Temporal Trends and Machine Learning-Based Risk Prediction of Female Infertility: A Cross-Cohort Analysis Using NHANES Data (2015-2023).

Background

Female infertility represents a significant global public health concern, yet its evolving trends and data-driven risk prediction remain under examined in nationally representative cohorts. This study investigates temporal changes in infertility prevalence and evaluates Machine Learning (ML) models for infertility risk prediction using harmonized clinical features from NHANES cycles (2015, 2016, 2017, 2018, 2021, 2022, and 2023).

Methods

Women aged 19 to 45 with complete data on infertility-related variables (including reproductive history, menstrual irregularity, Pelvic Infection Disease (PID), hysterectomy, and bilateral oophorectomy) were analyzed. Descriptive statistics and cohort comparisons employed ANOVA and Chi-square tests, while multivariate Logistic Regression (LR) estimated Adjusted Odds Ratios (OR) and informed feature importance. Predictive models (LR, Random Forest, XGBoost, Naive Bayes, SVM, and a Stacking Classifier ensemble) were trained and tuned via GridSearchCV with five-fold cross-validation. Model performance was evaluated using accuracy, precision, recall, F1-score, specificity, and AUC-ROC.

Results

We observed a notable increase in infertility prevalence from 14.8% in 2017-2018 to 27.8% in 2021-2023, suggesting potential post-pandemic impacts on reproductive health. In multivariate analysis, prior childbirth emerged as the strongest protective factor (Adjusted OR ≈0.00), while menstrual irregularity showed a significant positive association with infertility (OR =0.55, 95% CI 0.40 to 0.77, p<0.001). Unexpectedly, PID, hysterectomy, and bilateral oophorectomy were not significantly associated with infertility after adjustment (p>0.05), which may partly reflect the inherent definition of self-reported infertility used in this study. All six ML models demonstrated excellent and comparable predictive ability (AUC >0.96), reinforcing the effectiveness of even a minimal common predictor set for infertility risk stratification.

Conclusions

The rising prevalence of self-reported infertility among U.S. women underscores emerging public health challenges. Despite relying on a streamlined feature set, interpretable and ensemble ML models successfully predicted infertility risk, showcasing their potential applicability in broader surveillance and personalized care strategies. Future models should integrate additional sociodemographic and behavioral factors to enhance precision and support tailored interventions.