Hormonal and laboratory predictors of patent foramen ovale in cryptogenic ischemic events: a SHAP-enhanced logistic regression approach

Background: Patent foramen ovale (PFO) is associated with cryptogenic ischemic events. Detection relies on advanced imaging modalities, yet early detection tools remain limited. Methods: We developed a predictive model for PFO detection in cryptogenic stroke/TIA patients using clinical and laboratory data. This retrospective study included 300 female patients (18-65 years) with cryptogenic ischemic events (TOAST criteria) who underwent contrast-enhanced transcranial Doppler and echocardiography for PFO assessment. Forty-three variables were analyzed, with LASSO and mRMR feature selection identifying five key predictors: age, estradiol, follicle-stimulating hormone (FSH), D-dimer, and LDL cholesterol. Nine machine learning models were evaluated, with logistic regression selected as the final model. Performance was assessed via AUC, accuracy, sensitivity, specificity, F1 score, and decision curve analysis. SHAP analysis explained feature contributions. We also conducted a prespecified subgroup analysis by reproductive stage (reproductive-age 22-45 years vs. menopausal 45-65 years) to account for stage-related hormonal differences. Results: < 0.05), and the model maintained strong discrimination in both groups (AUC 0.988 in reproductive-age; 0.982 in menopausal). Conclusion: An interpretable, high-performing model can aid early PFO risk stratification in females; performance was robust across reproductive-age and menopausal strata, and external validation especially in larger female cohorts is warranted. Our model does not assess causality but aims to help clinicians identify those who are most likely to have a PFO, which can then be further investigated to determine the causal role of PFO in the ischemic event.