Adverse pregnancy outcomes (APOs), including preeclampsia, preterm birth, and maternal heart failure, pose substantial risks for women with pre-existing heart disease. Soluble ST2 (sST2), a biomarker of cardiac stress, is associated with adverse cardiovascular events, yet its predictive role in high-risk pregnancies remains poorly defined. We implemented a sequential, multi-layer workflow integrating exploratory analysis, multivariable regression, and advanced machine learning. Layer 1 involved exploratory and univariate screening of 1,528 patient-level clinical records to identify potential predictors of APOs. sST2, NT-proBNP, and functional cardiac measures showed strong discriminative patterns, guiding downstream modeling. Layer 2 applied ridge logistic regression with bootstrap resampling to handle variable sparsity and perfect separation. sST2 emerged as the most stable and influential predictor (OR = 6.82), followed by heart rate and NT-proBNP, confirming biological and clinical relevance. Layer 3 incorporated synthetic data augmentation via CTGAN to address class imbalance and developed the Stable Dendritic Neural Model (SDNM), integrating clinical, demographic, and biomarker features. SDNM outperformed conventional machine learning models—including decision trees, random forests, gradient boosting, XGBoost, and LightGBM—achieving accuracy 0.993, F1-score 0.993, Cohen’s Kappa 0.986, MCC 0.988, and AUC 0.997, with excellent calibration (lowest Brier Score, ECE, and MCE) and minimal variance across repeated experiments. Interpretability analyses (SHAP and LIME) confirmed sST2, low birthweight, and preterm birth as dominant contributors to APO risk, supporting clinically actionable insights. Our findings establish sST2 as a robust early-risk biomarker for APOs and demonstrate the utility of a multi-layer, biologically informed machine learning framework for high-dimensional clinical risk prediction, bridging observational insights with precise, interpretable predictive modeling.