JBRA Assisted Reproduction 2025;29(Suppl.2 SBRA 2025):29
Poster Presentation
29^th Annual Congress of the SBRA. São Paulo/SP - Brazil, 2025
doi: 10.5935/1518-0557.20263553

P-017. Artificial Intelligence Models for Predicting Pregnancy Outcomes in In Vitro Fertilization: A Systematic Review of Clinical and Laboratory Data Integration

Laís Gonçalves Aguiar¹, Laura Barros Possa¹, Anna Beatriz Santana Marques¹, Ana Márcia de Miranda Cota¹

¹Faculdade Ciências Médicas de Minas Gerais – Belo Horizonte - MG - Brasil

Objective: To evaluate the use of Artificial Intelligence (AI) algorithms for predicting pregnancy rates in In Vitro Fertilization (IVF) treatments, with an emphasis on the integration of clinical and laboratory data.
Methods: The search was conducted in the PubMed database in June of 2025 using the descriptors “artificial intelligence,” “pregnancy,” “in vitro fertilization,” and “clinical data.” Studies published in 2024 and 2025 were included. The search initially identified 35 articles, of which 21 were selected after reading the titles and abstracts. These 21 articles were then read in full. Of these, 10 were excluded for not being relevant to the research subject, and also due to the unavailability of the full texts.
Results: Artificial intelligence has proven highly effective in predicting clinical pregnancy in in vitro fertilization (IVF) treatments, analyzing a variety of data: clinical, demographic, laboratory, embryological, and ovulation-related variables, as well as other variables such as basal progesterone. In some studies, a particular artificial intelligence model performed better than others. Center-Specific Machine Learning Models (CSMs), when compared to the National Standard Model used in the US, based on the Society for Assisted Reproductive Technology (SART) registry, performed better in minimizing false positive and false negative results, according to F1 and PR-AUC metrics (p<0.05) for clinical pregnancy outcomes. AI was also used to predict the exact day of ovulation in frozen embryo transfer cycles in a natural cycle, to assess the impact of this prediction on clinical pregnancy rates. The model was able to predict the day of ovulation in over 92% of cases and resulted in improved clinical pregnancy rates. To evaluate the clinical pregnancy rate, Random Forest models (AUC=0.85 and 0.80 in different studies), deep neural networks (DNNs) that had an average accuracy of 0.855 and AUC of 0.86 and AUC=0.73 (external validation) and artificial neural networks (ANN) that presented AUC=0.83 were used. In the prediction of live births, Support Vector Machine (SVM) obtained AUC=0.862, with an accuracy of 90.6% and an F1-score of 82.1% in real tests, and AdaBoost, a model based exclusively on images, had AUC=0.749. For embryo implantation and live birth, Random Forest presented AUC=0.725. Gradient Boosting Machines (GBM) was a model chosen from the use of Large Language Models (LLMs) such as GPT-4, which achieved an AUC of 0.89, outperforming traditional models.
Conclusion: The use of AI models in reproductive medicine has made an important improvement in decision-making regarding the viability of IVF. It integrates clinical, demographic, hormonal, laboratory, and embryological data to predict clinical pregnancy and live birth outcomes. This allows healthcare professionals to provide personalized, evidence-based guidance to couples, helping them decide whether IVF is a worthwhile investment. However, the standardization of the procedures and multicenter validation are necessary before routine clinical implementation. As a result, AI not only enhances success rates but also contributes to cost reduction by minimizing unnecessary IVF cycles. Therefore, AI serves as a strategic tool in promoting more effective, personalized, and economically sustainable fertility treatments.