Prediction of birthweight with early and mid-pregnancy antenatal markers utilising machine learning and explainable artificial intelligence

Abstract

Low birthweight (LBW) is a significant health challenge worldwide, as these neonates experience both short- and long-term disabilities. Factors affecting maternal and fetal health during early to mid-pregnancy can greatly influence fetal development. Prediction of birthweight using machine learning (ML) models with antenatal data may help in better clinical management. However, the lack of explainability in these models has raised concerns within the medical community. To address this issue, our study aims to develop a more practical ML model by incorporating explainable artificial intelligence (XAI). We prospectively collected real-world clinical data of 19 maternal and fetal clinical features from 237 singleton pregnancies. Statistical analyses were conducted using Jamovi (version: 2.6.26) and JASP team (2024) JASP (version: 0.18.3). Multiple ML classifiers were employed. We developed a stacked ensemble model that integrated various algorithms, including a custom-stacked ensemble approach and three XAI methodologies: Shapley Additive Explanations (SHAP), Local Interpretable Model-agnostic Explanations (LIME), and Anchor. These methods provided meaningful explanations to help construct reliable and optimal clinical predictive models. Among the ML classifiers evaluated, the AdaBoost model achieved the highest performance, with a maximum accuracy of 77%, a precision of 73%, a recall of 77%, and an F1 score of 72%. The stacked model demonstrated an accuracy of 75%, indicating its possibility in clinical application. However, the accuracy of these models might be affected by the limited dataset, which included pregnant women undergoing treatment for thyroid abnormalities, diabetes, and hypertension. Our developed model identified several key attributes that influence birthweight, such as maternal height, nuchal translucency thickness, parity, crown-rump length, glycated hemoglobin, hypertensive disorders of pregnancy, and pregnancy-associated plasma protein A. This model can assist medical professionals in making more precise birthweight predictions using routinely collected antenatal parameters, enabling timely medical decisions and treatments.

Keywords: Antenatal care; Explainable artificial intelligence; Low birthweight; Machine learning.

Fig. 1

Illustration of etiology, complications, prevention, treatment, and management of LBW.

Fig. 2

Shows the interdisciplinary approach used to predict birthweight.

Fig. 3

The flowchart illustrates the participant selection process for the study and its outcomes.

Fig. 4

Representation of the characteristics of important attributes. The distributions of continuous variables are represented in the histogram, and categorical variables such as hypertension, gestational diabetes mellitus, and birthweight are represented in the pie diagram.

Fig. 5

Pearson’s matrix represents the correlation between birthweight and different attributes. Maternal height, CRL, NT, PAPP-A, and HTN are correlated with birthweight (BW).

Fig. 6

Representation of mutual information. It represents the importance of different variables according to the importance in descending order. NT, HC, HbA1c, age, and AC are considered important features and occupy the initial portion from left to right.

Fig. 7

Eight different types of ML classifiers applied in the study and a stacked model is developed.

Fig. 8

The flow chart represents the study procedure from ethical clearance to the study outcome in the development of birthweight prediction model.

Fig. 9

Confusion matrix of multi class obtained for the stacked model.

Fig. 10

SHAP plot indicating different attributes arranged in descending order. Important attributes such as height, parity, HTN, NT, HbA1C occupy the initial levels.

Fig. 11

Model explainability using LIME. Parity and height influence the outcome.