Deep Learning-Based Prediction of IVF Success: A Transformer Model Approach

Mahvash  Zargar; Seyed Masoud  Rezaeijo; Mahin  Najafian; Kobra   Shojaei; Vahideh  Yousefvand

doi:10.18502/fbt.v13i2.21932

Mahvash Zargar Fertility, Infertility and Perinatology Research Center, Department of Obstetrics and Gynecology, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran
Seyed Masoud Rezaeijo Department of Medical Physics, Faculty of Medicine, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran
Mahin Najafian Fertility, Infertility and Perinatology Research Center, Department of Obstetrics and Gynecology, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran
Kobra Shojaei Fertility, Infertility and Perinatology Research Center, Department of Obstetrics and Gynecology, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran
Vahideh Yousefvand Fertility, Infertility and Perinatology Research Center, Department of Obstetrics and Gynecology, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Iran

DOI: https://doi.org/10.18502/fbt.v13i2.21932

Keywords: Assisted Reproductive Technology; In Vitro Fertilization; Clinical Pregnancy Prediction; Endometrial Receptivity; Embryo Transfer Timing; Machine Learning; Deep Learning.

Abstract

Purpose: Predicting the success of assisted reproductive technology (ART) remains a significant challenge due to the complex interplay of clinical, embryological, and demographic factors. This study aimed to develop and evaluate machine learning models, particularly deep learning-based approaches, to identify key predictors of ART success and improve outcome prediction accuracy.

Materials and Methods: A retrospective study was conducted on 500 infertile couples undergoing ART treatment between 2019 and 2024. A comprehensive dataset, including 84 clinical, embryological, and demographic variables, was analyzed. The key predictors included endometrial thickness, endometrial pattern, embryo transfer day, and hormonal markers (PRL, LH). Four machine learning models were implemented: Decision Tree, Random Forest, XGBoost, and a Transformer-Based Model. Data preprocessing involved feature selection, missing data handling, normalization, and oversampling techniques to address class imbalance. The models were trained and validated using k-fold cross-validation, and performance was assessed using accuracy, precision, recall, and F1 score.

Results: The Transformer-Based Model achieved the highest accuracy (99.7%), outperforming traditional machine learning models. This performance was validated using k-fold cross-validation and oversampling to mitigate overfitting and ensure generalizability. Endometrial pattern (r = 0.69) and endometrial thickness (r = 0.82) were the strongest predictors of ART success, emphasizing the dominant role of uterine factors. While female age and infertility duration had a weak negative correlation, male infertility factors and lifestyle variables (smoking, alcohol consumption) showed minimal predictive significance. Model-based feature importance confirmed uterine and embryological factors as the primary determinants of ART success, indicating a potential shift in treatment strategies toward optimizing endometrial receptivity and embryo transfer timing.

Conclusion: This study highlights the superiority of deep learning models in ART success prediction, with uterine factors emerging as the strongest predictors. Integrating AI-driven predictive models into clinical practice can enable personalized ART treatment, improved patient counseling, and optimized embryo transfer strategies, ultimately enhancing fertility outcomes. However, the findings are based on data from a single medical center, and further multi-center validation is needed to confirm the model’s generalizability.