Novel Application of Machine Learning Algorithms and Model-Agnostic Methods to Identify Factors Influencing Childhood Blood Lead Levels

House Sparrows as Sentinels of Childhood Lead Exposure

Our understanding of connections between human and animal health has advanced substantially since the canary was introduced as a sentinel of toxic conditions in coal mines. Nonetheless, the development of wildlife sentinels for monitoring human exposure to toxins has been limited. Here, we capitalized on a three-decade long child blood lead monitoring program to demonstrate that the globally ubiquitous and human commensal house sparrow (Passer domesticus) can be used as a sentinel of human health risks in urban environments impacted by lead mining. We showed that sparrows are a viable proxy for the measurement of blood lead levels in children at a neighborhood scale (0.28 km2). In support of the generalizability of this approach, the blood lead relationship established in our focal mining city enabled us to accurately predict elevated blood lead levels in children from another mining city using only sparrows from the second location. Using lead concentrations and lead isotopic compositions from environmental and biological matrices, we identified shared sources and pathways of lead exposure in sparrows and children, with strong links to contamination from local mining emissions. Our findings showed how human commensal species can be used to identify and predict human health risks over time and space.

Using machine learning algorithms and techniques for defining the impact of affective temperament types, content search and activities on the internet on the development of problematic internet use in adolescents' population

Background

By using algorithms and Machine Learning - ML techniques, the aim of this research was to determine the impact of the following factors on the development of Problematic Internet Use (PIU): sociodemographic factors, the intensity of using the Internet, different contents accessed on the Internet by adolescents, adolescents' online activities, life habits and different affective temperament types.

Methods

Sample included 2,113 adolescents. The following instruments were used: questionnaire about: socio-demographic characteristics, intensity of the Internet use, content categories and online activities on the Internet; Facebook (FB) usage and life habits; The Internet Use Disorder Scale (IUDS). Based on their scores on the scale, subjects were divided into two groups - with or without PIU; Temperament Evaluation of Memphis, Pisa, Paris, and San Diego scale for adolescents (A-TEMPS-A).

Results

Various ML classification models on our data set were trained. Binary classification models were created (class-label attribute was PIU value). Models hyperparameters were optimized using grid search method and models were validated using k-fold cross-validation technique. Random forest was the model with the best overall results and the time spent on FB and the cyclothymic temperament were variables of highest importance for these model. We also applied the ML techniques Lasso and ElasticNet. The three most important variables for the development of PIU with both techniques were: cyclothymic temperament, the longer use of the Internet and the desire to use the Internet more than at present time. Group of variables having a protective effect (regarding the prevention of the development of PIU) was found with both techniques. The three most important were: achievement, search for contents related to art and culture and hyperthymic temperament. Next, 34 important variables that explain 0.76% of variance were detected using the genetic algorithms. Finally, the binary classification model (with or without PIU) with the best characteristics was trained using artificial neural network.

Conclusion

Variables related to the temporal determinants of Internet usage, cyclothymic temperament, the desire for increased Internet usage, anxious and irritable temperament, on line gaming, pornography, and some variables related to FB usage consistently appear as important variables for the development of PIU.

Mapping Blood Lead Levels in China during 1980-2040 with Machine Learning

Lead poisoning is globally concerning, yet limited testing hinders effective interventions in most countries. We aimed to create annual maps of county-specific blood lead levels in China from 1980 to 2040 using a machine learning model. Blood lead data from China were sourced from 1180 surveys published between 1980 and 2022. Additionally, regional statistical figures for 15 natural and socioeconomic variables were obtained or estimated as predictors. A machine learning model, using the random forest algorithm and 2973 generated samples, was created to predict county-specific blood lead levels in China from 1980 to 2040. Geometric mean blood lead levels in children (i.e., age 14 and under) decreased significantly from 104.4 μg/L in 1993 to an anticipated 40.3 μg/L by 2040. The number exceeding 100 μg/L declined dramatically, yet South Central China remains a hotspot. Lead exposure is similar among different groups, but overall adults and adolescents (i.e., age over 14), females, and rural residents exhibit slightly lower exposure compared to that of children, males, and urban residents, respectively. Our predictions indicated that despite the general reduction, one-fourth of Chinese counties rebounded during 2015-2020. This slower decline might be due to emerging lead sources like smelting and coal combustion; however, the primary factor driving the decline should be the reduction of a persistent source, legacy gasoline-derived lead. Our approach innovatively maps lead exposure without comprehensive surveys.

Predicting Low-Level Childhood Lead Exposure in Metro Atlanta Using Ensemble Machine Learning of High-Resolution Raster Cells

Low-level lead exposure in children is a major public health issue. Higher-resolution spatial targeting would significantly improve county and state-wide policies and programs for lead exposure prevention that generally intervene across large geographic areas. We use stack-ensemble machine learning, including an elastic net generalized linear model, gradient-boosted machine, and deep neural network, to predict the number of children with venous blood lead levels (BLLs) ≥2 to <5 µg/dL and ≥5 µg/dL in ~1 km2 raster cells in the metro Atlanta region using a sample of 92,792 children ≤5 years old screened between 2010 and 2018. Permutation-based predictor importance and partial dependence plots were used for interpretation. Maps of predicted vs. observed values were generated to compare model performance. According to the EPA Toxic Release Inventory for air-based toxic release facility density, the percentage of the population below the poverty threshold, crime, and road network density was positively associated with the number of children with low-level lead exposure, whereas the percentage of the white population was inversely associated. While predictions generally matched observed values, cells with high counts of lead exposure were underestimated. High-resolution geographic prediction of lead-exposed children using ensemble machine learning is a promising approach to enhance lead prevention efforts.

Prediction of Wave Energy Flux in the Bohai Sea through Automated Machine Learning

The rational assessment of regional energy distribution provides a scientific basis for the selection and siting of power generation units. This study, which focused on the Bohai Sea, set 31 research coordinate points in the Bohai sea for assessing the potential/trends of wave energy flux (WEF). We applied a point-to-point time series prediction method which modelled the different geographical coordinate points separately. Subsequently, we evaluated the performance of three traditional machine learning methods and three automated machine learning methods. To estimate WEF, the best model was applied to each research coordinate points, respectively. Then, the WEF was calculated and predicted based on the data of MWP, SWH, and water depth. The results indicate that, for all coordinates in the Bohai Sea, the H2O-AutoML algorithm is superior to the other five algorithms. Gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and stacked ensemble models yielded the best performance out of the H2O algorithms. The significant wave height (SWH), the mean wave period (MWP), and the WEF in the Bohai Sea tended to be concentrated in the center of the sea and dispersed in the nearshore areas. In the year 2000, 2010, 2020, and 2030, the maximum annual average WEF at each research coordinate in the Bohai Sea is around 1.5 kW/m, with a higher flux in autumn and winter. In summary, the results provide ocean parameter characterization for the design and deployment of wave energy harvesting devices. Moreover, the automated machine learning introduced herein has potential for use in more applications in ocean engineering.

Direct Prediction of Physicochemical Properties and Toxicities of Chemicals from Analytical Descriptors by GC-MS

With advances in machine learning (ML) techniques, the quantitative structure–activity relationship (QSAR) approach is becoming popular for evaluating chemicals. However, the QSAR approach requires that the chemical structure of the target compound is known and that it should be convertible to molecular descriptors. These requirements lead to limitations in predicting the properties and toxicities of chemicals distributed in the environment as in the PubChem database; the structural information on only 14% of compounds is available. This study proposes a new ML-based QSAR approach that can predict the properties and toxicities of compounds using analytical descriptors of mass spectrum and retention index obtained via gas chromatography–mass spectrometry without requiring exact structural information. The model was developed based on the XGBoost ML method. The root-mean-square errors (RMSEs) for log K_o-w, log (molecular weight), melting point, boiling point, log (vapor pressure), log (water solubility), log (LD₅₀) (rat, oral), and log (LD₅₀) (mouse, oral) are 0.97, 0.052, 51, 23, 0.74, 1.1, 0.74, and 0.6, respectively. The model performed well on a chemical standard mixture measurement, with similar results to those of model validation. It also performed well on a measurement of contaminated oil with spectral deconvolution. These results indicate that the model is suitable for investigating unknown-structured chemicals detected in measurements. Any online user can execute the model through a web application named Detective-QSAR (http://www.mixture-platform.net/Detective_QSAR_Med_Open/). The analytical descriptor-based approach is expected to create new opportunities for the evaluation of unknown chemicals around us.

Patterns of Children's Blood Lead Screening and Blood Lead Levels in North Carolina, 2011-2018-Who Is Tested, Who Is Missed?

BACKGROUND

No safe level of lead in blood has been identified. Blood lead testing is required for children on Medicaid, but it is at the discretion of providers and parents for others. Elevated blood lead levels (EBLLs) cannot be identified in children who are not tested.

OBJECTIVES

The aims of this research were to identify determinants of lead testing and EBLLs among North Carolina children and estimate the number of additional children with EBLLs among those not tested.

METHODS

We linked geocoded North Carolina birth certificates from 2011-2016 to 2010 U.S. Census data and North Carolina blood lead test results from 2011-2018. We estimated the probability of being screened for lead and created inverse probability (IP) of testing weights. We evaluated the risk of an EBLL of ≥3μg/dL at <30  months of age, conditional on characteristics at birth, using generalized linear models and then applied IP weights to account for missing blood lead results among unscreened children. We estimated the number of additional children with EBLLs of all North Carolina children using the IP-weighted population and bootstrapping to produce 95% credible intervals (CrI).

RESULTS

Mothers of the 63.5% of children (402,002 of 633,159) linked to a blood lead test result were disproportionately young, Hispanic, Black, American Indian, or on Medicaid. In full models, maternal age ≤20y [risk ratio (RR)=1.10; 95% confidence interval (CI): 1.13, 1.20] or smoking (RR=1.14; 95% CI: 1.12, 1.17); proximity to a major roadway (RR=1.10; 95% CI: 1.05, 1.15); proximity to a lead-releasing Toxics Release Inventory site (RR=1.08; 95% CI: 1.03, 1.14) or a National Emissions Inventory site (RR=1.11; 95% CI: 1.07, 1.14); and living in neighborhoods with more housing built before 1950 (RR=1.10; 95% CI: 1.05, 1.14) or before 1940 (RR=1.18; 95% CI: 1.11, 1.25) or more vacant housing (RR=1.14; 95% CI: 1.11, 1.17) were associated with an increased risk of EBLL, whereas overlap with a public water service system was associated with a decreased risk of EBLL (RR=0.85; 95% CI: 0.83, 0.87). Children of Black mothers were no more likely than children of White mothers to have EBLLs (RR=0.98; 95% CI: 0.96, 1.01). Complete blood lead screening in 2011-2018 may have identified an additional 17,543 (95% CrI: 17,462, 17,650) children with EBLLs ≥3μg/dL.

DISCUSSION

Our results indicate that current North Carolina lead screening strategies fail to identify over 30% (17,543 of 57,398) of children with subclinical lead poisoning and that accounting for characteristics at birth alters the conclusions about racial disparities in children's EBLLs. https://doi.org/10.1289/EHP10335.

Explainable artificial intelligence (XAI) for exploring spatial variability of lung and bronchus cancer (LBC) mortality rates in the contiguous USA

Machine learning (ML) has demonstrated promise in predicting mortality; however, understanding spatial variation in risk factor contributions to mortality rate requires explainability. We applied explainable artificial intelligence (XAI) on a stack-ensemble machine learning model framework to explore and visualize the spatial distribution of the contributions of known risk factors to lung and bronchus cancer (LBC) mortality rates in the conterminous United States. We used five base-learners-generalized linear model (GLM), random forest (RF), Gradient boosting machine (GBM), extreme Gradient boosting machine (XGBoost), and Deep Neural Network (DNN) for developing stack-ensemble models. Then we applied several model-agnostic approaches to interpret and visualize the stack ensemble model's output in global and local scales (at the county level). The stack ensemble generally performs better than all the base learners and three spatial regression models. A permutation-based feature importance technique ranked smoking prevalence as the most important predictor, followed by poverty and elevation. However, the impact of these risk factors on LBC mortality rates varies spatially. This is the first study to use ensemble machine learning with explainable algorithms to explore and visualize the spatial heterogeneity of the relationships between LBC mortality and risk factors in the contiguous USA.