Towards an explainable clinical decision support system for large-for-gestational-age births

Area of Science:

• Obstetrics and Gynecology
• Maternal-Fetal Medicine
• Biostatistics

Background:

• Delivery of large-for-gestational-age (LGA) infants is linked to numerous maternal and neonatal complications.
• LGA birth rates have risen globally, partly due to increasing maternal body mass index.
• Predictive models are needed to support clinical decision-making for LGA risk in overweight and obese pregnant individuals.

Purpose of the Study:

• To develop and validate probabilistic prediction models for LGA births.
• To create models specifically for clinical decision support in diverse populations.
• To enhance model explainability using Local Interpretable Model-agnostic Explanations (LIME).

Main Methods:

• Utilized data from 465 pregnant women with overweight/obesity from the PEARS study.
• Applied machine learning algorithms (Random Forest, SVM, AdaBoost, XGBoost) with Synthetic Minority Over-sampling Technique (SMOTE).
• Developed two models: one for white women (AUC-ROC 0.75) and one for all ethnicities (AUC-ROC 0.57); employed LIME for explainability.

Main Results:

• Identified key predictors: maternal age, mid-upper arm circumference, white cell count, fetal biometry, and gestational age at scan.
• Population-specific factors like Pobal HP deprivation index and fetal biometry centiles were also significant.
• Explainable models demonstrated effectiveness in predicting LGA probability.

Conclusions:

• Developed explainable predictive models for LGA births in overweight/obese women.
• Models can aid clinical decision-making and inform early pregnancy intervention strategies.
• These tools aim to reduce pregnancy complications associated with LGA infants.