P-394 Using machine learning to predict the risk of miscarriage in infertile couples undergoing assisted reproductive cycles

• Published in: Human reproduction (Oxford) Vol. 39; no. Supplement_1
• Main Authors: Zare, M, Hashemzadeh, M, Tabibnejad, N
• Format: Journal Article
• Language: English
• Published: 03.07.2024
• ISSN: 0268-1161; 1460-2350
• DOI: 10.1093/humrep/deae108.749

Abstract Study question Can an artificial intelligence (AI)-based model predict chemical abortion using basic clinical data of infertile couples undergoing assisted reproductive technology (ART)? Summary answer We have combined basic clinical data and machine learning algorithms to create a noninvasive AI model for prediction of chemical abortion. What is known already Approximately 10 to 15 percent of natural pregnancies encounter the issue of abortion, it significantly gets worse in pregnancies resulting from assisted reproductive methods. The emotional burden on infertile couples during their self-belief therapy journey is substantial. Predicting treatment outcomes can significantly decrease the anxiety experienced by these couples. This research aims to develop a machine learning-based model for predicting chemical abortion in infertile women who have conceived through assisted reproductive methods, offering an initial assessment. Study design, size, duration This study involved analysis of retrospectively collected records including basic clinical data and pregnancy outcomes of 1234 records between March 2020 and August 2022. It designs in two fundamental phases: data extraction and model implementation. The extracted data from the study group was evaluated with registered physical and electronic records to validate and correct the information. Participants/materials, setting, methods This study was conducted at Yazd Reproductive Sciences Institute, Yazd, Iran. Data pre-processing was performed for data mining and model implementation. This process yielded a dataset of 1234 samples containing 32 selected variables. Machine learning algorithms, including decision tree, random forest, logistic regression, and support vector machine, were employed for models’ assessment. Regarding the imbalance nature of the gathered dataset, balancing techniques were performed to enhance model performance. Balancing was applied using SMOTE, ADYNS algorithms. Main results and the role of chance Evaluation of the above-mentioned models across unbalanced, balanced, and reduced datasets revealed the random forest model as the most robust choice, achieving accuracies of 79.80%, 88.90%, and 89.90% in the respective test phases. In terms of balancing process, ADYNS was superior to SMOTE. In addition, this research identified 18 influential variables for predicting abortion. Limitations, reasons for caution This investigation was performed using retrospective data; some incomplete records were excluded. In addition, the data heterogenicity may affect the analysis. The model developed is limited to the analysis of basic clinical data, further modification of the model is needed to incorporate information from different stages of the ART cycle. Wider implications of the findings This study demonstrated a model to predict chemical miscarriage. The high accuracy of this model showed that this model could be implemented in ART clinics as a primary tool to assist clinicians in making an accurate initial judgment on the condition of patients with infertility and history of pregnancy loss. Trial registration number not applicable