Interpretable machine learning model for predicting low birth weight in singleton pregnancies: a retrospective cohort...

Area of Science:

• Obstetrics and Gynecology
• Neonatal Health
• Data Science in Medicine

Background:

• Low birth weight (LBW) is a significant global public health issue.
• Identifying risk and protective factors for LBW is crucial for intervention.
• This study focuses on predicting LBW in singleton pregnancies.

Purpose of the Study:

• To utilize predictive models for identifying critical factors associated with LBW.
• To compare the performance of logistic regression and machine learning algorithms in LBW prediction.
• To provide insights for clinical decision-making and public health strategies.

Main Methods:

• A retrospective cohort study of 10,227 singleton pregnancies was conducted.
• Logistic Regression (LR) and four machine learning (ML) models (Random Forest, LightGBM, SVM, XGBoost) were employed.
• SHAP analysis was used for feature importance interpretation in ML models.

Main Results:

• The XGBoost model showed superior performance in LBW prediction (AUC 0.741 on test set).
• Key factors identified by both LR and XGBoost include maternal age, gestational age, BMI, hypertensive disorders of pregnancy (HDP), and fetal distress.
• LBW infants face increased risks of hospitalization and conditions like congenital anomalies and respiratory distress syndrome (NRDS).

Conclusions:

• LR and XGBoost models effectively predict LBW and identify associated risk factors.
• Pregnant women with early gestational age (<37 weeks), low BMI (<18 kg/m²), younger maternal age (<25 years), HDP, or fetal distress are at higher risk.
• These findings can inform targeted interventions to reduce LBW rates and improve neonatal outcomes.