A Multivariate Statistical Modeling Approach for Non-Invasive Prenatal Testing

Abstract

The accuracy of non-invasive prenatal testing (NIPT) is significantly influenced by the fetal fraction and maternal physiological indicators. This study aims to enhance the precision of fetal health assessment by constructing multivariate statistical and machine learning models to optimize clinical decision-making. First, to analyze the correlation between maternal factors and Y-chromosome concentration, One-way ANOVA and Ordinary Least Squares (OLS) regression were employed; results indicated that gestational age and BMI are key predictors, with Random Forest Partial Dependence Plots (PDP) utilized to interpret non-linear relationships. Second, to address BMI-based stratification and timing, the Jenks Natural Breaks method was applied to divide BMI into eight groups, identifying an optimal testing window of 11–13 weeks, validated through Monte Carlo simulations to minimize detection risks. Third, a multi-factor optimization was achieved using an improved K-means clustering algorithm integrating height, weight, and age, which identified five refined BMI categories with an optimal testing point around 14 weeks. This model quantifies the contribution of technical errors through a weighted risk assessment. In conclusion, this integrated framework shifts NIPT from single-factor analysis to multi-dimensional optimization. Future work will incorporate multi-center data and Long Short-Term Memory (LSTM) networks to predict dynamic trends in fetal DNA concentrations.