1 of 23

Heavy Metals

Exposure and Diabetes

Risk Prediction using Machine Learning Approach

Sagar Shrestha, Felix Twum, Jennifer L. Lemacks, Sermin Aras

April 3^rd, 2025

Susan A. Siltanen Graduate Student Research Symposium

The University of Southern Mississippi

2 of 23

Introduction

Diabetes is a major global health concern
38.4 million people or 11.6% of the U.S. population have diabetes(CDC,2024)
Total Annual Cost of Diabetes (ADA, 2023)

$412.9 billion total cost in 2022

$306.6 billion in direct medical costs
$106.3 billion in indirect costs�

Diabetes mellitus is a growing global health concern, affecting millions worldwide. While traditional risk factors such as genetics, diet, and lifestyle are well established, emerging research suggests that environmental factors, particularly heavy metal exposure, may also play a crucial role in diabetes pathogenesis.

The cost burden disproportionately affects vulnerable and underserved communities (American Diabetes Association, 2023).
Direct medical costs of diabetes have risen by 35% since 2012 (adjusted for inflation) (American Diabetes Association, 2023).

Rise in Direct Medical Costs: Direct medical costs of diabetes have risen by 35% since 2012 (adjusted for inflation) (American Diabetes Association, 2023).
Indirect Costs and Productivity Losses: Indirect costs include substantial productivity losses:

$35.8 billion due to presenteeism
$28.3 billion from reduced employment due to disability
$32.4 billion in lost productivity from premature deaths (American Diabetes Association, 2023).

Annual Medical Expenditures: People with diabetes incur average annual medical expenditures of $19,736, which is 2.6 times higher than what would be expected without diabetes (American Diabetes Association, 2023).

In 2021, the International Diabetes Federation (IDF) reported that 537 million people around the world had diabetes, and they projected that 783 million people will have diabetes by 2045 [1].

3 of 23

Risk factors

Traditional Risk

Genetics
Diet
Lifestyle

Environmental factors

Heavy Metal exposure

4 of 23

Motivation of the Study

Explore Environmental influence
Early Detection
Reduce Healthcare Costs
Improve Quality of Life

5 of 23

Formative Study - OLS Regression Analysis

Manganese: Negative association with HbA1c (β = -0.0125, p < 0.001). Potential protective effect.
Selenium: Positive association with HbA1c (β = 0.0021, p < 0.001). Possible link to increased risk.
Lead: Marginal association with HbA1c (p = 0.075).
Cadmium & Mercury: Not significant predictors of HbA1c.

6 of 23

Objectives

Investigate the association between heavy metal exposure and diabetes
Determine the most effective ML model for predicting diabetes risk

7 of 23

Literature review

Not many studies on heavy metal exposure predicting diabetes risk using machine learning tools
Use of dietary copper (Cu), urinary cadmium (Cd) and urinary mercury (Hg) for improved prediction accuracy (Zhao et al., 2025).
Urinary antimony as a significant predictor of diabetic retinopathy risk (Gui et al.,2024).

8 of 23

Data Preprocessing

9 of 23

Dataset

Data Source: NHANES 2011-2018, initially consisting of 51,122 participants.

Dataset Size: 13667 survey responses

Original NHANES Data: 51,122 responses and 330 variables

10 of 23

Dataset Histogram

11 of 23

Dataset Histogram

12 of 23

Data Filtering

Data Imbalance:

87.38% no-diabetes, 12.62% people with diabetes.
Imbalance can pose challenges in modeling.

Feature Standardization:

Standardized nonbinary columns.
Ensures fair comparison and accurate analysis.

13 of 23

Key Variables

Demographics �(age, sex, BMI, etc.)
Blood heavy metal concentrations �(Lead, Mercury, Cadmium, Manganese)
Selenium in blood
Weight-related variables �(weight, height, weight in last year, etc )

14 of 23

Downsampling and Stratified K fold Cross Validation

Downsampling to prevent the model from being biased towards the majority class.
Satisfied K fold to divide the data into K folds.
Each fold maintains the same proportion of classes as the original dataset.
Provides a more accurate and unbiased evaluation of the model's ability to perform on all classes.

15 of 23

Principal component Analysis

Too many features in our weight-related data
Increase complexity, overfitting, and computational issue
PCA to reduce the number of features by transforming them into a smaller set of principal components
Understand the core factors influencing weight, streamline analysis

Too many features in our weight-related data (e.g., measurements, habits) lead to complexity, overfitting, and computational issues.

(Solution) Principal Component Analysis (PCA): A technique to reduce the number of features by transforming them into a smaller set of principal components that capture most of the data's variance.

(Benefits of PCA for Weight Data)

Reduces Dimensionality: Simplifies the data by focusing on the most important information.
Improves Analysis: Makes it easier to identify key patterns and relationships.
Potentially Enhances Modeling: Can reduce overfitting and improve model efficiency.
Aids Visualization: Allows for easier visualization of complex data.

(Impact) PCA helps us understand the core factors influencing weight, streamline analysis, and build better models despite having numerous initial features.

16 of 23

ML Models

CatBoost

LightGBM

Random Forest

Feedforward Neural Network (FNN)

17 of 23

Results for Risk Assessment

Metric	Random Forest	CatBoost	LightGBM	FNN
Accuracy	0.7166	0.7285	0.7163	0.668
Precision	0.6999	0.707	0.6985	0.6539
Recall	0.7596	0.7811	0.7626	0.7159
F1	0.7283	0.7421	0.729	0.6832
F2	0.7467	0.765	0.7487	0.7024

18 of 23

Feature importance for Cat Boost

Lead in blood (8.254) “Most significant heavy metal predictor”
Selenium in blood (5.688)
Cadmium in blood (4.842)
Mercury in blood (4.759)
Manganese in blood (4.232)

19 of 23

Feature importance for Cat Boost

20 of 23

Conclusions

Heavy metal exposure, particularly lead, is strongly associated with diabetes risk.

CatBoost is the most effective model for predicting diabetes based on environmental and physiological factors.

These findings underscore the need for stricter environmental regulations and further research into heavy metal toxicity and diseases.

1

2

3

21 of 23

Future Work

Exploring additional environmental and genetic factors

Implementing deep learning models to improve predictive accuracy.

Conducting longitudinal studies to establish causal relationships.

1

2

3

22 of 23

References

“A report card: Diabetes in the United States infographic,” Diabetes, May 15, 2024. https://www.cdc.gov/diabetes/communication-resources/diabetes-statistics.html

American Diabetes Association. (2023). Annual Report 2023. Retrieved from https://diabetes.org/sites/default/files/2024-06/ADA_2023_AnnualReport.pdf

Zhao, M., Wan, J., Qin, W., Huang, X., Chen, G., & Zhao, X. (2023). A machine learning-based diagnosis modeling of type 2 diabetes mellitus with environmental metal exposure. Computer Methods and Programs in Biomedicine, 107537. https://doi.org/10.1016/j.cmpb.2023.107537

Gui, Y., Gui, S., Wang, X., Li, Y., Xu, Y., & Zhang, J. (2024). Exploring the relationship between heavy metals and diabetic retinopathy: a machine learning modeling approach. Scientific Reports, 14, 13049.

Centers for Disease Control and Prevention (CDC), National Center for Health Statistics (NCHS). National Health and Nutrition Examination Survey (NHANES), 2011-2018. Hyattsville, MD: U.S. Department of Health and Human Services, Centers for Disease Control and Prevention. Available from: https://www.cdc.gov/nchs/nhanes/index.htm

23 of 23

Thanks