Multi-Task Machine Learning for Prenatal Risk Stratification: Integrating Biomarkers, Maternal Age, and Ultrasound Measurements to Predict the Risk of Down Syndrome, Trisomy 18, Trisomy 13, and Neural Tube Defects

This study developed a machine learning model for early risk stratification of Down syndrome by integrating maternal serum biomarkers and ultrasound measurements. A retrospective multicentre dataset was used, including maternal age, AFP, HCG, INHIBIN-A, and ultrasound parameters (NT, CRL). After imputing missing data and engineering features (e. g. , AgeNTᵢnteraction), a Gradient Boosting Machine (GBM) was trained and evaluated using AUROC, precision, recall, and F1-score. The model achieved high performance (AUROC: 0. 9921; precision: 1. 00; F1-score: 0. 91; accuracy: 0. 97). SHAP analysis identified key interactions—particularly AgeNT, AgeHCG, and AgePAPP-A—as major contributors. High maternal age combined with elevated HCG or low PAPP-A was linked to increased risk, aligning with clinical knowledge. The model offers a highly accurate and interpretable approach for Down syndrome risk prediction, supporting personalized, data-driven prenatal care. Prospective validation and clinical integration are recommended.