Abstract
Despite advancements in assisted reproductive technology (ART), recurrent implantation failure (RIF) continues to pose a significant challenge to achieving pregnancy. We included 2,463 retrospective RIF patients with no gynecological and anatomical anomalies who were referred to a clinical immunologist and received targeted immunotherapies. Twenty-three variables were used to develop a deep learning (TabNet) model to predict live births. Statistical analyses were used to compare characteristics between live birth and implantation failure groups. Model performance was evaluated using a confusion matrix, the receiver operating characteristic (ROC) curve, and calibration plots. Our model showed an accuracy of 87.4% and an AUROC of 0.952. According to the model, when there were no missing input variables, the most important features were age, Th1/Th2 ratio, BMI, anti-thyroid peroxidase (anti-TPO), antinuclear antibodies (ANA), anti-dsDNA, and anti-tissue transglutaminase (anti-TTG), respectively. In conclusion, the TabNet model yielded strong performance in predicting live births in RIF patients using a combination of 23 variables. This model can help improve understanding of the underlying mechanism of implantation failure and stratify patients who may benefit from immune modulation interventions.