreservation revealed a close association with PAFAH1B3. On the other hand, disruption of the enzyme activity either by P11, a selective inhibitor of PAFAH1B3, or by PAFAH1B3 antibody microinjection, caused arrest at the MI stage with defective spindle morphology and consequent failure of first polar body extrusion. In conclusion, our results show that one of the catalytic subunits of PAFAH1B, namely PAFAH1B3, is present in bovine, murine and human oocytes and that it plays a functional role in spindle formation and meiotic progression during bovine oocyte maturation.

Human oocyte calcium analysis predicts the response to assisted oocyte activation in patients experiencing fertilization failure after ICSI

STUDY QUESTION

Can human oocyte calcium analysis predict fertilization success after assisted oocyte activation (AOA) in patients experiencing fertilization failure after ICSI?

SUMMARY ANSWER

ICSI-AOA restores the fertilization rate only in patients displaying abnormal Ca2+ oscillations during human oocyte activation.

WHAT IS KNOWN ALREADY

Patients capable of activating mouse oocytes and who showed abnormal Ca2+ profiles after mouse oocyte Ca2+ analysis (M-OCA), have variable responses to ICSI-AOA. It remains unsettled whether human oocyte Ca2+ analysis (H-OCA) would yield an improved accuracy to predict fertilization success after ICSI-AOA.

STUDY DESIGN, SIZE, DURATION

Sperm activation potential was first evaluated by MOAT. Subsequently, Ca2+ oscillatory patterns were determined with sperm from patients showing moderate to normal activation potential based on the capacity of human sperm to generate Ca2+ responses upon microinjection in mouse and human oocytes. Altogether, this study includes a total of 255 mouse and 122 human oocytes. M-OCA was performed with 16 different sperm samples before undergoing ICSI-AOA treatment. H-OCA was performed for 11 patients who finally underwent ICSI-AOA treatment. The diagnostic accuracy to predict fertilization success was calculated based on the response to ICSI-AOA.

PARTICIPANTS/MATERIALS, SETTING, METHODS

Patients experiencing low or total failed fertilization after conventional ICSI were included in the study. All participants showed moderate to high rates of activation after MOAT. Metaphase II (MII) oocytes from B6D2F1 mice were used for M-OCA. Control fertile sperm samples were used to obtain a reference Ca2+ oscillation profile elicited in human oocytes. Donated human oocytes, non-suitable for IVF treatments, were collected and vitrified at MII stage for further analysis by H-OCA.

MAIN RESULTS AND THE ROLE OF CHANCE

M-OCA and H-OCA predicted the response to ICSI-AOA in 8 out of 11 (73%) patients. Compared to M-OCA, H-OCA detected the presence of sperm activation deficiencies with greater sensitivity (75 vs 100%, respectively). ICSI-AOA never showed benefit to overcome fertilization failure in patients showing normal capacity to generate Ca2+ oscillations in H-OCA and was likely to be beneficial in cases displaying abnormal H-OCA Ca2+ oscillations patterns.

LIMITATIONS, REASONS FOR CAUTION

The scarce availability of human oocytes donated for research purposes is a limiting factor to perform H-OCA. Ca2+ imaging requires specific equipment to monitor fluorescence changes over time.

WIDER IMPLICATIONS OF THE FINDINGS

H-OCA is a sensitive test to diagnose gamete-linked fertilization failure. H-OCA allows treatment counseling for couples experiencing ICSI failures to either undergo ICSI-AOA or to participate in gamete donation programs. The present data provide an important template of the Ca2+ signature observed during human fertilization in cases with normal, low and failed fertilization after conventional ICSI.

STUDY FUNDING/COMPETING INTEREST(S)

This work was supported by the Flemish fund for scientific research (FWO-Vlaanderen, G060615N). The authors have no conflict of interest to declare.