An immune‐inflammatory and thyroid function machine learning model for predicting adverse pregnancy outcomes: Development in a retrospective cohort and external validation in a prospective cohort

OBJECTIVE: Maternal thyroid dysfunction is associated with adverse pregnancy outcomes, yet thyroid status alone stratifies individual risk poorly and the immune-inflammatory contribution is rarely incorporated into prospectively validated prediction tools. We aimed to develop and externally validate a model predicting a composite adverse pregnancy outcome from antenatal thyroid and cytokine markers. METHODS: The model was developed in a retrospective cohort (n = 650; 131 events, 20.2%) and externally validated in an independent prospective cohort (n = 150; 26 events, 17.3%). Fifteen antenatal candidate predictors common to both cohorts were considered. Least absolute shrinkage and selection operator (LASSO) regression was used for variable selection; LASSO-logistic regression, random forest, extreme gradient boosting, and a support-vector machine were compared by repeated stratified cross-validation. Discrimination was quantified using the area under the receiver operating characteristic curve (AUC); calibration and clinical utility were also assessed, with internal-external cross-validation (IECV) as a sensitivity analysis. RESULTS: LASSO retained three predictors: thyroid peroxidase antibody (TPOAb), interleukin-6 (IL-6), and interleukin-10 (IL-10). LASSO-logistic regression was selected (development cross-validated AUC 0.762, 95% confidence interval [CI] 0.692-0.846). On external validation, the model achieved an AUC of 0.750 (95% CI 0.647-0.848), exceeding a thyroid group-only reference model (AUC 0.625); the Brier score was 0.127, and the calibration slope was 1.77. Random 10-fold IECV gave a pooled AUC of 0.761 (95% CI 0.720-0.800). Higher IL-6 was associated with increased risk and higher IL-10 with decreased risk. CONCLUSION: A parsimonious three-marker model combining TPOAb, IL-6, and IL-10 showed moderate discrimination for adverse pregnancy outcomes and outperformed thyroid status alone. Calibration requires refinement, and validation in larger multicenter cohorts is warranted before clinical implementation.