A Multi-Model Fusion and Risk Optimization Approach for Non-Invasive Prenatal Testing Based on Random Forest and K-means Clustering

Abstract

This paper focuses on the complex relationship between Y-chromosome concentration and multiple factors in non-invasive prenatal testing, and constructs a methodological framework that integrates multiple models. First, to address the significant nonlinear characteristics among variables, a random forest model is introduced to assess feature importance, identify key influencing factors, and enhance model robustness. Building on this foundation, segmented polynomial regression is employed to effectively improve overall fitting performance through interval partitioning and local modeling. Furthermore, the concentration threshold is transformed into a risk minimization problem to construct an optimization model. K-means clustering is utilized to achieve joint optimization of grouping and testing timing, thereby reducing overall risk while ensuring structural stability. Additionally, to address class imbalance, SMOTE is introduced for data augmentation, and a classification model is established using Gradient Boosted Decision Trees (GBDT) to effectively identify abnormal samples. Overall, this method integrates Random Forests, piecewise regression, clustering, and ensemble learning models, demonstrating strong generalization capabilities and stability, and holds significant practical value in complex data modeling and risk control.