Page 19 - AIH-2-3

Artificial Intelligence in Health AI in embryo selection for ART

3.11. Transformative algorithms in enhancing blastomere detection compared to previous computational
embryo selection methods. Similarly, the CHLOE model revealed no
82
significant bias between XX and XY embryos (U = 204621,
Within the field of ART, a wide range of ML and AI methods p=0.208). In contrast, manual morphological grading and
have been utilized to improve embryo selection and the KIDScore algorithm, a tool to support embryologists
predict the likelihood of successful implantation during in decision-making, tended to favor male embryos, with
IVF procedures. These technologies aim to enhance the
precision, consistency, and efficiency of evaluating embryo XY embryos receiving higher scores than XX embryos
2
viability. One prominent approach involves the use of deep (U = 207604, p=0.0182;  = 19.843, p<0.00001). These
findings suggest that deep-learning approaches may help
learning models, such as iDAScore v1.0, for the objective mitigate sex-selection bias. However, deep learning
83
ranking of blastocysts, introduced by Cimadomo et al. Its is not always superior to manual methods. In a recent
17
successor, iDAScore v2.0, developed by Theilgaard Lassen double-blind non-inferiority trial involving 1,066 patients,
et al., incorporated morphokinetic parameters into 533 were assigned to the iDAScore group and 533 to
16
embryo ranking, and achieved an AUC ranging from 0.621 the morphological grading group. The iDAScore group
to 0.707. Ueno et al. further demonstrated that increasing showed a clinical pregnancy rate of 46.5% (248 of
23
training data with Gardner grading for both IDA-V1 533 patients), compared to 48.2% (257 of 533 patients) in
and V2 significantly enhanced predictive performance, the morphological grading group (risk difference = −1.7%;
with an AUC value of 0.736 for ongoing pregnancy 95% CI = −7.7, 4.3; p=0.62). 84
predictions. Moreover, studies by Bori et al. and Johansen
19
et al. introduced AI models using time-lapse images 4. Discussion
18
80
to evaluate embryo viability. Benchaib et al. employed
shallow artificial networks (e.g., multilayer perceptron and This comprehensive analysis of the systematic review aims
recurrent neural network) based on morphokinetic time- to address the RQs, assess the quality of the review, and
lapse parameters to predict viable embryos for transfer. propose directions for future study.
Berntsen et al. employed a deep-learning AI model 4.1. General discussion
24
based on Python and TensorFlow to sort 115,832 embryos
from 18 IVF centers, achieving AUC values between The term “ART” is gaining significant recognition
0.63 and 0.69. In another approach, Chen et al. used a in the context of infertility treatment. ART refers to
25
random forest model on 345 paired blastocyst cultures, any procedure involving the in vitro manipulation of
demonstrating transplant suitability for A- and B-grade oocytes (immature ova or egg cells) for reproductive
38
embryos comparable to euploid ones. Glatstein et al. purposes. Common ART procedures include IVF,
20
used convolutional neural networks and support vector ICSI, ET, and luteal phase assistance, although these
machines to predict live birth probabilities, with AUC techniques may cause perinatal complications and
85
values of 0.63 – 0.83 and achieving up to 90% accuracy in outcomes. Primarily used for infertility treatments,
some models. Salih et al. showed that ML outperformed ART includes techniques such as artificial insemination,
21
86
embryologists in predicting embryo morphology, with IVF, surrogacy, and the use of fertility medication.
an accuracy of 75.5% compared to the embryologists’ The implementation of individualized treatment
34
65.4%. Meanwhile, Pons et al. applied logistic regressions protocols and multidisciplinary team management
significantly improves treatment outcomes and safety,
to update the ASEBIR system for predicting blastocyst making ART a viable option for couples and individuals
implantation and live birth. To predict fertilization failure facing fertility issues. The ART process encompasses
probabilities (Logistic Regression Function and Threshold several critical stages, including controlled ovarian
22
Fertilization Failure) in ART cycles, Tian et al. utilized
Bayesian network modeling and achieved an accuracy of stimulation, pituitary downregulation, oocyte retrieval,
fertilization, ET, embryo selection, and luteal phase
91.3%, aiming to optimize IVF and ICSI treatments.
support. In recent years, AI has been integrated into
39
Recent deep learning models show significant ART to enhance and automate embryo selection by
advancements in performance. For example, the analyzing microscopy images and identifying optimal
Embryo2live model outperformed traditional morphology embryos for transfer or cryopreservation. AI is also
77
grading by increasing live birth rates from 23.0% to 71.3% used to reduce interobserver variability, optimize drug
for top embryo selections. The Esava model, developed dosing, enhance sperm selection and oocyte quality
81
for the quantitative evaluation of IVF embryos, reported evaluation, and increase overall clinical efficiency. By
87
high rates for precision (0.9940), recall (0.9121), and enhancing outcomes and decision-making processes
mean average precision (0.9531), demonstrating superior through sophisticated algorithms and data analysis, AI
Volume 2 Issue 3 (2025) 13 https://doi.org/10.36922/aih.7170