Background: Early pregnancy loss is a common complication with significant implications for subsequent perinatal and neonatal health. Accurate first-trimester prediction remains challenging due to chromosomal, maternal, fetal, and environmental factors. Machine learning (ML) models integrating routinely collected clinical and biochemical data may improve risk stratification and guide personalized prenatal care. This study aimed to develop and evaluate ML models integrating first-trimester clinical, biochemical, and obstetric parameters for predicting early miscarriage in an Iranian population.

Methods: This retrospective study analyzed 2,500 first-trimester pregnancies (64 miscarriages, 2,436 ongoing). Features included maternal/paternal age, weight, pregnancy-associated plasma protein-A (PAPP-A), free β-human chorionic gonadotropin (free β-hCG), nuchal translucency multiples of the median (NT MoM), and obstetric history. Logistic regression (unregularized and Lasso/Ridge), support vector machine (SVM), XGBoost, random forest, gradient boosting, AdaBoost, and neural networks were compared using Monte Carlo simulation with 100 iterations and balanced subsampling (64 positive + 130 negative per iteration). Performance was assessed via accuracy, ROC-AUC, and PR-AUC.

Results:Support Vector Machine (SVM) achieved the highest performance (accuracy 0.62, F1-score 0.56, ROC-AUC 0.71, RMSE 0.44). Lower PAPP-A levels, higher maternal and paternal age, and higher maternal weight were significantly associated with miscarriage. Feature importance analysis highlighted paternal age, MoM PAPP-A, and first-trimester biomarkers as key predictors. Based on adjusted logistic regression, independent predictors of miscarriage were history of miscarriage (OR=2.85, p=0.01), increased NT (OR=5.52, p<0.001), and positive first-trimester Down syndrome screening (OR=5.71, p=0.02).

Conclusion: Machine learning models using routine first-trimester data showed only modest predictive performance for miscarriage in this setting. Severe class imbalance and the small number of miscarriage events constrain model generalizability. Larger multicenter prospective studies are needed before clinical deployment.