Artificial intelligence in assisted reproductive techniques

Introduction

Artificial intelligence (AI) transfers several tasks usually assigned to humans, relying on a series of machine learning algorithms with predefined functions to perform these tasks. The evolution of AI toward the development of neural networks determines the processing of a considerable volume of information⁽¹⁾.

Recent years have seen many advances in AI in assisted reproductive technology (ART). Thus, standard criteria were established regarding the automation of embryo evaluation in cultures using morphokinetic data, a real progress regarding the use of time-lapse microscopy (TLM), the early establishment of embryo viability, and the quality of selection for embryo transfer⁽²⁾.

These components of the AI system have not yet been validated in clinical practice; instead, they were used to develop embryo viability selection programs and to create TLM images. The first attempt to use artificial intelligence in the field of research was a study on an animal model from 1993, highlighting the penetration of sperm into the hamster oocyte. The application of AI in ART is essential because it can improve the analysis of the data obtained from the infertile couple, personalize the treatment related to infertility for each patient, the techniques for the evaluation and selection of embryos, oocytes and spermatozoa, as well as for the assessment of the ovarian reserve. Another possible application of neural networks used in artificial intelligence would be to find algorithms regarding the compatibility of recipients with egg donors based on initially defined criteria⁽³⁾.

To improve the performance of IVF laboratories, neural AI models must be able to identify embryo quality, predict blastocyst quality, and analyze uterine quality or other maternal factors regarding embryo implantation. In addition, developing a decision tree focused on each age group can optimize the actual chances of implantation.

Deep learning analysis of early embryonic biology is an emerging field. AI evaluation of images and intraembryonic cell mass objectively evaluates the embryo, allowing the prediction of implantation potential and the development of a personalized therapeutic strategy for in vitro fertilization (IVF) procedures. This goal has as its final objective the validation of some standardized techniques within ART by identifying some embryonic morphological and morphokinetic characteristics that increase the patient’s chances of having a normal pregnancy⁽⁴⁾.

The rapid implementation of AI techniques in clinical embryology is mainly due to the increased degree of variability regarding the experience of embryologists involved in IVF procedures. The artifacts that can appear result from the differences in the classification systems, the particularities of the IVF processes, and the personnel training. Another application of artificial intelligence is represented by the preimplantation genetic testing to detect possible aneuploidy for embryo transfer. To reduce the ambiguity of the results and eliminate the subjectivity of the staff in IVF clinics, the use of computerized analysis allows the evaluation of chromosomal integrity, the impact of microdeletions and mosaicism⁽⁵⁾.

Other roles regarding the application of AI in human reproductive medicine include the correct selection of patients, the efficiency of data from medical observation sheets, the personalized application of stimulation protocols, the identification of follicles and the assessment of their quality, the evaluation of the morphofunctional characteristics of sperm, the monitoring of the embryonic cell stage, the quality of the embryo for transfer, the prediction of pregnancies with favorable evolution, as well as the number of live newborns⁽¹⁾.

There are also a series of impediments related to implementing AI techniques: high costs, ethical aspects regarding the doctor’s degree of responsibility, the inertia of the medical system, and the lack of empathy regarding the patient. But, on the other hand, this technique can better help patients through a correct selection of them by increasing the accuracy of the diagnosis and the efficiency of the therapeutic management.

Providing personalized therapy to infertile patients increases the success rate of a pregnancy, improves their psychoemotional state, and reduces the costs related to ART procedures. Recently, artificial intelligence has been trying to establish an ART prediction model using noninvasive biomarkers to increase the therapeutic efficiency and implantation rate after embryo transfer⁽¹⁾. Currently, the AI procedures are semiautomatic, as no randomized clinical trials have been carried out to verify the validity of these neural models and analyze the data obtained on large groups of patients⁽⁴⁾.

In reproductive medicine, techniques based on artificial intelligence can optimize ART treatment schemes, but they are still not fully effective in predicting a normal pregnancy, nor can they identify the causes of implantation failure.

Another possible application of AI techniques would be in minimally invasive and robotic surgery that precedes ART, with the well-defined goal of improving the prognosis regarding future chances of fertility^(6,7).

The role of AI in oocytes evaluation

To increase the success rate of pregnancies, it is necessary to develop precise standards and techniques for selecting quality oocytes to identify the egg with the best development rate for the IVF process, as well as embryos without the risk of aneuploidy^(8,9).

In addition, compared to the prediction of oocyte quality, the degree of development of gametes must also be followed by applying AI methods, as well as the gene expression of human oocytes through omics technology or dynamic monitoring system^(8,10).

The identification of competent oocytes during the maturation process with high precision is carried out with an artificial neural network through the morphokinetic study at the cytoplasm level, using the velocimetric analysis of particle images⁽¹¹⁾. The noninvasive approach regarding oocyte development in clinical embryology, the prediction of blastocyst development and the transcriptomic changes in embryos with aneuploidy or death are just some of the potential applications of AI^(8,12).

AI analysis in sperm selection

The next step in reproductive medicine was represented by the use of AI of some support vector machine learning models for the classification of sperm morphology and motility, which by processing data on fertility prognosis, reached an overall accuracy of approximately 90%⁽¹³⁾. In an experimental study, Cavalera et al. developed a new predictive method based on the artificial neural network model regarding mouse sperm parameters with an accuracy of 91%⁽¹¹⁾. The processing of a large volume of information (images, laboratory analyses, personal data of patients, treatment schemes) required the adoption of a new processing method with deep learning, which analyzes the morphology and quality of spermatozoa, oocytes and embryos. The final goal is to organize a database that assesses at the population level the parameters regarding reproductive health at that time⁽¹⁴⁾.

The evaluation of the fertile potential of the spermatozoa is based on a system of analysis of their populations according to the morphokinetic study. The existence of idiopathic factors, the lack of laboratory standardization, inter- and intraobserver variability, multiple causes of infertility, and comorbidities require the development of objective and precise AI models^(3,13). For the prediction of the seminal quality, Badura et al. used a neural network related to sperm concentration, and the regain of motility was observed, with the role of early quantifying the male fertility rate⁽¹⁵⁾.

AI models and blastocyst/embryos selection

Although more than 44 years have passed since the birth of the first child obtained through IVF, the AI selection algorithm regarding the quality of the embryos has not been established⁽¹⁶⁾.

The prediction of the correct selection of the quality of the embryos requires the identification of a decision tree based on the morphokinetic parameters of the embryos, using the classic morphological score⁽¹⁷⁾. The decision-making strategy regarding the age of the patients for the elective cryopreservation of the oocytes and the success rate of the embryo transfer related to this can increase the accuracy of the IVF result^(18,19).

The lack of information on embryo implantation means that the methodology of machine learning models in predicting IVF results accurately does not exceed 68%⁽²⁰⁾. In an IVF cycle, the increase in the prediction index of embryo implantation using the naive Bayes model based on the number and morphological parameters of the transferred embryos increased the accuracy to 80%⁽²¹⁾. Bori et al., using the models validated by artificial neural networks at the level of the expanded diameter of the blastocyst and, respectively, at the cell cycle length of the trophectoderm of implanted and nonimplanted embryos, observed a prediction of implantation between 0.64 and 0.77⁽²²⁾. Currently, the selection of oocytes is performed visually or of embryos dynamically, following the morphology and development of the human zygote towards the blastocyst stage. In the conditions of carrying out the transfer of more than one embryo per cycle, there is an increased risk of complications relative to the potential increase in the success rate^(14,23).

On microscopic blastocyst images, Kragh et al. showed an automatic method for highlighting the intracellular mass’ and the trophectoderm’s segmentation process for the first time. The data on the embryo morphology (the quality of the intracellular mass and the trophectoderm) are basic elements regarding the selection criteria for embryo transfer⁽²⁴⁾.

Storr et al. have shown the effectiveness of the dynamic monitoring system of the growth stage of embryos⁽²⁵⁾, while other study questions, at this moment, the benefit of embryo selection using these AI algorithms⁽²⁶⁾. Studies on morphokinetic analysis can exclude those embryos with reduced implantation potential⁽¹⁷⁾.

The use of the AI technique in the blastocyst selection process by trained embryologists improved the identification rate of PGT-A euploid embryos intended for implantation⁽⁵⁾. After the analysis of videos and time-lapse images, the use of the EMA artificial neural network predictive model obtained a precision of 0.85 for blastocyst prediction and a precision of 0.72 for implantation on day 5/6⁽²⁷⁾.

Furthermore, the study carried out by VerMilyea et al., using the Life Whisperer AI model, showed a sensitivity of 70.1% for viable embryos and a specificity of 60.5% for nonviable embryos, improving embryologists’ accuracy by 24.7%⁽²⁸⁾.

Conclusions

Integrating the AI system into reproductive medicine can bring real benefits to infertile patients and doctors. AI algorithms applied in managing increasingly large databases and processing an increased volume of information can help enhance the accuracy of the selection process and the quality of oocytes, spermatozoa and embryos for ART. Thus, collecting a large amount of valid data and analyzing and integrating them can pave the way for further applications of artificial intelligence, such as machine learning. In addition, AI can collect, analyze, integrate and interpret medical data, genetics, laboratory analyses and medical images to support the doctor’s decision-making.

Automating IVF labs is another benefit of integrating AI models. In reproductive medicine, clinical models are constantly optimized to create expert systems based on the diagnosis assisted by artificial intelligence applications. 

 

Conflict of interest: none declared  
Financial support: none declared
This work is permanently accessible online free of charge and published under the CC-BY.