Machine Learning-Guided Noninvasive Embryo Selection for Clinical in Vitro Fertilization Treatment to Avoid Wasting Potentially Qualified Embryos

BackgroundThe success rates of in vitro fertilization (IVF) treatment are limited by the aneuploidy of human embryos. Pre-implantation genetic testing for aneuploidy(PGT-A) is often used to select embryos with normal ploidy but requires invasive embryo biopsy. MethodsWe performed chromosome sequencing of 345 paired blastocyst culture medium and whole blastocyst samples and developed a noninvasive embryo grading system based on the random forest machine-learning algorithm to predict blastocyst ploidy. The system was validated in 266 patients, and a blinded prospective observational study was performed to investigate clinical outcomes between machine learning-guided and traditional niPGT-A analyses. We graded embryos as A, B, or C using machine learning-guided niPGT-A analysis according to their euploidy probability levels predicted by noninvasive chromosomal screening. ResultsWe observed higher live birth rate in A- versus C-grade embryos (50.4% versus 27.1%, p=0.006) and B- versus C-grade embryos (45.3% versus 27.1%, p=0.022) and lower miscarriage rate in A- versus C-grade embryos (15.9% versus 33.3%, p=0.026) and B- versus C-grade embryos (14.3% versus 33.3%, p=0.021). The embryo utilization rate was significantly higher through machine learning strategy compared to the conventional dichotomic judgment of euploidy or aneuploidy in the niPGT-A analysis (78.8% versus 57.9%, p<0.001). We observed better outcomes in A- and B-grade embryos versus C-grade embryos and higher embryo utilization rates through machine learning strategies than traditional niPGT-A analysis. ConclusionThese results demonstrate that the machine learning-guided embryo grading system can optimize embryo selection and avoid wasting potential embryos.Trial registrationChinese Clinical Trial Registry,ChiCTR-RRC-17010396.Registered 11 January 2017, http://www.chictr.org.cn/ChiCTR-RRC-17010396

P–752 Embryo morphological grading across several IVF centers is not consistent but an interactive training is useful to improve its consistency

Study question Are the embryologists across several IVF clinics concordant when evaluating embryo morphology? Summary answer Embryo morphological grading is sufficiently consistent among embryologists from the same center, while an interactive training was essential to improve its accuracy across several clinics. What is known already Embryo morphology, mostly at the blastocyst stage, is the strongest non-invasive embryological feature that associates with implantation potential. This association is confirmed also when euploid blastocysts are transferred. At present, several embryo grading schemes exist but is still unclear which is the most effective among them. Moreover, many IVF clinics adopt internal embryo grading scores, further limiting the transferability of this crucial prognostic information across different laboratories. With the aim of assessing the level of concordance in embryo grading within and between IVF clinics, the Italian Society of Embryology, Reproduction and Research (SIERR) conceived this study. Study design, size, duration We photographed 40 cleavage-stage and 40 blastocyst-stage embryos (3 focal-planes=240 photos). Two embryologists (senior and junior) from 65 Italian IVF clinics were invited to grade them. Their evaluations were blindly collected as Phase-I (January2020-March2020). Phase-II consisted of an interactive-training on Google-Classroom during which 6 selected experts found a Consensus on the morphological evaluation of the 80 embryos (April2020). As Phase III (May2020-July2020), a second set of 240 pictures was sent to senior participants and experts. Participants/materials, setting, methods Eighteen centers agreed to participate, and 36 embryologists were included. The embryo grading scheme adopted was the Alpha-ESHRE Istanbul Consensus (parameters: cleavage-stage blastomeres’ symmetry and fragmentation, blastocyst’s expansion, inner-cell-mass and trophectoderm quality), conventionally used in 50% of the centers (N = 9/18). The concordance within (junior versus senior) and between (senior versus experts) centers was calculated through the Cohen’s-k. The concordance between centers was compared before and after the interactive training on the two sets of pictures. Main results and the role of chance The centers and embryologists included were representative of the Italian IVF scenario: oocyte-retrievals per year:711±636,range100–2200; cycles with cleavage-stage embryo-transfer:322±339,0–1300; cycles with blastocyst-stage embryo-transfer:390±403,0–1100; operators per center:5.6±4.0,2–13; senior embryologists’ experience:14.8±7.4yr,7–30; junior embryologists’ experience:2.7±0.6yr,1–3. The intra-center concordance was (i)for blastomeres’ symmetry 82±15% (38–100%), k 0.59±0.27 (0.02–1), (ii)for blastomeres’ fragmentation 88±9% (65–100%), k 0.71±0.2 (0.29–1), (iii)for blastocysts’ expansion 80±16% (48–100%), k 0.66±0.27 (0.19–1), (iv)for inner-cell-mass quality 73±16% (35–95%), k 0.58±0.24 (0.07–0.92), (v)for trophectoderm quality 71±19% (38–95%), k 0.54±0.32 (0.01–0.97). Linear regressions showed no association of centers’ and embryologists’ characteristics with all k-values. Among clinics with the highest mean number of cycles per year and intra-center concordance, we selected 6 experts for the interactive-training. We then calculated the inter-center concordance as the agreement rate between senior embryologists and the experts for phase-I and phase-III: (i)for blastomeres’ symmetry 67±15% (30–85%) and 73±17% (15–90%;Wilcoxon-signed-ranks-test=0.06), k 0.33±0.22 (–0.29–0.58) and 0.42±0.33 (–0.56–0.77); (ii)for blastomeres’ fragmentation 81±17% (23–95%) and 83±14% (50–95%;Wilcoxon-signed-ranks-test=0.8), k 0.54±0.22 (–0.05–0.84) and 0.55±0.22 (0.17–0.81); (iii)for blastocysts’ expansion 59±16% (35–85%) and 67±17% (23–90%;Wilcoxon-signed-ranks-test=0.04), k 0.35±0.20 (0.06–0.73) and 0.44±0.22 (–0.10–0.7); (iv)for inner-cell-mass quality 60±14% (33–80%) and 69±11% (48–85%;Wilcoxon-signed-ranks-test=0.02), k 0.40±0.20 (0.01–0.69) and 0.51±0.18 (0.18–0.77); (v)for trophectoderm quality 55±12% (23–70%) and 63±10% 48–78%;Wilcoxon-signed-ranks-test&lt;0.01), k 0.29±0.15 (–0.08–0.52) and 0.42±0.15 (0.21–0.66). Limitations, reasons for caution Only 28% (N = 18/65) of the Italian IVF centers invited to participate responded to the survey. The conventional adoption of grading schemes other than Istanbul-Consensus by 50% of the embryologists might have biased their evaluation. The experts were not fully-concordant in grading 13.8% of the embryos (N = 22/160), which required active discussions. Wider implications of the findings: Blastocyst-grading concordance was significantly improved after the training phase. Therefore, interactive consensus meetings and training platforms are keenly needed to standardize this practice across the centers. The “avant-garde” of artificial intelligence applied to embryo image analysis might help overcoming this issue in the future. Trial registration number N.A.

P–267 Characterising a novel embryo grading system for 4-cell embryos including symmetry, fragmentation, cell configuration, cell contacts per cell, distance between cells, and cell adhesion strength

Study question What other cell junction factors can be easily characterised from imagery of 4-cell embryos to assist in embryo classification and prediction of viability? Summary answer 4-cell embryo grading should not only account for symmetry and fragmentation, but also cell configuration and cell adhesion quality. What is known already 4-cell embryos are clinically classified according to cell number, symmetry and fragmentation without accounting for cell orientation or quality of cell junctions. Our previous work has focused on classification of 4-cell embryos according to overall embryo shape (“tetrahedral” versus “planar”) using artificial intelligence. Our work, as confirmed by others, has demonstrated that embryo shape at the 4-cell stage is an important determinant of the ability of the embryo to reach blastocyst, utilisation, pregnancy and live birth. This is thought to be because of variations in intracellular communication between cells in embryos with different orientations, and consequently, different intracellular junction phenotypes. Study design, size, duration Using geometrical principles, possible permutations of 4-cell embryos (excluding redundant mirrored permutations) were identified and further classified based on shape and number of cell junctions. For ease of calculations, cells were assumed to be spherical, with at least one intracellular cell contact and symmetrical in size with other cells in the same embryo. Participants/materials, setting, methods The six distances between centroids of permutations for each configuration were calculated relative to the size of the cell, and the shortest distance between cell membranes. Adhesion was characterised from embryo imagery based on the overall shape of the cell, external angle between cells and the length of cell contact (the more spherical the cell, the larger the angle and the longer the cell contact point, the stronger the adhesion, adapted from Winklbauer,2015). Main results and the role of chance 4-cell embryos may be classified into 13 variant configurations: 1 typical Tetrahedral, 2 quasi-tetrahedral, 10 planar. These variants were classified according to number of cells with 0,1,2 and 3 intracellular contacts, leading to six possible configurations: 0004(tetrahedral), 0022(quasi tetrahedral/planar), 0040 (planar), 0121(quasi tetrahedral/planar), 0301(planar), 0220(planar). The number of total cell junctions in the embryo in each of these configurations was 12,10,8,8,6,6 respectively, with tetrahedral embryos (0004) having twice the cell contacts compared to planar embryos (p &lt; 0.001). Tetrahedral embryos have an advantage over the other embryo configurations in terms of better embryo communication, as demonstrated by the lower average and variation in distance and shorter sum of all intracellular distances between centroids (mean: 0.78 vs 0.94,0.98,1.04,1.09,1.19; stdev:0.06 vs 0.2,0.3,0.3,0.4,0.4,0.3; sum:4.7 vs 5.2,6.3,6.1,6.7,7.0,6.0 cell lengths) and between cells (mean: 0 vs 0.34,0.07,0.39,0.42,0.56; stdev: 0 vs 0.56,0.11,0.51,0.53,0.66; sum: 0 vs 2.71,0.43,3.13,3.38,4.46 cell lengths) observed in tetrahedral embryos versus other five configurations respectively (p &lt; 0.001). Cell junctions were classified according to degree of cell adhesion: A:none (cells remain spherical in shape); B:weak (external angle between the cells is acute, there is a narrow visible cell junction); C:strong (external angle between the cells is obtuse with a wide visible cell junction). Limitations, reasons for caution: Follow-up studies will evaluate the impact of different cell shapes,cells without intracellular contact,and asymmetrical embryos. The proposed classification will be validated against a database of known outcome from 8 clinics from 6 countries to quantify the clinical implications of this classification,and the consistency of assessment by humans and AI . Wider implications of the findings: It is clear that differences in intracellular communication between cells in embryos with different orientations, and different intracellular junction phenotypes is an important determinant of embryo viability. Our classification system allows for an easy to use and mathematically sound criteria for classifying 4-cell embryo cell junction quality. Trial registration number NA