异常检测
计算机科学
数据挖掘
逻辑回归
聚类分析
机器学习
采样(信号处理)
异常
个性化医疗
支持向量机
决策支持系统
匹配(统计)
人工智能
胎儿游离DNA
决策树
事件(粒子物理)
临床决策支持系统
统计分类
特征(语言学)
班级(哲学)
异常(物理)
功能(生物学)
集成学习
假阳性悖论
作者
Haoyu Liu,Anqi Feng,Zhao Congli
出处
期刊:
[Association for Computing Machinery]
日期:2026-03-27
卷期号:: 147-152
标识
DOI:10.1145/3819113.3819138
摘要
The accuracy of Non-invasive prenatal testing (NIPT) is highly dependent on fetal cell-free DNA (cffDNA) concentration, which is significantly influenced by maternal physiological heterogeneity, such as Body Mass Index (BMI). Traditional uniform testing windows often lead to detection failures or delayed interventions. This study proposes a data-driven collaborative framework for personalized NIPT timing optimization and fetal abnormality detection based on clinical data from 1,086 cases. First, a log-polynomial hybrid nonlinear regression model was established to quantify the systematic dilution effect of BMI on fetal fraction. Building upon this, we developed a dual-objective risk function that balances detection failure risks against delayed treatment costs. Utilizing K-means++ clustering and dynamic programming, we derived optimal testing trajectories tailored to distinct BMI cohorts. Results indicate that the optimal testing window shifts from 12.5 weeks for low-BMI groups to 16.0 weeks for high-BMI extreme groups to ensure concentration compliance. Furthermore, to address fetal anomaly classification under class imbalance, we integrated SMOTE-Tomek hybrid sampling with a triple ensemble learning model (Random Forest, Gradient Boosting, and Logistic Regression). The ensemble model achieved an AUC of 0.860 and a diagnostic specificity of 98.6%, significantly reducing unnecessary invasive procedures. This framework provides a systematic quantitative decision-making method for personalized prenatal screening and can be extended to other time-sensitive medical screening scenarios.
科研通智能强力驱动
Strongly Powered by AbleSci AI