Predictive Models of Pregnancy Based on Male Partner Data From a Preconception Cohort Study.

Background

Male factors contribute to approximately 50% of couple infertility, yet few studies have used machine learning with comprehensive male partner lifestyle and behavioral data to predict spontaneous conception.

Objectives

To develop and validate machine learning models predicting spontaneous conception using detailed self-reported male partner characteristics from a North American prospective preconception cohort.

Materials And Methods

We analyzed 2843 couples from the Pregnancy Study Online (PRESTO; 2013-2024). Using 185 male partner variables spanning sociodemographic, lifestyle, dietary, anthropometric, medical, and reproductive history domains (with female age and BMI included as covariates), we developed models for: (1) per-cycle probability of conception (Model I) using a discrete-time proportional hazards model; (2) pregnancy in fewer than 12 cycles (Model II); and (3) pregnancy within 6 cycles (Model III) using regularized logistic regression, gradient boosted trees, and random forests. We evaluated model performance using fivefold cross-validation. Age-stratified analyses (≥30 vs. <30 years) assessed heterogeneity in predictive factors.

Results

Among 2843 couples, 87.3% reported pregnancy in fewer than 12 cycles and 69.7% within 6 cycles. Model I achieved a concordance index of ∼0.59 (SD 0.012). The sparse regularized logistic regression model, selected via recursive feature elimination for interpretability, achieved ROC AUCs of 0.653 (SD 0.027) for Model II and 0.639 (SD 0.016) for Model III-comparable to the best full models using all candidate predictors (AUC 0.665 and 0.648, respectively). Prior impregnation history was the strongest predictor (Model II OR: 1.48, 95% CI: 1.34-1.63; Model III OR: 1.34, 95% CI: 1.21-1.47). Additional predictors included sugar-sweetened beverage consumption, environmental tobacco smoke exposure, gastroesophageal reflux disease, and physical activity. Male age emerged as a significant predictor among males ≥30 years, whereas female BMI, cycling ≥ 3 h/week, and tobacco smoke exposure were key predictors among younger males.

Discussion

AND

Conclusions

Machine learning models incorporating male partner characteristics demonstrated moderate discrimination for predicting pregnancy within 6-12 cycles. Predictive factors differed by male age, highlighting heterogeneity that may warrant further investigation. Modifiable factors, particularly sugar-sweetened beverage consumption, tobacco smoke exposure, and physical activity, were among the predictors identified and may represent candidate targets for future intervention research.