A recent study published in the journal ClinicalMedicine explores the use of artificial intelligence (AI) to non-invasively predict embryonic ploidy from medical images. Embryonic aneuploidy, which refers to an abnormal chromosome count in the embryo, is a critical factor in in vitro fertilization (IVF), as its high prevalence can cause implantation failure, pregnancy loss, and congenital anomalies.
The incidence of aneuploidy in embryos in the early stage of IVF varies between 25% and 40%, increasing with maternal age. Currently, preimplantation genetic testing for aneuploidy (PGT-A) is the most widely used method to detect chromosomal abnormalities, although its high cost and invasive procedure make its accessibility difficult and raise ethical considerations.
AI has proven its value in other areas of medicine thanks to its ability to analyze large amounts of data quickly and accurately, and now appears to offer a potential solution in the field of assisted reproduction.
Through machine learning and deep learning models, AI could be able to assess the ploidy of an embryo without needing to perform a biopsy. However, this study highlights that while current AI techniques show potential for identifying ploidy, additional research is still needed to increase their reliability and clinical applicability before they can replace invasive methods such as PGT-A.
To examine the effectiveness of these algorithms, the research team conducted an extensive analysis of previous studies on AI and embryonic ploidy. Recognized scientific databases, such as PubMed, MEDLINE, Embase, IEEE, SCOPUS and Cochrane, were reviewed, evaluating publications until August 10, 2024.
The review included articles that presented specific diagnostic results, such as sensitivity and specificity, or that contained relevant contingency data. In total, 4,774 studies were examined, of which those that did not meet the required criteria were excluded. Finally, 20 studies were selected for analysis, with only 12 studies providing data suitable for detailed meta-analysis.
The analysis showed that the AI algorithms used in ploidy prediction achieved a sensitivity (Se) of 67%, a specificity (Sp) of 58% and an area under the curve (AUC) of 0.67 overall. However, by focusing on the most accurate contingency tables, the values improved to an Se of 71%, an Sp of 75%, and an AUC of 0.80, suggesting a significant improvement in the accuracy of the models.
For quality assessment, the researchers used the QUADAS-AI tool, designed to examine studies of diagnostic accuracy in AI, finding that 19 of the studies had a high risk or unclear risk of bias, mainly due to missing data from open source and external validation.
Likewise, factors that affected accuracy were identified, such as the type of algorithm used, maternal age and sample size, in addition to observations that models that integrated clinical data together with images had a higher AUC of 0.71 , compared to those that only used images, with an AUC of 0.62.
AI models were classified into three types of decision support systems (DSS): black box, matte box, and glass box. This classification reflects the transparency of each model in explaining its predictions, and five studies were found to use glass box models, five matte box models, and four black box models. More recent studies showed an increase in specificity and AUC, suggesting progressive improvement in AI models for ploidy prediction.
Currently, PGT-A remains the standard in the detection of aneuploidy in IVF, despite its risks. In addition to costs and ethical limitations, embryo biopsy, an invasive procedure, carries risks such as preeclampsia and placenta previa, without a significant increase in successful pregnancy rates. For these reasons, researchers are optimistic about the development of non-invasive AI methods for ploidy prediction, as they could help embryo selection without resorting to invasive methods.
However, this study concludes that current AI models do not yet have the necessary precision to replace PGT-A and that, for now, they should only be considered supportive tools in assisted reproduction. External validation and greater model transparency are critical to reducing the risk of bias and improving prediction accuracy. As AI technology advances, these models are expected to offer a practical and less invasive alternative to embryo evaluation, which would represent a significant change in the field of in vitro fertilization and ethical approaches to assisted reproduction.
Keep reading:
- Cod liver oil and its benefits: what we should know
- Penicillin use is related to lower risk of Parkinson’s disease
- Experts warn of increased harm from gambling on mental health