Machine learning and explainable artificial intelligence to predict and interpret lead toxicity in pregnant women and unborn baby

Abstract

Introduction: Lead toxicity is a well-recognised environmental health issue, with prenatal exposure posing significant risks to infants. One major pathway of exposure to infants is maternal lead transfer during pregnancy. Therefore, accurately characterising maternal lead levels is critical for enabling targeted and personalised healthcare interventions. Current detection methods for lead poisoning are based on laboratory blood tests, which are not feasible for the screening of a wide population due to cost, accessibility, and logistical constraints. To address this limitation, our previous research proposed a novel machine learning (ML)-based model that predicts lead exposure levels in pregnant women using sociodemographic data alone. However, for such predictive models to gain broader acceptance, especially in clinical and public health settings, transparency and interpretability are essential.

Methods: Understanding the reasoning behind the predictions of the model is crucial to building trust and facilitating informed decision-making. In this study, we present the first application of an explainable artificial intelligence (XAI) framework to interpret predictions made by our ML-based lead exposure model.

Results: Using a dataset of 200 blood samples and 12 sociodemographic features, a Random Forest classifier was trained, achieving an accuracy of 84.52%.

Discussion: We applied two widely used XAI methods, SHAP (SHapley additive explanations) and LIME (Local Interpretable Model-Agnostic Explanations), to provide insight into how each input feature contributed to the model's predictions.

Keywords: classification; explainable AI; lead toxicity; machine learning; predictive modelling.

Figure 1

Influence diagram showing features that impact maternal blood lead levels. Reproduced with permission from “Influence diagram of features impacting maternal BLL and toxicity exposure” by Priyanka Chaurasia, Sally I. McClean, Abbas Ali Mahdi, Pratheepan Yogarajah, Jamal Akhtar Ansari, Shipra Kunwar and Mohammad Kaleem Ahmad, licensed under CC BY 4.0.

Figure 2

XAI framework for automated lead toxicity prediction model.

Figure 3

Summary bar plot showing important variables, indicating the magnitude of each feature in the prediction of the class labels in all the instances of the test data.

Figure 4

Beeswarm summary plots of representative SHAP values for the 12 features in our model, from most significant to the least significant one (top to bottom) for each of the class: ND_5, Between5_10, Between10_15 and GreaterThan15.

Figure 5

Example of the output class ND_5: (a) Interpretation of model prediction results based on SHAP, (b) Explanation provided by the LIME model.

Figure 6

Example of SHAP decision plot for the output class ND_5.

Figure 7

Example of the output class Between5_10: (a) Interpretation of model prediction results based on SHAP, (b) Explanation provided by the LIME model.

Figure 8

Example of SHAP decision plot for the output class Between5_10.

Figure 9

Example of the output class Between10_15: (a) Interpretation of model prediction results based on SHAP, (b) Explanation provided by the LIME model.

Figure 10

Example of SHAP decision plot for the output class Between10_15.

Figure 11

Example of the output class GreaterThan15: (a) Interpretation of model prediction results based on SHAP, (b) Explanation provided by the LIME model.

Figure 12

Example of SHAP decision plot for the output class “GreaterThan15.”

Figure 13

Web interface showing results for an input instance for which the predicted class is very high level (GreaterThan15) and the explanation of the outcome using the force and decision plots.