Introduction Approximately 1–5% of pregnant women experience recurrent pregnancy loss (RPL). Early detection and evaluation of high-risk variables allow for the initiation of recommended treatments while also reducing the likelihood of RPL in these couples. Objective The goal of this study was to build a deep learning model to identify an immune-lifestyle pattern in RPL patients based on clinical and laboratory findings. Methods We retrospectively collected the data from 16,818 RPL patients and 19,979 healthy women from across five clinics throughout Iran from December 2014 to April 2024. Fundamental population size and laboratory symptoms were gathered from the all-available data of participants. Data preprocessing involved cleaning and partitioning the data into training and validation groups. We used 22 characteristics to identify patterns via a deep learning model (TabNet). The model performance was evaluated using a confusion matrix, precision-recall curve, calibration plot, and Receiver Operating Characteristics (ROC) curve. Results The model showed robust and practical performance in identifying an immunological pattern based on the selected variables. The model yielded an AUC of 0.985, an accuracy of 0.946, a precision of 0.936, a specificity of 0.921, and a sensitivity of 0.968. Over-fitting was mitigated via repeated 5-fold cross validation (CV); no feature leakage detected. Conclusion Patterns linking immune and lifestyle factors to pregnancy loss were identified with high reliability using a deep learning approach. These findings may support a deeper understanding of the biological mechanisms underlying RPL and help guide the development of targeted treatment strategies.
Dashti et al. (Thu,) studied this question.