P-586 Development and validation of an interpretable machine learning model for predicting pregnancy outcomes based on mitochondrial DNA (mtDNA) content and clinical features: a retrospective study

Abstract

Abstract Study question Can machine learning models be developed and validated based on mtDNA content and other relevant clinical characteristics to predict the probability of successful embryo implantation? Summary answer On the basis of large-sample data, we developed and validated an explainable machine learning (ML) model to predict the probability of successful embryo implantation. What is known already Currently, Preimplantation Genetic Testing (PGT) has been extensively applied in assisted reproductive cycles worldwide. Nevertheless, despite its use, 25% to 50% of euploid blastocysts still fail to result in successful implantation during euploid blastocyst transfers. Some researchers suggest that the mtDNA content can be a biomarker for embryo viability. However, in previous studies exploring the correlation between mtDNA content and clinical outcomes, conventional data analysis methods have struggled to establish valuable predictive models. Study design, size, duration This is a retrospective study. Information on a total of 11,672 embryos was collected. Eventually, 2,012 embryos with clinical pregnancy and delivery outcomes were included in the study and randomly divided into a training set of 1,410 embryos and a test set of 602 embryos. Data collection spanned from 2016 to 2024. Participants/materials, setting, methods Clinical characteristics and embryo mtDNA content data of PGT patients admitted from 2016 to 2024 at Jiangsu Province Hospital were collected. Eight ML methods including Neural Network (NNET) were used to build models. The performance of the models was compared using evaluation indicators such as the area under the receiver operating characteristic curve (AUC), and the SHapley Additive exPlanation (SHAP) method was used to rank the feature importance and explain the final model. Main results and the role of chance Among eight ML models [Generalized Linear Model (GLM), Multilayer Perceptron (MLP), Neural Network (NNET), eXtreme Gradient Boosting (XGB), Naive Bayes (NB), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF)], NNET and SVM models demonstrated the best performance. In the training set [AUC: 0.728, 95% confidence interval (CI) (0.702 - 0.754); AUC: 0.773, 95% CI (0.749 - 0.797)] and the test set [AUC: 0.823, 95% CI (0.790 - 0.855); AUC: 0.818, 95% CI (0.785 - 0.852)], both the NNET and SVM models accurately predicted the pregnancy outcomes of patients undergoing PGT. Furthermore, we discovered a significant non-linear correlation between the mtDNA content and the pregnancy outcome. Notably, when the mtDNA content exceeded 0.182, the probability of successful delivery started to decline. The SHAP analysis identified the variables that contributed to the model’s prediction. The final model incorporated 16 variables: mtDNA content, anti - Müllerian hormone, female age, female body mass index (BMI), male age, male BMI, history of adverse pregnancy outcomes, chromosomal abnormalities, history of thyroid diseases, disturbance of ovulation, history of endometriosis, history of autoimmune diseases, history of endocrine and metabolic diseases, gonadotropin total dose, developmental days at the time of biopsy, and the number of sinus follicles. Limitations, reasons for caution This study is a single-center retrospective study. The generalizability of the model needs to be verified in future multi-center, prospective studies. Wider implications of the findings To our knowledge, this is the first study combining PGT’s mtDNA content with ML to build a clinical model. It clarifies the mtDNA-pregnancy outcome correlation in in vitro fertilization and offers insights for future clinical analyses of similar data. Trial registration number No

Ähnliche Arbeiten