Highlights
- •The combination of morphology and proteomic profile is a powerful tool for selecting embryos.
- •IL-6 and MMP-1 concentrations in culture medium are helpful to identify successful embryos.
- •Artificial neural networks can detect the best embryos from a euploid cohort.
- •Artificial intelligence is effective in predicting live birth from IVF treatments.
ABSTRACT
Research question
The study aimed to develop an artificial intelligence model based on artificial neural networks (ANNs) to predict the likelihood of achieving a live birth using the proteomic profile of spent culture media and blastocyst morphology.
Design
This retrospective cohort study included 212 patients who underwent single blastocyst transfer at IVI Valencia. A single image of each of 186 embryos was studied, and the protein profile was analysed in 81 samples of spent embryo culture medium from patients included in the preimplantation genetic testing programme. The information extracted from the analyses was used as input data for the ANN. The multilayer perceptron and the back-propagation learning method were used to train the ANN. Finally, predictive power was measured using the area under the curve (AUC) of the receiver operating characteristic curve.
Results
Three ANN architectures classified most of the embryos correctly as leading (LB+) or not leading (LB–) to a live birth: 100.0% for ANN1 (morphological variables and two proteins), 85.7% for ANN2 (morphological variables and seven proteins), and 83.3% for ANN3 (morphological variables and 25 proteins). The artificial intelligence model using information extracted from blastocyst image analysis and concentrations of interleukin-6 and matrix metalloproteinase-1 was able to predict live birth with an AUC of 1.0.
Conclusions
The model proposed in this preliminary report may provide a promising tool to select the embryo most likely to lead to a live birth in a euploid cohort. The accuracy of prediction demonstrated by this software may improve the efficacy of an assisted reproduction treatment by reducing the number of transfers per patient. Prospective studies are, however, needed.
Graphical abstract
Graphical Abstract
Key words
Introduction
The two main factors responsible for the success of an IVF treatment are the endometrium and the embryo (
Edwards et al., 1984
). Non-invasive methods (morphological and morphokinetic) as well as invasive methods (genetic testing) are currently used in IVF laboratories in embryo selection. However, new approaches to select embryos are still being investigated due to the limited improvement in live birth rate over the last few years (Dyer et al., 2016
; De Geyter et al., 2018
).New non-invasive methods based on “-omic” sciences, such as metabolomics and proteomics, have emerged to assess embryo viability and improve clinical outcomes (
Krisher et al., 2015
). There is evidence for the importance of soluble ligands and receptors in both the embryo and the female reproductive tract during the embryonic preimplantation stage (Thouas et al., 2015
). The group of proteins secreted or metabolized by the embryo is known as the secretome (Hathout, 2007
); by analysing the culture medium, this is a promising source of information about the protein and metabolic state of the embryo (Hollywood et al., 2006
).Thirty years ago, platelet-activating factor in spent embryo culture media was reported as the first evidence of the embryo secretome (
Punjabi et al., 1990
). Since then, several proteins have been identified during the preimplantation stage, for example IFN-α2 (interferon-α2; Jones et al., 1992
), SP1 (pregnancy-specific P-l glycoprotein; Saith et al., 1994
), HLA-G (human leukocyte antigen G ; Noci et al., 2005
), Apo-A1 (apolipoprotein A1; Mains et al., 2011
), HCG (human chorionic gonadotrophin; Butler et al., 2013
) and GM-CSF (granulocyte-macrophage colony-stimulating factor; Ziebe et al., 2013
). Methods for protein analysis have improved in terms of sensitivity, leading to the simultaneous study of multiple proteins (e.g. using mass spectrometry techniques or microarray technology). This has made it possible to observe associations among protein expression patterns, developmental stages and embryo morphology (Katz-Jaffe and Gardner, 2007
; Katz-jaffe and McReynolds, 2013
).Analysis of the embryo culture medium has revealed the secretion and consumption of several proteins by the embryo. The increase in the culture medium of TNF (tumour necrosis factor) R1, IL (interleukin) 10, IL-6, VEGFA (vascular endothelial growth factor A), EMMPRIN (extracellular matrix metalloproteinase inducer), PLGF (placental growth factor), EpCAM (epithelial cell adhesion molecule), caspase-3, HE-4 (Human epididymis protein 4) and IL-6 meant that the embryo had secreted those proteins (
Dominguez et al., 2008
, Dominguez et al., 2015
; Lindgren et al., 2018
). Similarly, a decrease in CXCL (C-X-C motif chemokine ligand) 13, SCF (stem cell factor), TRAIL-R3 (receptor for the cytotoxic ligand TRAIL), MIP-1β (macrophage inflammatory protein-1β) and MSP-α (macrophage stimulating protein-α) in the culture medium indicated that the embryo had consumed or metabolized those proteins (Dominguez et al., 2008
). Furthermore, different protein patterns were observed according to implantation potential and embryo quality. Embryos that implanted consumed more CXCL13 and GM-CSF than non-implanted ones (Dominguez et al., 2010
; Robertson, 2007
;). In addition, culture medium from blastocysts showed higher concentrations of EMMPRIN than samples from arrested embryos. Higher concentrations of caspase-3 were found in culture medium from good quality blastocysts (Lindgren et al., 2018
).Artificial intelligence, in terms of the machine capacity to learn and exercise intelligent behaviour (
Simopoulou et al., 2018
), is making headway in human embryology. Publications increased seven-fold in 1 year (Curchoe and Bormann, 2019
), which shows the potential of using artificial intelligence to improve the efficiency of assisted reproductive treatments.It has recently been shown that artificial intelligence is a non-invasive tool with a high potential for predicting a live birth due to its capacity for recognizing patterns (
Miyagi, 2019
; Qiu et al., 2019
). Among artificial intelligence techniques, the most outstanding are the following: deep learning, which requires large databases (Chen et al., 2019
; Miyagi, 2019
) and hard computational efforts (Najafabadi et al., 2015
); convolutional neural networks, which are a subset of deep learning widely used in image analysis (Matusevičius et al., 2017
); the multilayer perceptron, whose artificial neural network (ANN) architecture contains intermediate layers between input and output data (Abiodun et al., 2018
; Milewski et al., 2017
; Zador, 2019
); and Bayesian networks, which are probabilistic models (Hernández-González et al., 2018
; Uyar et al., 2015
). All these techniques use artificial neural networks as the basis of the model.ANNs are computational techniques inspired by the functioning of the human brain whose main objective is to learn tasks that seek to solve complex problems (
Vanneschi and Castelli, 2018
). This technique simulates a biological neuron and imitates its ability to learn through trial and error. The characterization of an ANN is achieved via its self-learning through training. A commonly used algorithm for supervised learning is back-propagation (Gupta and Sexton, 1999
).Nowadays, the potential of ANNs is being increasingly explored mainly by associating other techniques, such as genetic algorithms, to optimize the results. The latter are evolutionary algorithms based on both Darwin's theory of evolution and Gregor Mendel's laws of genetics. This methodology generates a random population of individuals that is evaluated during the evolutionary process. In this process, the most qualified individuals are maintained in the next generations. After the creation of a new population, the process is repeated until a satisfactory solution is found (
Ghaheri, 2015
; Rigla et al., 2018
; Rosa and Luz, 2009
). In this case, the individuals are the ANNs (Takahashi et al., 2016
).The introduction of artificial intelligence to IVF laboratories would allow the analysis of raw embryo images without previous manual annotations. Even though the incorporation of continuous monitoring increases the objectivity of embryo assessment, manual annotations are subject to disparity among embryologists (
Martínez-Granados et al., 2017
; Storr et al., 2017
; Sundvall et al., 2013
). The introduction of systems with computer vision capable of extracting objective and standardized image information (Danuser, 2011
) and performing automatic annotations could reduce the subjectivity of embryo selection (Manna et al., 2013
). Macroscopic characteristics such as the number of cells, texture or movement patterns could be learned through training data.Methods to fully automatize the evaluation of mammalian embryos have been already proven to work (
Rocha et al., 2017
), and the next step is to transfer this knowledge to human embryology (Blank et al., 2019
). Recently, several approaches using artificial intelligence techniques for embryo classification and prediction of clinical outcomes have been published (see, for example, Bormann et al., 2020
; Chavez-Badiola et al., 2020
; Dirvanauskas et al., 2019
; Khosravi et al., 2019
; Rad et al., 2018
; Rocha et al., 2018
; Tran et al., 2019
; Zaninovic et al., 2018
).- Zaninovic N.
- Rocha J.C.
- Zhan Q.
- Toschi M.
- Malmsten J.
- Nogueira M.
- Meseguer M.
- Rosenwaks Z.
- Hickman C.
Application of artificial intelligence technology to increase the efficacy of embryo selection and prediction of live birth using human blastocysts cultured in a time-lapse incubator.
Fertil. Sterility Sci. Congr. Suppl. Oral. Poster Sess. Abstr. 2018; 110: e372
The objective of the current study was to develop a combinative predictive model based on artificial intelligence. First, morphological data from several embryos were used to create an ANN capable of predicting live birth. Donated oocytes were used, assuming that, as they were obtained from young women, they would result in mostly euploid embryos (
Dang et al., 2019
; Rubio et al., 2003
;). Second, information from blastocyst images and proteomic information was used to predict the potential of a euploid embryo to lead to a live birth. To the authors’ knowledge, this is the first approach to predict the likelihood of achieving a live birth by relying on machine learning and proteomic data.Materials and methods
Study design and participants
In this single-centre project, two populations were enrolled (Figure 1): 131 recipients of the oocyte donation programme of IVI Valencia (no preimplantation genetic testing [PGT]) and 81 women using autologous eggs in the PGT for aneuploidies (PGT-A) programme. The resulting embryos were individually cultured up to blastocyst stage in a continuous monitoring system (EmbryoScope; Vitrolife, Denmark).
Figure 1The study design shown as a flow diagram representing the distribution of embryos in the project. ANN, artificial neural network; PGT, preimplantation genetic testing.
A single time-lapse image from each embryo acquired at 111.5 ± 1.5 h of development was analysed using computational vision to extract information (as described below). Additionally, 20 µl of culture medium (Gems; Genea Biomedx, Australia) was collected for the proteomic analysis from the biopsied embryos and eight control samples (medium in which no embryos had been cultured). Samples were obtained on day 5 of development and were stored at –80°C until the proteomic analysis. Only the medium from euploid embryos was analysed after single-embryo transfer.
A total of 212 embryos was selected for application of the artificial intelligence technique: the first group consisted of 131 embryos obtained from the oocyte donation programme, and the second group included 81 embryos from autologous treatments with proteomic information. After the image analysis, 26 embryos were excluded from the second group: 19 blastocysts were outside the zona pellucida at 111 ± 1.5 h of development, which made image analysis difficult and not comparable with the 186 images finally remaining, and seven embryos did not reach the blastocyst stage at the proposed time and were discarded due to their early developmental stage. Therefore, the second group included 55 embryos for analysis, of which 11 were used for the blind test. Thus, the database totalled 186 embryos to undergo application of the artificial intelligence technique.
Ovarian stimulation in treatments with donated oocytes
Donors received stimulatory treatment using the conventional ovarian stimulation protocol with gonadotrophin-releasing hormone (GnRH) agonist treatment. GnRH agonist (Decapeptyl; Ipsen Pharma, Spain) was administered by intramuscular injection until more than eight follicles had reached a mean diameter of 18 mm or more. Transvaginal oocyte retrieval was scheduled for 36 h later. Endometrial preparation was undertaken using hormone replacement therapy as described by Cerrillo and colleagues (
Cerrillo et al., 2017
). After embryo transfer, oocyte recipients received a daily dose of 400 mg of vaginal micronized progesterone (Progeffik; Lab. Effik, Spain) every 12 h as luteal phase support.Ovarian stimulation in treatments with autologous oocytes
GnRH-antagonist treatments were applied, the GnRH-agonist being administered when at least three leading follicles had reached a mean diameter of 18 mm. Transvaginal oocyte retrieval was scheduled for 36 h later through follicular aspiration, and oocytes were washed in gamete medium (Cook Medical, Australia).
Oocyte retrieval and embryo incubation
Oocytes were cultured in fertilization medium (Origio; CooperSurgical, Denmark) in 5% CO2 and 5% O2 at 37°C. Denudation was carried out just before intracytoplasmic sperm injection (ICSI), 4 h after oocyte retrieval, using mechanical and chemical procedures (pipetting in 40 IU/ml hyaluronidase). ICSI was performed in a HEPES-buffered gamete medium at × 400 magnification using an Olympus IX7 microscope (Olympus Corporation, Japan). Finally, oocytes were placed in EmbryoSlides (Vitrolife, Denmark) pre-equilibrated to blastocyst stage with 28 µl of single-step medium (Gems; Genea Biomedx, Australia) and 1.6 ml of mineral oil.
Embryos were cultured individually up to the fifth or sixth day of development in the time-lapse system EmbryoScope (Vitrolife, Denmark). Successful fertilization was assessed at 16–19 h after ICSI.
Embryo morphology was evaluated on day 3 (62–72 h after ICSI) based on digital images, taking into consideration the number, the symmetry of the blastomeres, the percentage of fragmentation and the degree of compaction. Blastocysts were scored on day 5 (120 h after ICSI) based on the Association for the Study of Biology of Reproduction (ASEBIR) criteria (Supplementary Tables 1, 2 and 3) and the KIDScore Day 5 (with EmbryoViewer software; Vitrolife, Denmark). Embryologists annotated the morphokinetic parameters: the timings of cell divisions to the 2-cell (t2), 3-cell (t3), 4-cell (t4), 5-cell (t5), 8-cell (t8), 9-cell (t9), compaction (CP), morula (tM), start of blastulation (tSB), blastocyst (tB) and expanding blastocyst (tEB) were calculated, with the start of ICSI being used as t0. The durations of the cell cycle intervals t3–t2 (cc2), t4–t3 (s2), t5–t2, t5–t3 (cc3) and t8–t5 (s3) were also calculated. If there was more than one embryo of the same morphological quality, the score provided by the KIDScore Day 5 decided which would be transferred.
PGT-A
Embryos were taken out of the incubator on day 3 of development to undergo a small laser incision traversing the zona pellucida (assisted hatching) using a Lykos laser (Hamilton Thorne, USA). This procedure made the trophectoderm biopsy on day 5 of culture easier, when a biopsy pipette was used to remove approximately 5 cells. Chromosome analysis was performed using next-generation sequence technology (Thermo Fisher Scientific, USA).
Embryo transfer and clinical outcome
A single-embryo transfer of one blastocyst was performed for all patients; for those in the PGT-A programme, the blastocyst had previously been vitrified and warmed using the Cryotop method (Kitazato Biopharma, Japan). Embryo selection for transfer was based on chromosomal status, morphology and morphokinetics. The β-HCG concentration was determined 10 days after embryo transfer, and clinical pregnancy was confirmed by the presence of gestational sac at the fifth week of pregnancy. Finally, patients inform about live birth after the delivery.
Protein analysis in spent culture media
The relative concentrations of 92 proteins from 81 samples of spent embryo culture medium and eight control samples (medium in which no embryos had been cultured) were analysed using Proseek Multiplex Assays (Olink Bioscience, Sweden) based on proximity extension assay (PEA) technology.
Proteins were measured using the Olink Inflammation panel (Olink Proteomics, Sweden) according to the manufacturer's instructions. The PEA technology used for the Olink protocol has been well described (
Assarsson et al., 2014
); it enables 92 proteins to be analysed simultaneously, using 1 µl of each sample. In brief, pairs of oligonucleotide-labelled antibody probes bind to their target protein. If the two probes are closely located, the oligonucleotides will hybridize in a pairwise manner. The addition of a DNA polymerase leads to a proximity-dependent DNA polymerization event, generating a unique polymerase chain reaction (PCR) target sequence. The resulting DNA sequence is subsequently detected and quantified using a microfluidic real-time PCR instrument (Biomark HD; Fluidigm, USA). Data are then quality controlled and normalized using an internal extension control and an inter-plate control, to adjust for intra- and inter-run variation. The final assay readout is presented as Normalized Protein eXpression (NPX) values, which are arbitrary units on a log2 scale, where a high value corresponds to a higher protein expression. Hence, a 1 NPX difference means a doubling of protein concentration. All assay validation data (detection limits, intra- and inter-assay precision data, reproducibility and specificity) are available on manufacturer's website (www.olink.com).Artificial intelligence model: image analysis, collinearity analysis, ANN and genetic algorithms
Standardization of the 186 blastocyst images was necessary before applying the artificial intelligence technique. The methodology used was as described by Rocha and colleagues (
Rocha et al., 2017
). The images were initially imported automatically into Matlab software (MathWorks, USA) and a standardization algorithm was run to normalize them in terms of contrast and resolution (). Afterwards, blastocyst images were segmented into regions of interest. Using the Hough transform, the algorithm analysed three regions separately (Hassanein et al., 2015
; Mukhopadhyay and Chaudhuri, 2014
): the area of the expanded blastocyst, the inner cell mass and the trophectoderm. Finally, the algorithm obtained 33 mathematical variables by using the measurement of the area, the number of pixels present in each segmented portion of the blastocyst image, the binary patterns (Huang et al., 2011
) and the texture analysis (Bino et al., 2012
). These variables were chosen seeking to represent all the relevant characteristics of the blastocyst for embryo quality assessment and live birth prediction through artificial intelligence.After the embryo images had been standardized and isolated, and the mathematical variables had been extracted, a collinearity analysis was used to check if the variables were correlated with one another based on the variance inflation factor (VIF). The VIF represents the degree of independence or non-redundancy between a variable and another independent variable. According to previous studies, collinear variables were considered to be those with VIF values higher than 10 (
O'Brien, 2007
; Walczak and Cerpa, 1999
); removal of collinear variables reduced the number of variables representing the human embryo to 20 (Supplementary Table 4).Likewise, a collinearity analysis was performed to discriminate the independent and non-redundant proteins. As it is shown in detail in ‘Collinearity analysis for proteins’, below, seven proteins were suitable for using as input to the ANN, in conjunction with the 20 morphological variables. The artificial intelligence technique associated ANNs (multilayer perceptron) with the genetic algorithm by using the back-propagation learning algorithm (
Gupta and Sexton, 1999
) for the training phase. The genetic algorithm used the ANNs as individuals in a population, which, over generations, end up selecting the best ANN (i.e. the one with the highest accuracy for live birth prediction).The dataset of 131 embryos was randomly divided into 70% for training, 15% for validation and 15% for testing the ANN. Of the dataset of 55 embryos, 20% were used for the blind test, and the remaining 44 embryos were randomly divided into 68% for training, 16% for validation and 16% for testing the ANN (
Kalpana et al., 2015
).Statistical analysis
Statistical tests were applied to probe significant differences in the values of each protein in conditioned compared with control media and in conditioned media from implanted compared with non-implanted embryos. A t-test was used for parameters with a normal distribution, and a Wilcoxon rank sum test for those with a non-normal distribution. Values of P that were <0.05 were considered statistically significant.
The image analysis and the final model were tested using two techniques. First, receiver operating characteristic (ROC) curves were used to analyse the artificial intelligence results for pattern recognition. The resulting graph represents the ratio of true test positives to total positives (the sensitivity) per the false positive fraction, i.e. the ratio of false test positives to total negatives (1 – specificity). An area under the ROC curve (AUC) of greater than 0.5 might indicate a predictive power to identify embryos that lead (LB+) or do not lead (LB–) to a live birth. The greater the AUC, the more favourable the compensation between sensitivity and specificity. It tells how much the model is capable of distinguishing among LB+ and LB– embryos. Second, confusion matrix methodology was used to analyse the intersection between the data provided by the model (the artificial intelligence system) and the real results. The authors considered as true positive the number of embryos that achieved a live birth, the model classifying them as positive, and as true negatives those embryos that did not achieve a live birth, the model classifying them as negative. The embryos wrongly classified as leading to a positive or negative live birth were described as false positive and false negative, respectively.
Ethical approval
An Institutional Review Board (IRB reference 1802-VLC-012-MM), which regulates and approves database analysis and clinical IVF procedures for research at IVI, approved the procedure and protocol on 10 April 2018. Additionally, the project complies with the Spanish law governing assisted reproductive technologies (14/2006).
Results
The mean age of the patients included in the study was 41.6 years, with a mean body mass index (BMI) of 23.2 kg/m3 for the women receiving autologous oocytes. Regarding the clinical outcome, this group showed a positive β-HCG of 63.0%, an implantation rate of 56.8% and a live birth rate of 47.0%. The patients included in the oocyte donation programme had a mean age of 37.9 years with a mean BMI of 22.9 kg/m3 and their treatments achieved a positive β-HCG of 68.7%, an implantation rate of 54.20% and a live birth rate of 40.5%.
Proteomic profile of preimplantation embryos
Of the total of 92 proteins, 67 had identical NPX values in all the samples analysed (conditioned and control media). The lack of variation in the signal for each sample was decisive for not including these proteins in the following analyses. Only 25 of the total protein samples analysed had different NPX values. The means of the NPX values for these 25 proteins in control and conditioned medium are shown in Figure 2. Higher concentrations of three proteins were detected in the spent embryo culture media compared with background concentrations in control media. These proteins were IL-8 (P = 0.025), IL-6 (P = 0.001) and uPA (urokinase-type plasminogen activator; P = 0.006). Furthermore, lower concentrations of 14 proteins were detected in spent embryo culture media compared with background concentrations. These proteins were DNER (Delta/Notch-like EGF-related receptor; P < 0.001), CSF-1 (macrophage colony-stimulating factor 1; P < 0.001), Flt3L (FMS-like tyrosine kinase 3 ligand; P < 0.001), SCF (P < 0.001), CD40 (P < 0.001), MCP-1 (methyl-accepting chemotaxis protein; P < 0.001), CX3CL1 (P < 0.001), CD6 (P < 0.001), TRAIL (P = 0.002), TNFRSF9 (tumour necrosis factor receptor superfamily member 9; P < 0.001), CD244 (P < 0.001), IL-18 (P < 0.001), CCL23 (P < 0.001) and IL-18R1 (P < 0.001).
Figure 2The grouped column chart represents the average of the Normalized Protein eXpression (NPX) value obtained using the proximity extension assay technique for the 25 useful proteins analysed in culture medium from 81 embryos generated from autologous eggs. Grey columns represent control media (n = 8), and blue columns represent conditioned media collected on day 5 of embryo culture.
The only protein with a different NPX value in implanted (NPX value 2.44) and non-implanted embryos (NPX value 2.76) was VEGFA (P = 0.017).
Collinearity analysis for proteins
The collinearity analysis of the 25 proteins demonstrated that most of them were highly correlated with one another (data not shown). Thus, after correcting the collinearity, seven independent and non-redundant proteins remained for use in the ANN: matrix metalloproteinase-1 (MMP-1), IL-6, VEGFA, uPA, TNF-related activation-induced cytokine (TRANCE), Flt3L and DNER. The relative NPX values for each protein are shown in Table 1.
Table 1NPX values obtained using the PEA technique for the seven independent proteins resulting from the collinearity analysis
| Protein | NPX value (mean ± SD) | ||
|---|---|---|---|
| LB+ (n = 38) | LB– (n = 43) | P-value | |
| MMP-1 | –0.39 ± 0.83 | –0.49 ± 0.53 | 0.579 |
| IL-6 | 0.94 ± 0.53 | 0.75 ± 0.67 | 0.261 |
| VEGFA | 2.40 ± 0.37 | 2.59 ± 0.53 | 0.137 |
| uPA | 0.67 ± 0.42 | 0.61 ± 0.52 | 0.634 |
| TRANCE | 0.19 ± 0.51 | –0.05 ± 0.15 | 0.023 |
| Flt3L | 5.87 ± 0.20 | 6.00 ± 0.27 | 0.051 |
| DNER | 4.57 ± 0.29 | 4.71 ± 0.27 | 0.077 |
LB+, positive for live birth; LB–, negative for live birth; NPX, Normalized Protein eXpression; PEA, proximity extension assay; SD, standard deviation.
Artificial intelligence model
An extraction of relevant variables from the 131 blastocyst images included in the first population was required to design the ANN using morphological data. The training performed with 70% of the images was capable of correctly classifying 89% of the embryos as either LB+ or LB– (true positive, 33; true negative, 48; false positive, 6; false negative, 4), with 95% correctly classified in the test (true positive, 6; true negative, 13; false positive, 1; false negative, 0).
Regarding the validation of the ANN with the second embryo population and the additional proteomic data, the three most efficient architectures obtained using the genetic algorithm technique are shown in Table 2. Considering only the test data, the ANN was successful in predicting positive and negative live birth (mean of total success 89.67%).
Table 2Test accuracies of the three most efficient ANN architectures in live birth prediction
| ANN architecture | Morphology from image analysis | Proteomic data | Testing data (n = 7) | ||
|---|---|---|---|---|---|
| Success for LB+ (%) (n = 4) | Success for LB– (%) (n = 3) | Total success (%) | |||
| 1 | 20 variables | MMP-1, IL-6 | 100 (AUC = 1) | 100 (AUC = 1) | 100 |
| 2 | 20 variables | MMP-1, IL-6, VEGFA, uPA, TRANCE, Flt3L and DNER | 100 (AUC = 0.9) | 80 (AUC = 0.9) | 85.7 |
| 3 | 20 variables | 25 proteins | 87.5 (AUC = 0.83) | 80 (AUC = 0.84) | 83.3 |
a The variables are described in Supplementary Table 4.
b IL-8, VEGFA, CD244, OPG, uPA, IL-6, MCP-1, TRAIL, CST5, IL-1α, CXCL1, CD6, SCF, IL-18, FGF-23, MMP-1, IL-18R1, TRANCE, CCL23, Flt3L, DNER, CD40, CX3CL1, TNFRSF9, CSF-1.ANN, Artificial Neural Network; LB+, positive for live birth; LB–, negative for live birth.
Regarding the test dataset, the ROC curves for the three architectures are shown in Figure 3. The architecture developed using IL-6 and MMP-1 achieved a correct classification of all the embryos as a positive or negative live birth in the training, validation and test phases (total success 100%). The resulting AUC to predict the positive and negative live birth reached the highest value, 1.0.
Figure 3Receiver operating characteristic (ROC) curves for live birth (LB) prediction using the testing dataset (n = 7) and the artificial neural network with architecture 1 (A), architecture 2 (B) and architecture 3 (C) (see Table 2 for information on the architectures). The y-axis represents the sensitivity, and the x-axis refers to 1 – Specificity. Class 1, positive live birth; Class 2, negative live birth.
The blind test for architecture 1 was performed with 11 embryos that had not previously been used, and reached an accuracy of prediction of 72.7% (Figure 4). It correctly classified eight embryos out of the total number (true positive 4, true negative 4, false positive 1 and false negative 2).
Figure 4Confusion matrix for the blind testing dataset using the artificial neural network (ANN) with architecture 1. The y-axis refers to the output value predicted by the ANN, and the x-axis to the real value. Green areas represent the embryos classified correctly, and orange ones the embryos classified incorrectly. Dark blue areas represent the total hits (72.7%) and total mistakes (27.3%), and light blue areas provide the right:wrong ratio for each column. LB+, positive for live birth; LB–, negative for live birth.
Discussion
Information from the blastocyst images and proteomic information gained from the analysis of the embryonic secretome were used to predict the potential of a euploid embryo to lead to a live birth.
The proteomic analysis of the culture media showed that, out of 92 proteins measured, only IL-6, uPA and IL-8 were differentially secreted by the developing human embryos. Previous studies revealed that the concentration of IL-6 in the culture medium could be useful in selecting the embryo for implantation (
Dominguez et al., 2015
). Another research group has demonstrated higher concentrations of IL-6 in embryos that reach blastocyst stage than in arrested ones (Lindgren et al., 2018
). In addition, there is evidence about the secretion of IL-6 by the endometrial epithelial cells that emphasizes the importance of this cytokine in embryo development (Dominguez et al., 2010
). The other prominent proteins found in the current work to be differentially secreted have also previously been detected. Whereas the first evidence of uPA in embryo culture media was in 1996 (Khamsi et al., 1996
), the presence of IL-8 was not identified in the embryo secretome until 2018 (Lindgren et al., 2018
), using PEA technology.At the end of the collinearity analysis, seven proteins remained for use in the artificial intelligence model. According to the current study's results, two of these proteins (IL-6 and uPA) were secreted by the embryos, two others (DNER and Flt3L) were the ones consumed, and one protein (VEGFA) was related to poor implantation. Regarding the other two proteins, the authors consider that they have passed the collinearity analysis due to their exclusive characteristics: MMP-1 was the only one with a negative mean value in both control medium and medium from the embryos; and TRANCE was not found in control media, but was present in all media that had contained an embryo (see Figure 2). Furthermore, TRANCE was the only one out of the seven proteins with significant difference in relative concentration between LB+ and LB– samples (Table 1).
The image analysis performed in the present research on the blastocyst pictures has been proven in a previous study that showed good results in classifying the quality of bovine blastocysts (
Rocha et al., 2017
). This software considered that 20 morphological variables (Supplementary Table 4) of human embryos were enough to predict the likelihood of achieving a live birth. These parameters were combined with the data from the proteomic analysis to develop the current artificial intelligence model of prediction.The current results demonstrated that the accuracy of prediction was higher as the ANN architecture improved. First, the model created using 25 proteins correctly classified 83.3% of the embryos used in the test. Second, the collinearity analysis was shown to be efficient as the predictive power improved to 85.7% when only the resulting seven independent proteins were used. Finally, the artificial intelligence model achieved the highest accuracy in predicting live birth (AUC = 1) when considering IL-6, MMP-1 and the 20 morphological variables. It is reported that these proteins play an important role in reproductive function. IL-6 is relevant for embryonic development (
Dominguez et al., 2010
, Dominguez et al., 2015
; Iles, 2019
), and MMP-1 has been detected in mammalian ovaries (Hulboy, 1997
) and human follicular fluid (Lee et al., 2005
).A recent publication demonstrated that, for embryo selection an objective time-lapse imaging algorithm is superior to the subjective blastocyst morphological scoring system (
Fishel et al., 2019
). The algorithm used had previously been published (Fishel et al., 2017
) and obtained an AUC of 67.43% for live birth prediction, compared with 61.74% using the blastocysts’ morphological grade. The accuracy of embryo selection using the current proposed model is higher than using standard morphological selection alone, as well as using algorithms developed with time-lapse images.This study differs from previously published approaches to artificial intelligence in terms of the experimental design, including the proteomic analysis, and the outcome. The first application of ANN to predict the outcome of an IVF treatment achieved an accuracy of 59% (
Kaufmann et al., 1997
). Since then, several predictive models have been developed based on different populations of patients, such as in cycles associated with male factor infertility (Wald, 2005
) or women with endometriosis (Ballester et al., 2012
). In addition, embryo morphokinetic parameters were exclusively used as input data for an ANN that predicted 70% of pregnancies (Milewski et al., 2017
). Machine learning methods have been used to combine morphokinetic algorithms and factors of infertility such as sperm motility, anti-Müllerian hormone (AMH) concentration and blastomere size on day 3 (Blank et al., 2019
).Embryo images provided by time-lapse systems have now become the subject of studies based on artificial intelligence. Deep learning techniques have been used to predict blastocyst quality and select the most appropriate embryo to transfer (
Khosravi et al., 2019
). Embryos classified as good quality by these authors’ deep neural network, called STORK, showed higher probabilities of leading to a live birth than those classified as bad quality (61.4% and 50.9%, respectively). Nevertheless, STORK cannot estimate the pregnancy rate, although it showed a very high AUC (0.98) in predicting blastocyst quality. Another deep learning model was recently developed by Tran and colleagues (Tran et al., 2019
) that used the complete video of embryo development to predict the likelihood of pregnancy, with an AUC of 0.93 in 5-fold stratified cross-validation. The likelihood of an IVF treatment ending up with a live birth has recently been analysed considering patient and treatment characteristics (Vogiatzi et al., 2019
).Limitations of the current study that should be considered are reflected in the design of the research. This model was trained with proteomic data and morphological variables from the image analysis of one time-lapse system. The total population of embryos was distributed into two groups with different phases (training, validation and testing), resulting in a small number of embryos in each. Additionally, only one laboratory and a unique culture medium were involved in this study. This could be considered as an advantage in the proteomic analysis, but applicability to other laboratories remains unclear. In addition, there is evidence of inter-batch protein and pH variability with the same medium (
Dyrlund et al., 2014
; Leonard et al., 2013
; Tarahomi et al., 2018
). In general, the models developed with ANNs could also be affected by the overfitting phenomenon. The current study tried to avoid this by defining the input and output variables. The overall success on the blind test may be considered as evidence (72.7% for the blind test versus 100% for training, validation and testing). It is also necessary to highlight that the model with autologous oocytes was built using only euploid embryos, so the clinical value lies in distinguishing the most viable embryo among those that test as euploid. Further studies should have a large sample size and multicentric nature, and should include data from different time-lapse systems to standardize the artificial intelligence model and globalize its use.In conclusion, the introduction of artificial intelligence to IVF laboratories would help embryologists to predict the success of an embryo for achieving a live birth. The combination of proteomic analysis of the embryo culture medium and morphological information from the blastocyst images has never previously been assessed using artificial intelligence techniques. The present preliminary research has shown the predictive power of this combination. The ANN achieved excellent accuracy for detecting euploid embryos capable of resulting in a live birth, especially in terms of IL-6 and MMP-1. In fact, the model proposed in this manuscript is a promising tool to select the most successful embryo of a euploid cohort. In further studies, the ANN should be retrospectively tested with an appropriately sized study to confirm the effectiveness of this innovative method before its prospective application.
Acknowledgements
The authors acknowledge the embryologists and technicians of the IVF laboratory at IVI-RMA Global Valencia for their clinical support. The authors also thank the contribution of the IVI Foundation and the State University of São Paolo in this project. This work was supported by the Ministry of Science and Universities CDTI (IDI-20191102) awarded to M.M.; grants # 2017/19323-5, 2018/19371-2 and 2018/24252-2 from the São Paulo Research Foundation (FAPESP); and from the Spanish Ministry of Economy and Competitiveness through the Miguel Servet programme [CPII018/00002] awarded to F.D.
Appendix. Supplementary materials
REFERENCES
- State-of-the-art in artificial neural network applications : A survey.Heliyon. 2018; 4: e00938
- Homogenous 96-Plex PEA Immunoassay Exhibiting High Sensitivity, Specificity, and Excellent Scalability.PLoS One. 2014; 9: e95192
- Nomogram to predict pregnancy rate after ICSI – IVF cycle in patients with endometriosis.Hum. Reprod. 2012; 27: 451-456
- Gray Level Co - Occurence Matrices: Generalisation and some new features.Int. J. Comput. Sci. Eng. Inf. Technol. 2012; 2: 151-157
- Prediction of implantation after blastocyst transfer in in vitro fertilization : a machine-learning perspective.Fertil. Steril. 2019; 111: 318-326
- Consistency and objectivity of automated embryo assessments using deep neural networks.Fertil. Sterility Sci. Congr. Suppl. Oral Poster Sess. Abstr. 2020; 113: 781-787
- Human chorionic gonadotropin (hCG) in the secretome of cultured embryos: Hyperglycosylated hCG and hCG-free beta subunit are potential markers for infertility management and treatment.Reprod. Sci. 2013; 20: 1038-1045
- Impact of Endometrial Preparation Protocols for Frozen Embryo Transfer on Live Birth Rates.Rambam Maimonides Med. J. 2017; 8: e0020
- Predicting pregnancy test results after embryo transfer by image feature extraction and analysis using machine learning.Sci. Rep. 2020; 10: 4394
- Using Deep Learning with Large Dataset of Microscope Images to Develop an Automated Embryo Grading System.Fertil. Reprod. 2019; 1: 51-56
- Artificial intelligence and machine learning for human reproduction and embryology presented at ASRM and ESHRE 2018.J. Assist. Reprod. Genet. 2019; 36: 591-600
- Preimplantation Genetic Testing of Aneuploidy by Next Generation Sequencing : Association of Maternal Age and Chromosomal Abnormalities of Blastocyst.Open Access Maced J. Med. Sci. 2019; 7: 4427-4431
- Essay Computer Vision in Cell Biology.Cell. 2011; 147: 973-978
- Embryo development stage prediction algorithm for automated time lapse incubators.Comput. Methods Programs Biomed. 2019; 177: 161-174
- Comparative protein-profile analysis of implanted versus non-implanted human blastocysts.Hum. Reprod. 2008; 23: 1993-2000
- Embryologic outcome and secretome profile of implanted blastocysts obtained after coculture in human endometrial epithelial cells versus the sequential system.Fertil. Steril. 2010; 93: 774-782e1
- New strategy for diagnosing embryo implantation potential by combining proteomics and time-lapse technologies.Fertil. Steril. 2015; 104: 908-914
- International Committee for Monitoring Assisted Reproductive Technologies world report : Assisted Reproductive Technology 2008, 2009 and 2010.Hum. Reprod. 2016; 31: 1588-1609
- Unconditioned commercial embryo culture media contain a large variety of non-declared proteins : a comprehensive proteomics analysis.Hum. Reprod. 2014; 29: 2421-2430
- Factors influencing the success of in vitro fertilization for alleviating human infertility.J. Vitr. Fertil. Embryo. Transf. 1984; 1: 3-23
- Evolution of Embryo Selection for IVF from Subjective Morphology Assessment to Objective Time-Lapse Algorithms Improves Chance of Live Birth.Reprod. Biomed. Online. 2019; 40 (Elsevier Ltd): 61-70
- Live births after embryo selection using morphokinetics versus conventional morphology: a retrospective analysis.Reprod. Biomed. Online. 2017; 35: 407-416
- ART in Europe, 2014: Results generated from European registries by ESHRE.Hum. Reprod. 2018; 33: 1586-1601
- The Applications of Genetic Algorithms in Medicine.Oman Med. J. 2015; 30: 406-416
- Comparing backpropagation with a genetic algorithm for neural network training.Omega. 1999; 27: 679-684
- A Survey on Hough Transform, Theory, Techniques and Applications.Int. J. Comput. Sci. Issues. 2015; 12: 139-156
- Approaches to the study of the cell secretome.Expert Rev. Proteomics. 2007; 4: 239-248
- Fitting the data from embryo implantation prediction: Learning from label proportions.Stat Methods Med. Res. 2018; 27: 1056-1066
- Metabolomics : Current technologies and future trends.Proteomics. 2006; 6: 4716-4723
- Local Binary Patterns and Its Application to Facial Image Analysis: A Survey.EEE Trans Syst. Man., Cybern. 2011; 4: 1-17
- Matrix metalloproteinases as mediators of reproductive function.Mol. Hum. Reprod. 1997; 3: 27-45
- Secretome profile selection of optimal IVF embryos by matrix-assisted laser desorption ionization time-of-flight mass spectrometry.J. Assist. Reprod. Genet. 2019; 36: 1153-1160
- Immunosuppressive activity and alpha interferon concentrations in human embryo culture media as an index of potential for SUC- cessfal implantation.Fertil. Steril. 1992; 57: 637-640
- Pattern classification of EEG signals on different states of cognition using linear and nonlinear classifiers.Res. J. Appl. Sci. Eng. Technol. 2015; 11: 623-629
- Embryology in the era of proteomics.Theriogenology. 2007; 1: S125-S130
- Embryology in the era of proteomics.Fertil. Steril. 2013; 99: 1073-1077
- The application of neural networks in predicting the outcome of in-vitro fertilization.Hum. Reprod. 1997; 12: 1454-1457
- Expression of urokinase-type plasminogen activator in human preimplantation embryos.Mol. Hum. Reprod. 1996; 2: 273-276
- Deep learning enables robust assessment and selection of human blastocysts after in vitro fertilization.npj Digit. Med. 2019; 2: 1-9
- Omics as a window to view embryo viability.Fertil. Steril. 2015; 103: 333-341
- The expression of matrix metalloproteinase-9 in human follicular fluid is associated with in vitro fertilisation pregnancy.BJOG An. Int. J. Obstet. Gynaecol. 2005; 112: 946-951
- Variability in protein quality used for embryo culture: Embryotoxicity of the stabilizer octanoic acid.Fertil. Steril. 2013; 100: 544-549
- Differences in secretome in culture media when comparing blastocysts and arrested embryos using multiplex proximity assay.Ups J. Med. Sci. 2018; 123: 143-152
- Identification of apolipoprotein A1 in the human embryonic secretome.Fertil. Steril. 2011; 96: 422-427
- Artificial intelligence techniques for embryo and oocyte classification.Reprod. Biomed. Online. 2013; 26: 42-49
- Inter-laboratory agreement on embryo classification and clinical decision: Conventional morphological assessment vs. time lapse.PLoS One. 2017; 12: 1-13
- Embryo cell detection using regions with convolutional neural networks.CEUR Workshop Proc. 2017; 1856: 89-93
- How much information about embryo implantation potential is included in morphokinetic data? A prediction model based on artificial neural networks and principal component analysis.Adv. Med. Sci. 2017; 62: 202-206
- Feasibility of deep learning for predicting live birth from a blastocyst image in patients classified by age.Reprod. Med. Biol. 2019; 18: 190-203
- A survey of Hough Transform.Pattern Recognit. 2014; 48: 993-1010
- Deep learning applications and challenges in big data analytics.J. Big Data. 2015; 2: 1-21
- Embryonic soluble HLA-G as a marker of developmental potential in embryos.Hum. Reprod. 2005; 20: 138-146
- A caution regarding rules of thumb for variance inflation factors.Qual. Quant. 2007; 41: 673-690
- Embryo-Derived Platelet Activating Factor, a Marker of Embryo Quality and Viability Following Ovarian Stimulation for in Vitro Fertilization.J. Vitr. Fertil. Embryo. Transf. 1990; 7: 3-8
- Personalized prediction of live birth prior to the first in vitro fertilization treatment : a machine learning method.J. Transl. Med. 2019; 18: 1-8
- Blastomere cell counting and centroid localization in microscopic images of human embryo.IEEE 20th Int. Work Multimed Signal Process MMSP 2018. 2018; (IEE): 1-6
- Artificial Intelligence Methodologies and Their Application to Diabetes.J. Diabetes Sci. Technol. 2018; 12: 303-310
- GM-CSF regulation of embryo development and pregnancy.Cytokine Growth Factor Rev. 2007; 18: 287-298
- Is AI assessment of morphokinetic data and digital image analysis from time-lapse culture predictive of implantation potential of human embryos?.Fertil Sterility Sci. Congr. Suppl. Oral. Poster Sess. Abstr. 2018; 110: E373
- A Method Based on Artificial Intelligence To Fully Automatize The Evaluation of Bovine Blastocyst Images.npj Digit. Med. 2017; 7: 1-10
- Conceitos Básicos de Algoritmos Genéticos: Teoria e Práctica.Ulbra-ToBr. 2009; : 27-37
- Chromosomal abnormalities and embryo development in recurrent miscarriage couples.Hum. Reprod. 2003; 18
- The image processing handbook.CRC press, 2016 (editor)
- The role of pregnancy-specific P-l glycoprotein (SP1) in assessing human blastocyst quality in vitro.Hum. Reprod. 1994; 11: 1038-1042
- Are computational applications the “crystal ball” in the IVF laboratory? The evolution from mathematics to artificial intelligence.J. Assist. Reprod. Genet. 2018; 35: 1545-1557
- Inter-observer and intra-observer agreement between embryologists during selection of a single Day 5 embryo for transfer: A multicenter study.Hum. Reprod. 2017; 32: 307-314
- Inter- and intra-observer variability of time-lapse annotations.Hum. Reprod. 2013; 28: 3215-3221
- Optimization of artificial neural network by genetic algorithm for describing viral production from uniform design data.Process Biochem. 2016; 51: 422-430
- pH stability of human preimplantation embryo culture media: effects of culture and batches.RBMO. 2018; 37: 409-414
- Soluble ligands and their receptors in human embryo development and implantation.Endocr. Rev. 2015; 36: 92-130
- Deep learning as a predictive tool for fetal heart pregnancy following time-lapse incubation and blastocyst transfer.Hum. Reprod. 2019; 34: 1011-1018
- Predictive modeling of implantation outcome in an in vitro fertilization setting: an application of machine learning methods.Med. Decis. Mak. 2015; 35: 714-725
- Multilayer Perceptrons.Encycl Bioinforma Comput Biol. 2018; 1: 1-9
- An artificial neural network for the prediction of assisted reproduction outcome.J. Assist. Reprod. Genet. 2019; 36: 1441-1448
- Heuristic principles for the design of artificial neural networks.Inf. Softw. Technol. 1999; 41: 107-117
- Computational models for prediction of IVF / ICSI outcomes with surgically retrieved spermatozoa.Reprod. Biomed. Online. 2005; 11: 325-331
- A critique of pure learning and what artificial neural networks can learn from animal brains.Nat. Commun. 2019; 10
- Application of artificial intelligence technology to increase the efficacy of embryo selection and prediction of live birth using human blastocysts cultured in a time-lapse incubator.Fertil. Sterility Sci. Congr. Suppl. Oral. Poster Sess. Abstr. 2018; 110: e372
- A randomized clinical trial to evaluate the effect of granulocyte-macrophage embryo culture medium for in vitro fertilization.Fertil. Steril. 2013; 99: 1600-1609
Biography
Lorena Bori is doctoral researcher at IVIRMA Valencia, Spain. She received her biology degree in 2016 and her Master's degree in the biotechnology of human assisted reproduction in 2018 from the University of Valencia. Her primary field of research is embryo evaluation and selection using non-invasive methodologies, especially artificial intelligence.
Article Info
Publication History
Published online: October 07, 2020
Accepted:
September 30,
2020
Received in revised form:
September 17,
2020
Received:
May 11,
2020
Declaration: The authors report no financial or commercial conflicts of interestIdentification
Copyright
© 2020 Reproductive Healthcare Ltd. Published by Elsevier Ltd. All rights reserved.
30537-X&id=fx1.jpg){kind=link}
30537-X&id=gr1.jpg){kind=link}
30537-X&id=gr2.jpg){kind=link}
30537-X&id=gr3.jpg){kind=link}
30537-X&id=gr4.jpg){kind=link}