Development and evaluation of clinical pregnancy prediction models for intrauterine insemination using three machine...

Predicting clinical pregnancy following intrauterine insemination (IUI) remains challenging because of the complex interplay of multiple biological and treatment-related factors. Machine learning approaches may improve predictive accuracy compared with conventional statistical methods. This study aimed to compare the predictive performance of logistic regression (LR), random forest (RF), and multilayer perceptron (MLP) models and to explore their clinical interpretability.

A total of 957 IUI cycles were randomly divided into a training set (80%) and an independent validation set (20%). Synthetic Minority Oversampling Technique (SMOTE) was applied to the training set to address class imbalance. Model performance was evaluated using 10-fold cross-validation, and discrimination was assessed by area under the receiver operating characteristic curve (AUC). Model interpretability was examined using SHAP (Shapley Additive Explanations) analysis.

Across 10-fold cross-validation, all three models demonstrated modest and comparable discriminative ability. The MLP achieved a mean AUC of 0.547 (cross-validation mean AUC 0.560 ± 0.062), while RF and LR yielded mean AUC values of 0.549 and 0.541, respectively. Although overall discrimination was limited, the MLP showed relatively greater stability across folds. SHAP analysis consistently identified female age, infertility etiology, body weight, antral follicle count, estradiol, and treatment regimen as important predictors across models.

The MLP model showed modest discriminative ability in predicting IUI clinical pregnancy and identified several relevant clinical features. While its predictive performance is limited, the model may contribute to risk stratification and exploratory individualized assessment in assisted reproduction.