Machine learning approach for the prediction of macrosomia

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Fetal macrosomia poses risks for maternal and newborn health due to inaccurate weight estimation.
• Traditional methods struggle with precise fetal measurements in late pregnancy.
• Early detection of macrosomia is crucial for optimizing delivery strategies and reducing complications.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting fetal macrosomia.
• To compare the efficacy of various ML algorithms against traditional methods.
• To identify key features for accurate macrosomia prediction in the third trimester.

Main Methods:

• Trained and validated six ML algorithms (logistic regression, k-NN, SVM, RF, XGBoost, LightGBM) on data from 3244 pregnant women.
• Utilized the information gain method to identify significant predictive features.
• Conducted experiments using ultrasound data from 1-7 and 8-14 days before delivery.
• Developed an ensemble model combining RF, XGBoost, and LightGBM.

Main Results:

• The ensemble ML model demonstrated superior performance in predicting macrosomia compared to individual algorithms and the Hadlock formula.
• The model achieved high accuracy based on recall and area under the curve metrics.
• Key risk-relevant features were identified, enhancing predictive power.

Conclusions:

• Machine learning algorithms, particularly the proposed ensemble model, offer a promising approach for accurate fetal macrosomia prediction.
• These ML models can significantly aid obstetricians in selecting appropriate delivery methods.
• Improved prediction of fetal macrosomia can lead to better maternal and neonatal outcomes.