Partial Dependence Research Articles

The application of explainable artificial intelligence (XAI) in electronic health record research: A scoping review.

Machine Learning (ML) and Deep Learning (DL) models show potential in surpassing traditional methods including generalised linear models for healthcare predictions, particularly with large, complex datasets. However, low interpretability hinders practical implementation. To address this, Explainable Artificial Intelligence (XAI) methods are proposed, but a comprehensive evaluation of their effectiveness is currently limited. The aim of this scoping review is to critically appraise the application of XAI methods in ML/DL models using Electronic Health Record (EHR) data. In accordance with PRISMA scoping review guidelines, the study searched PUBMED and OVID/MEDLINE (including EMBASE) for publications related to tabular EHR data that employed ML/DL models with XAI. Out of 3220 identified publications, 76 were included. The selected publications published between February 2017 and June 2023, demonstrated an exponential increase over time. Extreme Gradient Boosting and Random Forest models were the most frequently used ML/DL methods, with 51 and 50 publications, respectively. Among XAI methods, Shapley Additive Explanations (SHAP) was predominant in 63 out of 76 publications, followed by partial dependence plots (PDPs) in 11 publications, and Locally Interpretable Model-Agnostic Explanations (LIME) in 8 publications. Despite the growing adoption of XAI methods, their applications varied widely and lacked critical evaluation. This review identifies the increasing use of XAI in tabular EHR research and highlights a deficiency in the reporting of methods and a lack of critical appraisal of validity and robustness. The study emphasises the need for further evaluation of XAI methods and underscores the importance of cautious implementation and interpretation in healthcare settings.

Interpretable machine learning for predicting chronic kidney disease progression risk.

Chronic kidney disease (CKD) poses a major global health burden. Early CKD risk prediction enables timely interventions, but conventional models have limited accuracy. Machine learning (ML) enhances prediction, but interpretability is needed to support clinical usage with both in diagnostic and decision-making. A cohort of 491 patients with clinical data was collected for this study. The dataset was randomly split into an 80% training set and a 20% testing set. To achieve the first objective, we developed four ML algorithms (logistic regression, random forests, neural networks, and eXtreme Gradient Boosting (XGBoost)) to classify patients into two classes-those who progressed to CKD stages 3-5 during follow-up (positive class) and those who did not (negative class). For the classification task, the area under the receiver operating characteristic curve (AUC-ROC) was used to evaluate model performance in discriminating between the two classes. For survival analysis, Cox proportional hazards regression (COX) and random survival forests (RSFs) were employed to predict CKD progression, and the concordance index (C-index) and integrated Brier score were used for model evaluation. Furthermore, variable importance, partial dependence plots, and restrict cubic splines were used to interpret the models' results. XGBOOST demonstrated the best predictive performance for CKD progression in the classification task, with an AUC-ROC of 0.867 (95% confidence interval (CI): 0.728-0.100), outperforming the other ML algorithms. In survival analysis, RSF showed slightly better discrimination and calibration on the test set compared to COX, indicating better generalization to new data. Variable importance analysis identified estimated glomerular filtration rate, age, and creatinine as the most important predictors for CKD survival analysis. Further analysis revealed non-linear associations between age and CKD progression, suggesting higher risks in patients aged 52-55 and 65-66 years. The association between cholesterol levels and CKD progression was also non-linear, with lower risks observed when cholesterol levels were in the range of 5.8-6.4 mmol/L. Our study demonstrated the effectiveness of interpretable ML models for predicting CKD progression. The comparison between COX and RSF highlighted the advantages of ML in survival analysis, particularly in handling non-linearity and high-dimensional data. By leveraging interpretable ML for unraveling risk factor relationships, contrasting predictive techniques, and exposing non-linear associations, this study significantly advances CKD risk prediction to enable enhanced clinical decision-making.

Multitask learning to predict successful weaning in critically ill ventilated patients: A retrospective analysis of the MIMIC-IV database.

Weaning is an essential issue in critical care. This study explores the efficacy of multitask learning models in predicting successful weaning in critically ill ventilated patients using the Medical Information Mart for Intensive Care (MIMIC) IV database. We employed a multitask learning framework with a shared bottom network to facilitate common knowledge extraction across all tasks. We used the Shapley additive explanations (SHAP) plot and partial dependence plot (PDP) for model explainability. Furthermore, we conducted an error analysis to assess the strength and limitation of the model. Area under receiver operating characteristic curve (AUROC), calibration plot and decision curve analysis were used to determine the performance of the model. A total of 7758 critically ill patients were included in the analyses, and 78.5% of them were successfully weaned. Multitask learning combined with spontaneous breath trial achieved a higher performance to predict successful weaning compared with multitask learning combined with shock and mortality (area under receiver operating characteristic curve, AUROC, 0.820 ± 0.002 vs 0.817 ± 0.001, p < 0.001). We assessed the performance of the model using calibration and decision curve analyses and further interpreted the model through SHAP and PDP plots. The error analysis identified a relatively high error rate among those with low disease severities, including low mean airway pressure and high enteral feeding. We demonstrated that multitask machine learning increased predictive accuracy for successful weaning through combining tasks with a high inter-task relationship. The model explainability and error analysis should enhance trust in the model.

Development and Validation of an Explainable Radiomics Model to Predict High-Aggressive Prostate Cancer: A Multicenter Radiomics Study Based on Biparametric MRI.

In the last years, several studies demonstrated that low-aggressive (Grade Group (GG) ≤ 2) and high-aggressive (GG ≥ 3) prostate cancers (PCas) have different prognoses and mortality. Therefore, the aim of this study was to develop and externally validate a radiomic model to noninvasively classify low-aggressive and high-aggressive PCas based on biparametric magnetic resonance imaging (bpMRI). To this end, 283 patients were retrospectively enrolled from four centers. Features were extracted from apparent diffusion coefficient (ADC) maps and T2-weighted (T2w) sequences. A cross-validation (CV) strategy was adopted to assess the robustness of several classifiers using two out of the four centers. Then, the best classifier was externally validated using the other two centers. An explanation for the final radiomics signature was provided through Shapley additive explanation (SHAP) values and partial dependence plots (PDP). The best combination was a naïve Bayes classifier trained with ten features that reached promising results, i.e., an area under the receiver operating characteristic (ROC) curve (AUC) of 0.75 and 0.73 in the construction and external validation set, respectively. The findings of our work suggest that our radiomics model could help distinguish between low- and high-aggressive PCa. This noninvasive approach, if further validated and integrated into a clinical decision support system able to automatically detect PCa, could help clinicians managing men with suspicion of PCa.

랜덤포레스트 기법 활용 대학 중도탈락 기관변인 탐색

This study analyzed institutional factors affecting college student retention rates using a machine learning predictive model. The data of 2022, 2021, and 2022 university information disclosure data consisting of all four-year universities, capital region, and non-capital region were analyzed with the randomForest algorithm in R. First, the study findings indicate that the randomForest model demonstrated an accuracy rate of approximately 70%. Second, 10 key variables were identified in the overall university data set and 9 key variables in the capital region and non-capital region university data sets. However, there were some differences in the composition and ranking of the key factors in the data sets. Third, the partial dependence plots reveal the value ranges of the variables where the retention rate either decreased or remained stable. Institutions may apply these results by utilizing randomForest machine learning techniques for data-driven decision-making in formulating university retention policies.

An interpretable clustering approach to safety climate analysis: Examining driver group distinctions

The transportation industry, particularly the trucking sector, is prone to workplace accidents and fatalities. Accidents involving large trucks accounted for a considerable percentage of overall traffic fatalities. Recognizing the crucial role of safety climate in accident prevention, researchers have sought to understand its factors and measure its impact within organizations. While existing data-driven safety climate studies have made remarkable progress, clustering employees based on their safety climate perception is innovative and has not been extensively utilized in research. Identifying clusters of drivers based on their safety climate perception allows the organization to profile its workforce and devise more impactful interventions. The lack of utilizing the clustering approach could be due to difficulties interpreting or explaining the factors influencing employees’ cluster membership. Moreover, existing safety-related studies did not compare multiple clustering algorithms, resulting in potential bias. To address these problems, this study introduces an interpretable clustering approach for safety climate analysis. This study compares five algorithms for clustering truck drivers based on their safety climate perceptions. It also proposes a novel method for quantitatively evaluating partial dependence plots (QPDP). Then, to better interpret the clustering results, this study introduces different interpretable machine learning measures (Shapley additive explanations, permutation feature importance, and QPDP). The Python code used in this study is available at https://github.com/NUS-DBE/truck-driver-safety-climate. This study explains the clusters based on the importance of different safety climate factors. Drawing on data collected from more than 7,000 American truck drivers, this study significantly contributes to the scientific literature. It highlights the critical role of supervisory care promotion in distinguishing various driver groups. Moreover, it showcases the advantages of employing machine learning techniques, such as cluster analysis, to enrich the scientific knowledge in this field. Future studies could involve experimental methods to assess strategies for enhancing supervisory care promotion, as well as integrating deep learning clustering techniques with safety climate evaluation.

Daily concentration of snowfalls in the mountains of the Iberian Peninsula

AbstractThe temporal concentration of snowfalls has direct implications on the management of water resources as well as on the economic activity of mountain areas, conditioning, for example, the seasonal performance of ski resorts. This work uses the daily concentration index (CI) for analysing the frequency concentration of snowfalls in the Iberian Peninsula Mountain ranges. First, we provide a spatiotemporal analysis of the CI patterns and trends for the 1980–2014 period. Subsequently, we determine the atmospheric circulation patterns that control the CI variability. In addition, we determine the geographical and low‐frequency climate modes that control the CI for this mid‐latitude area. In addition, we have estimated the partial dependence relationship between the CI and several geographical factors by fitting a multiple linear regression. The results from these analyses show that elevation as well as the distance from the Atlantic are the main geographical pattern that controls the CI in the Iberian Peninsula Mountain ranges. These geographical factors also reflect the role of the main atmospheric circulation patterns in the Iberian Peninsula in controlling the spatial dynamics of the CI. The Cantabrian, Iberian, and northern slopes of the Pyrenees show the lowest CI due to their exposition to northern and Atlantic circulation weather types. On the contrary, the highest CI values are found in the southern and eastern slopes of the Pyrenees, eastern slopes of Sierra Nevada, and southern slopes of the Central system. Trend analysis shows a slight increase of CI in the Central system and in the western Sierra Nevada. However, eastern Sierra Nevada, Cantabrian, Central, and Iberian show a downward CI trend. CI is principally driven by the East‐Atlantic/Western Russia pattern and the North Atlantic Oscillation (NAO) in the Cantabrian, Iberian, and northern slopes of the Central range. The CI values in the Pyrenees show a different relationship with the Western Mediterranean Oscillation (WeMO) depending on whether it is the southern or the northern slope. In addition, the positive phase of the NAO oscillation controls the higher values of CI for the whole Pyrenees, especially in the mid‐south part. Finally, in Sierra Nevada the CI dynamics are controlled mostly by the WeMO.

Exploring the Behavior-Driven Crash Risk Prediction Model: The Role of Onboard Navigation Data in Road Safety

Driving behavior has frequently been overlooked in previous road traffic crash research. Hereby, abnormal (extreme) driving behavior data transmitted by the onboard navigation systems were collected for vehicles involved in traffic crashes, including sharp-lane-change, sharp-acceleration, and sudden-braking behaviors. Using these data in conjunction with expressway crash records, multiple classification learners were trained to establish a behavior-driven risk prediction model. To further investigate the influence of driving behavior on crash risk, partial dependence plots (PDPs) were applied. Regression analyses indicate that models have a stronger effect when derivative features such as frequency of specific deviant behavior, speed, and acceleration in the behavior process are included. The behavioral RUSBoost model surpasses other models, achieving an AUC prediction metric of 0.782 and outperforming traditional traffic-flow-driven machine learning models. PDP analysis demonstrates that the sudden-braking behavior is the leading contributory factor of expressway crashes, particularly when the acceleration exceeds 0.5 G. This study confirms the potential of predicting crash risks through augmenting behavior data from navigation software; the findings lay a foundation for countermeasures.

Synergistic effects of supplementary cementitious materials and compressive strength prediction of concrete using machine learning algorithms with SHAP and PDP analyses

In order to reduce the CO2 associated with cement production, this study explored the potential of rice husk ash (RHA) and fly ash (FA) as supplementary cementitios materials for partially replacing cement in concrete production. The study aimed to analyze the synergistic effects of a cement-based mixture consisting of RHA and FA in different proportions on concrete's fresh, hardened, non-destructive and microscopic properties. In addition to the experimental work, this study successfully applied machine learning to predict the compressive strength of RHA-FA concrete using three types of algorithms: ANN (Analytical Neural Network), XGB (Extreme Gradient Boosting), and GBM (Gradient Boosting Model). A total of 138 data points were used for this prediction, and statistical and parametric analyses were performed to define the impact of input parameters on the outcome. Furthermore, both destructive and non-destructive tests were conducted on hardened concrete, including compressive strength, split tensile strength, ultrasonic pulse velocity (UPV), and rebound hammer. The scanning electron microscope (SEM) was operated to analyze the microstructural characteristics of concrete. The compressive and split-tensile strength test results showed that the mixture with a higher percentage of fly ash and a lower percentage of rice husk ash achieved maximum strength. The X-ray Fluorescence (XRF) analysis revealed that both ashes contained a significant amount of silica, which gave them excellent pozzolanic properties. After 28 days, both the UPV and the rebound hammer strength align with the destructive compressive strength results. The study also employed SHAP (SHapley Additive exPlanations) and PDP (Partial Dependence Plot) analyses to identify the optimal range for each parameter's contribution to strength improvement. The machine learning models exhibited a strong correlation with the test results, achieving an R2 value of 0.84 for the XGBoost model.

Optimization of structural and electrical properties of graphene-based TiO2 thin film device using Bayesian machine-learning approach

The optimization of laser-induced graphene (LIG) patterning is crucial for achieving desirable electronic properties in flexible electronic devices. In this study, we applied state-of-the-art automated parameter optimization techniques, specifically Bayesian optimization, to enhance the electrical resistance of LIG patterns. By iteratively optimizing the laser power, irradiation time, pressure, and gas type, we achieved minimum LIG resistance within eight batch configurations suggested by the Bayesian optimization (BO) approach. Our method eliminates the reliance on skilled operators, as the initial surrogate models were trained with random parameter evaluations. Notably, our system enables the optimization of material properties even when characterizations are only available outside the experiment loop. Furthermore, the surrogate model provided insights into the underlying mechanisms of LIG growth on quartz substrates. Through partial dependence analysis, we identified relevant physical domains for further investigation, leading to the discovery of negative capacitance and a correlation between structural and electrical properties in LIG. These findings were supported by XPS, Raman, and optical characterizations. Our approach streamlines the experimental design, reducing time and cost while accelerating materials research, and offers human-interpretable conclusions for a deeper understanding of LIG patterning processes.

Analysis of the relationship between metro ridership and built environment: A machine learning method considering combinational features

Limited studies have examined the relationship between combinational features of the built environment and metro ridership. In this study, we applied the gradient boosting regression tree (GBRT) to explore the non-linearity effects for metro commuter ridership and non-commuter ridership, respectively. The results show that the number of bus stops is an important predictor of urban metro commuter ridership, whereas residential and community services, road density, and leisure and entertainment are important predictors of urban metro non-commuter ridership. Based on the partial dependence plots generated by GBRT, the results confirm that combinational features of the built environment have a larger impact on commuter ridership (e.g., high-high level road density-life services on the destination side). Combinational features at a higher level exhibit a more substantial influence on commuter ridership, whereas combinational features at a lower level demonstrate a greater impact on non-commuter ridership.

What determines the real-world CO2 emission reductions of ridesplitting trips?

Ridesplitting has great potential to alleviate the negative impacts of ridesourcing on the environment. However, not all ridesplitting trips are effective in reducing CO2 emissions. The actual trip-level environmental impact of ridesplitting and its determinants remain unclear. To bridge this gap, this study reveals what determines the real-world CO2 emission reductions of ridesplitting trips based on the observed data of ridesourcing services in Chengdu, China. First, the CO2 emission reduction of each ridesplitting trip compared with its substituted single rides is calculated using a vehicle emission model (COPERT) coupled with GPS trajectory data. Then, four implementations of gradient boosting machines, namely GBDT, XGboost, LighGBM, and CatBoost models, are employed to explore the relationships between the ridesplitting CO2 emission reduction rate and its explanatory variables. Finally, the nonlinear and interaction effects of these variables are analyzed based on partial dependence plots and SHapley Additive ExPlanations. Empirical results show that the ridesplitting CO2 emission reduction rate varies from trip to trip, with an average of 43.15 g/km. In the real world, approximately 15 % of ridesplitting trips even increase CO2 emissions, mainly due to the low overlap rate and high detour rate of the shared rides. To enhance the CO2 emission reduction rate of ridesplitting, the overlap rate, number of shared rides, average speed, and ride distance ratio should be increased, while the detour rate, actual trip distance, and ride distance gap should be decreased. This study can help the ridesourcing companies better match the shared rides and improve the environmental benefits of ridesplitting.

Interpretability study on prediction models for alloy pitting based on ensemble learning

Ensemble learning (EL) models were developed to predict the pitting potential (Ep) of alloys. By carefully selecting features and optimizing parameters, the predictive accuracy of the models was improved and experimentally validated. Interpretability methods, SHapley Additive exPlanations (SHAP), Partial Dependence Plot (PDP), and Accumulated Local Effect (ALE), explained the model's prediction logic effectively and portrayed the effects of alloying elements, thermodynamic parameters, and environmental factors on Ep in detail. Furthermore, performance differences between interpretability methods, errors in model prediction, and failure cases of interpretability methods were discussed, which provided insights for model optimization and appropriate Interpretability methods selection.

Estimation of compressive strength for spiral stirrup-confined circular concrete column using optimized machine learning with interpretable techniques

The compressive strength (CS) of a concrete column confined with spiral stirrups is an important indicator for assessing the safety and stability of concrete structures. However, achieving accurate CS estimation for confined concrete remains challenging due to the complex confinement mechanism provided by spiral stirrups. In this study, three robust machine learning (ML) algorithms—support vector regression (SVR), random forest (RF) and extreme gradient boosting (XGBoost)—are employed to predict the CS value of the spiral stirrup-confined circular concrete columns. The hyperparameters of the ML models undergo fine-tuning via Bayesian optimization with 10-fold cross-validation, and the optimized ML models are evaluated for their predictive capabilities. Results show that compared to SVR and RF, XGBoost exhibits more stable generalization performance, achieving an average coefficient of determination (R 2) of 0.944 for the 10-fold cross-validation, and demonstrates superior accuracy on the testing dataset with an R 2 value of 0.967. To provide insights into the relationship between input features and the output CS value, Individual Conditional Exception (ICE) plots, one/two-dimensional Partial Dependence Plots (PDPs), and Shapley Additive Explanation (SHAP) techniques are utilized to interpret the optimized XGBoost model. Additionally, a friendly online graphical user interface (GUI) has been specially developed based on the optimized XGBoost model to facilitate convenient CS estimation for spiral stirrup-confined circular concrete column.

Wine quality analysis by the structural causal model (SCM)

Bayes network modelling for structural causal analysis between wine physicochemical data and quantitative human quality blind assessments is applied. The large dataset of white and red "Vinho Verde'' wine samples from Portugal, which was available from an open data repository for machine learning at the University of California at Irving, was analysed. The dataset contains 4898 white and 1599 red samples evaluated by blind tastes by a minimum of 3 sensory assessors and 12 physicochemical properties. The casual effects of wine analytic data on human quality evaluations are evaluated numerically by Bayes neural networks for adjusted sets of the covariates as marginal distributions and presented graphically as partial dependence plots. Structural causal analysis revealed important differences between the most important variables for quality predictions and the individual causal effects. Bayes neural network models of the partial dependencies show more pronounced nonlinear effects for red wines compared to white wine quality. The artificial intelligence models with boosted random decision tree forests for untrained wine samples yield a 5% relative standard error of predictions compared to 12% for the linear models and ordinary least squares estimation. For red wine, the most important direct causal quality effects are caused by alcohol, volatile acidity, and sulphates. Alcohol improves quality with a maximum plateau at 14%, while volatile acidity has a strong proportional negative effect. The effect of sulphates is highly nonlinear with maximum positive effect at a concentration of 1 g/L of K2SO4. For the white wine samples causal effects are linear with positive effects of alcohol and negative effects of volatile and fixed acidity. The developed structural causal model enables evaluation of targeted wine production interventions, named as “doing x, do(x) models”, as restructured adjusted Bayes networks. It leads to potential applications of artificial intelligence in wine production technology and process quality control.

Improved climate time series forecasts by machine learning and statistical models coupled with signature method: A case study with El Niño

The different phases of ENSO (El Niño Southern Oscillation) directly influence the occurrence of natural disasters and global warming. To limit the socio-economic impact, it is essential to develop simple and fast numerical models that can predict these different cycles. Here, we aimed to improve the predictive performance and extracting relevant information from climatic events by applying the signature method to time series models. After transforming the data using this signature method, we performed a comparative analysis of the statistical and machine learning models. In addition, we used PDP (Partial Dependence Plot) and SPRC (Standard Partial Regression Coefficient) to better understand the interactions between different climate indices.Our results showed that the best predictive performance was obtained when we use the signature method with the LSTM (R2 = 0.74) and Lasso (R2 = 0.79) models. Two complementary methods were used to highlight the influence of the following climate indices on ENSO cycle changes: NINO3, NINO3.4 and NPI (North Pacific Index). These methodologies also enabled us to determine the switchover thresholds, and order temporal variations. With this first application of the signature method to this type of time series, we obtained accurate forecasts on a 6-month scale with reduced computation time. This suggests that our methodology can be applied to many other fields of research that use multivariate time-series analyses.

Using machine learning to explore oxyanion adsorption ability of goethite with different specific surface area

In this study, we developed prediction models for the adsorption of divalent and trivalent oxyanions on goethite based on machine learning algorithms. After verifying the reliability of the models, the importance of goethite specific surface area (SSA) and the average oxyanion adsorption capacities of goethite with different SSAs were calculated by shapley additive explanations (SHAP) importance analysis and partial dependence (PD) analysis. Despite there were differences in the feature importance of divalent and trivalent oxyanions, the contribution of goethite's SSA to the adsorption amount ranked the fourth based on SHAP importance, indicating SSA played the important role in oxyanion adsorption. Meanwhile, the PD values of SSA and the optimized complexation constants from surface complexation modeling (SCM) both indicated a non-monotonic relationship between the goethite with different SSA and its oxyanions binding capacity. When the total site concentration and crystal face composition were used as the machine learning model input features, the SHAP importance values of crystal faces and the PD decomposition results indicated that the (001) face showed the crucial influence on oxyanions adsorption amount. These findings demonstrated the important role of crystal face composition in goethite's adsorption ability, and provided a theoretical explanation for the variations of oxyanions adsorption amount on different SSA goethite.

Exploring the association between two groups of metals with potentially opposing renal effects and renal function in middle-aged and older adults: Evidence from an explainable machine learning method

BackgroundMachine learning models have promising applications in capturing the complex relationship between mixtures of exposures and outcomes. ObjectiveOur study aimed at introducing an explainable machine learning (EML) model to assess the association between metal mixtures with potentially opposing renal effects and renal function in middle-aged and older adults. MethodsThis study extracted data from two cycle years of the National Health and Nutrition Examination Survey (NHANES). Participants aged 45 years or older with complete data on six metals (lead, cadmium, manganese, mercury, and selenium) and related covariates were enrolled. The EML model was developed by the optimized machine learning model together with Shapley Additive exPlanations (SHAP) to assess the chronic kidney disease (CKD) risk with metal mixtures. The results from EML were further compared in detail with multiple logistic regression (MLR) and Bayesian kernel machine regression (BKMR). ResultsAfter adjusting for included covariates, MLR pointed out the lead and arsenic were generally positively associated with CKD, but manganese had a negative association. In the BKMR analysis, each metal was found to have a non-linear association with the risk of CKD, and interactions can exist between metals, especially for arsenic and lead. The EML ranked the feature importance: lead, manganese, arsenic and selenium were close behind in importance after gender, age or BMI for participants with CKD. Strong interactions between mercury and lead, manganese and cadmium and arsenic and manganese were identified by partial dependence plot (PDP) of SHAP and bivariate exposure-response effect plots of BKMR. The EML model determined the “trigger point” at which the risk of CKD abruptly changed. ConclusionCo-exposure to metals with different nephrotoxicity could have different joint association with renal function, and EML can be a powerful method for studying complex exposure mixtures.

Understanding the driving mechanisms of site contamination in China through a data-driven approach

China currently faces significant environmental risks stemming from contaminated sites. The driving mechanism of site contamination, influenced by various drivers, remain obscured due to a dearth of quantitative methodologies and comprehensive data. Here, we used a data-driven causality inference approach to construct an interpretable random forest (RF) model. Results show that: (1) the trained RF model demonstrated remarkable predictive accuracy for identifying contaminated sites, with an accuracy rate of 0.89. In contrast to conventional correlation analysis, the RF model excels in discerning the key drivers through non-linear and genuine causal relationships between these drivers and site contamination. (2) Among the 25 potential drivers, we identified 18 key drivers of site contamination. These drivers encompass a broad spectrum of factors, including production and operational data, pollutant control level, site protection capability, pollutant characteristics, and physical-geographical conditions. (3) Each key driver exerts varying impacts on site pollution, with diverse directions, intensities, and underlying patterns. The partial dependence plots (PDPs) illuminate the role of each key driver, its critical value contributing to site pollution, and the interplay between these drivers. The key drivers facilitate the realization of three primary contamination processes: uncontrolled release, effective migration, and persistent accumulation. In light of our findings, environmental managers can proactively prevent site contamination by regulating single, dual, and multiple key drivers to disrupt critical pollution processes. This research offers valuable insights for devising targeted strategies and interventions aimed at mitigating environmental risks associated with contaminated sites in China.

Agroecologies defined by species distribution models improve model fit of genotype by environment interactions to identify the best performing chicken breeds for smallholder systems

Animal performance is an outcome of genetic effects, environmental influences, and their interaction. Understanding the influences of the environment on performance is important to identify the right breeds for a given environment. Agroecological zonation is commonly used to classify environments and compare the performance of breeds before their wider introduction into a new environment. Environmental classes, also referred to as agroecologies, are traditionally defined based on agronomically important environmental predictors. We hypothesized that our own classification of agroecologies for livestock at a species level and incorporating the most important environmental predictors may improve genotype by environment interactions (GxE) estimations over conventional methodology. We collected growth performance data on improved chicken breeds distributed to multiple environments in Ethiopia. We applied species distribution models (SDMs) to identify the most relevant environmental predictors and to group chicken performance testing sites into agroecologies. We fitted linear mixed-effects models (LMM) to make model comparisons between conventional and SDM-defined agroecologies. Then we used Generalized Additive Models (GAMs) to visualize the influences of SDM-identified environmental predictors on the live body weight of chickens at species level. The model fit in LMM for GxE prediction improved when agroecologies were defined based on SDM-identified environmental predictors. Partial dependence plots (PDPs) produced by GAMs showed complex relationships between environmental predictors and body weight. Our findings suggest that multi-environment performance evaluations of candidate breeds should be based on SDM-defined environmental classes or agroecologies. Moreover, our study shows that GAMs are well-suited to visualizing the influences of bioclimatic factors on livestock performance.