MULTI-FEATURE INTEGRATED INTELLIGENT DIAGNOSIS MODEL FOR NON-INVASIVE PRENATAL TESTING BASED ON ENSEMBLE LEARNING

Abstract

To address the challenges of high false-negative rates and low diagnostic reliability in current Non-Invasive Prenatal Testing (NIPT) for detecting chromosomal abnormalities in female fetuses, this study proposes an innovative multi-feature intelligent diagnosis model based on ensemble learning. The key innovations of this research include: (1) Multi-dimensional feature integration: For the first time, 16 critical features covering chromosome Z-values (13, 18, 21, X), GC content, sequencing metrics, and maternal physiological indicators were systematically integrated to comprehensively characterize fetal chromosomal status. (2) Advanced ensemble framework: We developed a novel hybrid ensemble approach combining Random Forest, XGBoost, and LightGBM algorithms through soft voting, effectively addressing data challenges of high dimensionality, small sample size, and severe class imbalance (only 10.7% abnormal samples). (3) Dual optimization strategy: The model was optimized using both SMOTE oversampling and random undersampling techniques for data balance, combined with grid search and five-fold cross-validation for parameter tuning. Experimental results demonstrate that our ensemble model achieved superior performance with 91.59% accuracy, 0.9583 AUC, 91.49% precision, 89.58% recall, and 90.53% F1-score, significantly outperforming single-algorithm models. Feature importance analysis revealed that BMI, chromosome 18 Z-values, and maternal age were the most influential predictors. This model provides a clinically applicable, highly accurate diagnostic tool that substantially improves the reliability of NIPT-based female fetal abnormality detection.