Research on NIPT Timing Selection and Fetal Abnormality Determination Based on Optimized Classification Models

Abstract

With the growing demand for prenatal screening, selecting appropriate timing for non -invasive prenatal testing (NIPT) and improving the accuracy of abnormality determination have become key issues in clinical practice and public health. This study, based on cl inical NIPT data from a specific region, constructs and compares a Beta regression-based proportional response model with a classification model based on Gradient Boosting Decision Trees (GBDT). The research process includes data cleaning and standardization, Shapiro – Wilk normality testing, Spearman correlation screening, and stepwise Beta regression modeling to quantify the relationship between Y chromosome concentration and factors such as gestational age and BMI. Subsequently, a multi -objective optimization classification model is designed for male fetus samples, using BMI grouping critical points and corresponding NIPT timing as decision variables, solved with fmincon to simultaneously minimize potential risks and maximize detection accuracy. This is then extended to a multi - factor comprehensive discrimination index incorporating age, height, weight, and other variables. Finally, a GBDT classifier is constructed for female fetuses, comparing three sample balancing strategies —random oversampling, random un dersampling, and SMOTE —and optimizing hyperparameters through grid search. The main results show that the proposed methods can significantly improve detection accuracy under different settings (the model achieves notably higher accuracy in several configurations, with female fetus classification reaching a high level under the optimal balancing strategy). The study provides actionable quantitative support for developing stratified NIPT timing and abnormality determination protocols in clinical settings. The research discusses limitations such as the single data source and the lack of dynamic modeling, and proposes future directions including expansion to multi -center data and prospective validation.