Constructing Prediction Models for Small for Gestational Age Based on Multimodal Clinical and Ultrasonographic Data
10.13471/j.cnki.j.sun.yat-sen.univ（med.sci）.20240617.001
- VernacularTitle:结合临床与超声影像多模态数据构建小于胎龄儿预测模型
- Author:
Xinyu CHEN
1
;
Yunxiao ZHU
1
Author Information
1. Department of Medical Ultrasonics， The Seventh Affiliated Hospital of Sun Yat-sen University， Shenzhen 518107， China
- Publication Type:Journal Article
- Keywords:
small for gestational age;
prediction model;
machine learning;
first-trimester;
second-trimester
- From:
Journal of Sun Yat-sen University(Medical Sciences)
2024;45(4):637-648
- CountryChina
- Language:Chinese
-
Abstract:
ObjectiveTo explore the predictive value of multimodal clinical and ultrasonographic data in first- and second-trimester for small for gestational age （SGA）， so as to build and internally validate SGA prediction models based on multiple machine learning algorithms. MethodsThis retrospective study enrolled 1，307 pregnant women with singleton pregnancies， diagnosed SGA according to INTERGROWTH-21st fetal growth criteria， and collected multimodal clinical data including general clinical information， biochemical test data， and prenatal ultrasound screening data. Extreme gradient boosting （XGBoost） algorithm was used to calculate the importance of variables. Seven machine learning algorithms were used to construct and internally verify the prediction models. The area under the receiver operating characteristic curve （AUC） was used as the main indicator to measure the prediction performance and used to compare predictive performance between models with the sensitivity at a 10% false positive rate. ResultsThe optimal prediction model built based on general clinical information and biochemical test data had an AUC of 0.70， 95%CI （0.609， 0.791） and a sensitivity of 0.38， 95%CI （0.236， 0.519）. The optimal prediction model based on prenatal ultrasound screening data was better than the former， with an AUC of 0.77， 95%CI （0.687， 0.858） and a sensitivity of 0.62， 95%CI （0.457， 0.743）. The two data sets were combined to form the multimodal clinical dataset， and the performance of the best prediction model was further improved with an AUC of 0.91， 95%CI （0.851， 0.972） and a sensitivity of 0.88， 95%CI （0.745， 0.947）， and the model calibration showed good goodness of fit. ConclusionBy using machine learning algorithms to fully explore the predictive value of different types of clinical data for SGA in first- and second-trimester， this study proves the absolute advantages of multimodal clinical data for SGA screening， and provides an accurate and effective reference for personalized management of pregnant women.
