Dependence Plots Research Articles

Future-Proofing EU-27 Energy Policies with AI: Analyzing and Forecasting Fossil Fuel Trends

The energy sector plays a pivotal role in economic development, societal progress, and environmental sustainability, yet heavy reliance on fossil fuels remains a major challenge for achieving climate neutrality. Within this context, the European Union (EU-27) has committed to ambitious climate goals, including achieving carbon neutrality by 2050, making it a critical region for studying energy transition. This study analyzes the determinants of fossil fuels’ share (SFF) in final energy consumption at the aggregate EU-27 level over a 19-year period (2004–2022) and forecasts trends in the region’s energy transition through 2030. Using a random forest (RF) regressor, complex nonlinear relationships between SFF and six key predictors—GDP, population, industrial production, CO2 emissions, renewable energy share (SRE), and energy intensity—were modeled. Model interpretability was enhanced through Shapley additive explanations (SHAP) and partial dependence plots (PDPs), revealing CO2 emissions and SRE as the dominant predictors with opposing effects on SFF. Interaction effects highlighted the synergistic role of emission reduction and renewable energy adoption in minimizing fossil fuel reliance. GDP, while less influential overall, exhibited a significant negative relationship with SFF during early growth stages. Forecasts indicate a steady decline in fossil fuel reliance, from 1.8% in 2022 to 1.33% by 2030, supporting the EU’s climate objectives by emphasizing the importance of renewable energy adoption and emission control. This study demonstrates the transformative potential of machine learning and explainable AI (XAI) techniques in providing actionable insights to advance the EU-27’s sustainability journey.

Comparing Traditional Methods and Modern Statistical Techniques for Tree Height Prediction

Forest mensuration is important to gain knowledge and information about forest stands. Because tree height often proves more difficult to measure than diameter, different statistical models are used for their estimation instead. In this paper, the data of 986 spruce trees (Picea abies KARST. (L.)), measured in the federal states of Salzburg and Tyrol (Austria), were used to train and compare random forest with more traditional approaches such as linear and non-linear mixed models and a classical uniform height curve. For model comparison, RMSE, percent RMSE, percent bias, and bias are used. For further visualization of the differences, residual plots, partial dependence plots, and conditional dependence plots are shown. The results show that random forest (RMSE 2.23 m) can compete with more traditional methods, such as non-linear (RMSE 2.14 m) and linear (RMSE 2.24 m) mixed models or uniform height curves (RMSE 2.92 m), but is not able to outperform those methods, especially when it comes to extrapolation or prediction in areas where training data are sparse or not available. Furthermore, the results show that the incorporation of additional covariates can improve the prediction of certain models.

Analysis of geometrical parameter sensitivity and mechanism of flow loss in exhaust volute based on particle swarm optimization

To investigated the mechanism of flow loss occurs in gas-turbine exhaust volute, comprehensive analysis was conducted based on the traces of parametric optimizations. An original volute was parameterized by depicting the core-part with 6 parameters. Concerning the coefficients of total pressure loss and static pressure recovery, the volute was optimized, as a start of the investigation, by Particle Swarm Optimization (PSO) to generate 56 exhaust volute designs and 8 geometric parameter traces. The geometry evolution history was fully recorded by these traces during the optimization. Based on this, parameter sensitivity analyses were conducted by single, double and K-means-clustering-based comprehensive parameters. Furthermore, partial dependence plot (PDP) and individual conditional expectation plot (ICEP) were applied to enhance the expression of parameter sensitivity. This enables the detailed discussion of the mechanisms of flow loss occurs in exhaust volute. The results demonstrate that the Particle Swarm Optimization (PSO) algorithm is highly effective for optimizing the exhaust volute. After six rounds of optimization with eight particles per round, the total pressure loss coefficient at the exhaust volute outlet was reduced by 35%, while the static pressure recovery coefficient increased by 79%. The sensitivity analysis reveals that geometric parameters exhibit varying degrees of influence on aerodynamic performance, with diffuser length being the most critical factor. Notably, a shorter diffuser, constrained by the same external dimensions, tends to result in lower flow losses. Flow loss within the collector accounts for 71.7% of the total loss, which can be attributed to surface friction and turbulent dissipation. The latter is primarily driven by turbulent viscous dissipation, predominantly occurring in the collector section. Additionally, the size of large-scale vortices in the curved parts of the diffuser and collector, which contribute to turbulent dissipation, as well as the ratio of the average flow velocity to circulation speed in the collector, which affects wall friction, are key factors in flow loss generation.

Assessing the influence of green space morphological spatial pattern on urban waterlogging: A case study of a highly-urbanized city.

The extensive expansion of impervious surfaces encroaches on green spaces and causes frequent urban waterlogging disasters. Previous studies have focused mainly on the influence of green space landscape pattern on waterlogging, with less attention given to green space morphological spatial pattern (MSPA). MSPA can be used to differentiate various types of land use morphologies from a microscopic perspective and reveal visualized spatial characteristics. Therefore, this study selected Shenzhen, a city with serious waterlogging problems, as the study area. The anthropogenic/natural environments and green space morphological spatial pattern were considered. Pearson correlation analysis and random forest regression were combined to investigate the influence of these drivers on the density of waterlogging hotspots and quantify the degree of importance for each driver. The results were supplemented with explanations using SHapley Additive exPlanations and Partial Dependence Plots. Pearson correlation analysis revealed that green space morphological spatial pattern, the proportion of green spaces, and the proportion of impervious surfaces were the dominant drivers. Additionally, the random forest regression showed that incorporating green space morphological spatial pattern and average tree height as potential drivers could strengthen the model's goodness-of-fit. While the proportion of impervious surfaces, the proportion of green spaces, and population density were important drivers, the green space morphological spatial pattern, specifically the "loop", "edge", and "core", was even more crucial and had an optimal design range. Therefore, green space morphological spatial pattern should be emphasized during the planning of "sponge cities" to maximize the ability of green spaces to mitigate waterlogging. In summary, our findings are expected to provide feasible suggestions for waterlogging control and green space planning.

Construction of a health literacy prediction model for diabetic patients: A multicenter study.

To achieve a rapid assessment of health literacy (HL) levels among diabetic patients. A questionnaire survey was conducted among diabetic patients from nine communities in Nantong City, Jiangsu Province, China, using convenient sampling. Based on the survey results, data from three communities were randomly selected as the test set, with the remaining data used as the training set. Feature selection was performed using recursive feature elimination. Predictive models were established and compared using logistic regression (LR), random forest (RF), and support vector machine (SVM). Calibration curves, decomposition plots, and partial dependence plots were drawn to evaluate and interpret the models. In November 2023, a total of 802 valid questionnaires were received. Eight variables were selected for modeling: educational level, exercise habits, average monthly household income, dietary control, age, medication for blood sugar control, duration of diabetes, and number of cohabitants. The recall for LR in the three communities was 0.778, 0.800, and 0.862 [area under the curve (AUC): 0.810, 0.792, and 0.775]. For RF, the recall values were 0.879, 0.877, and 0.923 (AUC: 0.781, 0.710, and 0.710). For SVM, the recall values were 0.859, 0.862, and 0.877 (AUC: 0.813, 0.759, and 0.770). Model evaluation showed that as the data volume increased, the calibration curves became more ideal. As one of the few HL prediction models for diabetic patients in mainland China that is built based on multi-center survey data and evaluated through multi-center assessment, this model can quickly identify patients with insufficient HL using a small amount of objective personal information.

Simulation and prediction of PM2.5 concentrations and analysis of driving factors using interpretable tree-based models in Shanghai, China.

PM2.5 is a critical air pollutant, and understanding its drivers is essential for regional air quality control. This study employed meteorological and pollutant variables to predict PM2.5 concentrations in Shanghai using interpretable tree-based models. The random forest (RF) model performed best, achieving MAE, RMSE, MBE, and R2 values of 3.279, 4.609, 1.254, and 0.971, respectively, improving accuracy by 42.1%-85.5% compared to AdaBoost. Shapley additive explanations (SHAP) analysis identified CO, SO2, and O3 as the most influential factors. Partial dependence plots (PDPs) showed SO2 had the strongest impact below 40 μg/m³, while NO2 exhibited a linear positive correlation with PM2.5 up to 60 μg/m³. Atmospheric pressure and rainfall were negatively correlated with PM2.5, with notable reductions in concentrations under high-pressure conditions and rainfall levels between 0 and 20 mm. Temperature and relative humidity showed complex relationships, with sharp increases in PM2.5 at temperatures between -5°C and 15°C and SHAP values declining for humidity above 90%. Wind speed exhibited a non-linear effect, with minimal influence at higher velocities. The combined effects of different pollutants can be intensified significantly at higher levels. These findings offer valuable guidance for urban air quality management and pollution mitigation strategies.

Investigation of the Effect of Alloying Elements on the Density of Titanium-Based Biomedical Materials Using Explainable Artificial Intelligence

Titanium alloys are widely preferred in the healthcare sector as biocompatible materials due to their superior properties such as low density and exceptional mechanical strength. Their low density provides lightweight solutions, and their density is closer to that of human bone compared to other metallic alloys with similar strength. This similarity facilitates a balanced load distribution between the bone and the implant, enhancing biomechanical compatibility. This study investigates the effects of alloying elements on the density of titanium-based biomedical materials using a computational materials science approach. A total of 72 different compositions of Ti-Al-V alloys were modeled using JMatPro software, and their densities were simulated at room temperature (25°C). The simulation produced a comprehensive dataset, which was utilized to train an explainable artificial intelligence (XAI) model. Advanced interpretability techniques, including SHAP (SHapley Additive exPlanations), LIME (Local Interpretable Model-agnostic Explanations), and Partial Dependence Plots (PDP), were employed to elucidate the influence of each alloying element on the density. The dataset was analyzed using an XAI-based regression model implemented with the Artificial Neural Network (ANN) algorithm. The interpretability graphs provided insights into the individual contributions of the alloying elements, revealing their positive or negative effects on the density. The findings offer a deeper understanding of the role of alloying elements in optimizing the performance of titanium-based biomedical materials, particularly in achieving lightweight designs. This study highlights the potential of integrating computational material modeling with explainable AI to advance the design and development of high-performance lightweight materials for biomedical applications.

Daily power generation forecasting for a grid-connected solar power plant using transfer learning technique

Deep learning is efficiently used for photovoltaic power generation forecasting to handle the intermittent nature of solar energy. However, big data are required for training deep networks which are not available for newly installed plants. Therefore, in this study, a novel strategy is proposed to train a deep learning model using a transfer learning technique to cop up with the unavailability of enough training datasets. A new 400 kWp solar power plant installed in the Himalayan region is considered as a case study to evaluate the proposed model. The proposed approach utilizes solar radiation data to train a deep neural network and then fine-tune the model using the power generation data from the plant. The network architecture is optimized using grey wolf optimizer to find the best suitable model for the data. The evaluation results show that the same model can achieve higher performance in generation forecasting with percentage error improved by 2% and R-value increased by 7.7% after applying transfer learning. Moreover, SHapley Additive exPlanation and Partial Dependence Plots are used to interpret the model behavior and showed that the model is mostly dependent on the previous generation values (up to 4 days) followed by the temperature and solar radiation.

Prediction of mechanical properties of eco-friendly concrete using machine learning algorithms and partial dependence plot analysis

Rice husk ash concrete (RHAC) shows promise as a beneficial supplementary material in concrete. However, determining mechanical properties such as compressive strength (CS) and splitting tensile strength (STS) of RHAC through conventional lab-scale methods is laborious and time-consuming. In this research, seven important variables were selected as inputs to predict CS and STS using machine learning (ML) models, including Gaussian Process Regression (GPR), Random Forest Regression (RFR), and Decision Tree Regression (DTR) with grid search optimization. The result presented revealed that selected machine learning models provide well accuracy for CS and STS estimates. Among these, the DTR model demonstrated superior performance, with CS prediction R2, RMSE, MAE, and MAPE values of 0.964, 3.314, 2.225, and 5.068, at the testing stage respectively. For STS at the testing stage, DTR achieved R2 of 0.969, RMSE of 0.177, MAE of 0.1322, and MAPE of 3.413. GPR and RFR models also performed well, with R2 values of 0.9434 and 0.9530 for CS prediction. The partial dependence plot (PDP) analysis revealed the optimal mix design parameters for achieving the desired strength. These results offer valuable insights for sustainable construction, allowing engineers to efficiently predict and optimize material performance, reducing the reliance on time-consuming lab methods.

Predicting biomass conversion and COD removal in wastewater treatment by phototrophic bacteria with interpretable machine learning.

Photosynthetic bacteria (PSB) excel in wastewater treatment by removing pollutants and generating biomass but are challenging to optimize due to complex operational and environmental interactions. Neural Ordinary Differential Equations, Elastic Net, Stacking, and Categorical Boosting were applied as artificial intelligence methods to predict chemical oxygen demand (COD) removal efficiency, biomass productivity, biomass yield, and energy yield. Among these, the Stacking model demonstrated superior predictive performance across all targets. Interpretable machine learning methods were employed to identify key features and establish their workable ranges, which included dissolved oxygen (0.3-2.8mgL⁻1), illuminance (2995.3-6000.0 lux), and light energy (20.0-40.0 kWh) for COD removal efficiency; organic loading rate (OLR, 5.7-7.5g COD L⁻1d⁻1), hydraulic retention time (HRT, 0.2-3.2d), and COD concentration (5.3-10.1gL⁻1) for biomass productivity; COD/N ratio (609.0-800.0), OLR (0.1-2.4g COD L⁻1d⁻1), and illuminance (2661-6000 lux) for biomass yield; and pH (6.5-7.9) and HRT (1.2-2.6d) for energy yield. The two-dimensional partial dependence plots revealed that optimal interactions between two key input features resulted in COD removal efficiency >72%, biomass productivity >28gL⁻1d⁻1, biomass yield> 0.96g CODbiomass g CODremoved⁻1, energy yield> 0.49g kWh⁻1. This work advances the understanding of PSB optimization in wastewater treatment through a combination of advanced machine learning and interpretability analysis, offering potential for more efficient resource recovery and process optimization.

Explainable AI Using OBDII Data for Urban Buses Maintenance Management: A Study Case About the DPF Regeneration

Industry 4.0, leveraging tools like AI and the massive generation of data, is driving a paradigm shift in maintenance management. Specifically, in the realm of Artificial Intelligence (AI), traditionally “black box” models are now being unveiled through explainable AI techniques, which provide insights into model decision-making processes. This study addresses the underutilization of these techniques alongside On-Board Diagnostics data by maintenance management teams in urban bus fleets for addressing key issues affecting vehicle reliability and maintenance needs. In the context of urban bus fleets, diesel particulate filter regeneration processes frequently operate under suboptimal conditions, accelerating engine oil degradation and increasing maintenance costs. Due to limited documentation on the control system of the filter, the maintenance team faces obstacles in proposing solutions based on a comprehensive understanding of the system’s behavior and control logic. The objective of this study is to analyze and predict the various states during the diesel particulate filter regeneration process using Machine Learning and explainable artificial intelligence techniques. The insights obtained aim to provide the maintenance team with a deeper understanding of the filter’s control logic, enabling them to develop proposals grounded in a comprehensive understanding of the system. This study employs a combination of traditional Machine Learning models, including XGBoost, LightGBM, Random Forest, and Support Vector Machine. The target variable, representing three possible regeneration states, was transformed using a one-vs-rest approach, resulting in three binary classification tasks where each target state was individually classified against all other states. Additionally, explainable AI techniques such as Shapley Additive Explanations, Partial Dependence Plots, and Individual Conditional Expectation were applied to interpret and visualize the conditions influencing each regeneration state. The results successfully associate two states with specific operating conditions and establish operational thresholds for key variables, offering practical guidelines for optimizing the regeneration process.

Predicting postoperative adhesive small bowel obstruction in infants under 3 months with intestinal malrotation: a random forest approach.

This study aimed to develop a predictive model using a random forest algorithm to determine the likelihood of postoperative adhesive small bowel obstruction (ASBO) in infants under 3 months with intestinal malrotation. A machine learning model was used to predict postoperative adhesive small bowel obstruction using comprehensive clinical data extracted from 107 patients with a follow-up of at least 24 months. The Boruta algorithm was used for selecting clinical features, and nested cross-validation tuned and selected hyper-parameters for the random forest model. The model's performance was validated with 1000 bootstrap samples and assessed using receiver operating characteristic (ROC) analysis, the area under the ROC curve (AUC), sensitivity, specificity, precision, and F1 score. The random forest model demonstrated high diagnostic accuracy with an AUC of 0.960. Significant predictors of ASBO included pre-operative white blood cell count (pre-WBC), mechanical ventilation (MV) duration, surgery duration, and post-operative albumin levels (post-ALB). Partial dependence plots showed non-linear relationships and threshold effects for these variables. The model achieved high sensitivity (0.805) and specificity (0.952), along with excellent precision (0.809) and a robust F1 score (0.799), indicating balanced recall and precision performance. This study presents a machine learning model to accurately predict postoperative ASBO in infants with intestinal malrotation. Demonstrating high accuracy and robustness, this model shows great promise for enhancing clinical decision-making and patient outcomes in pediatric surgery.

Development and validation of machine learning models for MASLD: based on multiple potential screening indicators.

Multifaceted factors play a crucial role in the prevention and treatment of metabolic dysfunction-associated steatotic liver disease (MASLD). This study aimed to utilize multifaceted indicators to construct MASLD risk prediction machine learning models and explore the core factors within these models. MASLD risk prediction models were constructed based on seven machine learning algorithms using all variables, insulin-related variables, demographic characteristics variables, and other indicators, respectively. Subsequently, the partial dependence plot(PDP) method and SHapley Additive exPlanations (SHAP) were utilized to explain the roles of important variables in the model to filter out the optimal indicators for constructing the MASLD risk model. Ranking the feature importance of the Random Forest (RF) model and eXtreme Gradient Boosting (XGBoost) model constructed using all variables found that both homeostasis model assessment of insulin resistance (HOMA-IR) and triglyceride glucose-waist circumference (TyG-WC) were the first and second most important variables. The MASLD risk prediction model constructed using the variables with top 10 importance was superior to the previous model. The PDP and SHAP methods were further utilized to screen the best indicators (including HOMA-IR, TyG-WC, age, aspartate aminotransferase (AST), and ethnicity) for constructing the model, and the mean area under the curve value of the models was 0.960. HOMA-IR and TyG-WC are core factors in predicting MASLD risk. Ultimately, our study constructed the optimal MASLD risk prediction model using HOMA-IR, TyG-WC, age, AST, and ethnicity.

Prediction of outpatient rehabilitation patient preferences and optimization of graded diagnosis and treatment based on XGBoost machine learning algorithm.

The Department of Rehabilitation Medicine is key to improving patients' quality of life. Driven by chronic diseases and an aging population, there is a need to enhance the efficiency and resource allocation of outpatient facilities. This study aims to analyze the treatment preferences of outpatient rehabilitation patients by using data and a grading tool to establish predictive models. The goal is to improve patient visit efficiency and optimize resource allocation through these predictive models. Data were collected from 38 Chinese institutions, including 4,244 patients visiting outpatient rehabilitation clinics. Data processing was conducted using Python software. The pandas library was used for data cleaning and preprocessing, involving 68 categorical and 12 continuous variables. The steps included handling missing values, data normalization, and encoding conversion. The data were divided into 80% training and 20% test sets using the Scikit-learn library to ensure model independence and prevent overfitting. Performance comparisons among XGBoost, random forest, and logistic regression were conducted using metrics, including accuracy and receiver operating characteristic (ROC) curves. The imbalanced learning library's SMOTE technique was used to address the sample imbalance during model training. The model was optimized using a confusion matrix and feature importance analysis, and partial dependence plots (PDP) were used to analyze the key influencing factors. XGBoost achieved the highest overall accuracy of 80.21% with high precision and recall in Category 1. random forest showed a similar overall accuracy. Logistic Regression had a significantly lower accuracy, indicating difficulties with nonlinear data. The key influencing factors identified include distance to medical institutions, arrival time, length of hospital stay, and specific diseases, such as cardiovascular, pulmonary, oncological, and orthopedic conditions. The tiered diagnosis and treatment tool effectively helped doctors assess patients' conditions and recommend suitable medical institutions based on rehabilitation grading. This study confirmed that ensemble learning methods, particularly XGBoost, outperform single models in classification tasks involving complex datasets. Addressing class imbalance and enhancing feature engineering can further improve model performance. Understanding patient preferences and the factors influencing medical institution selection can guide healthcare policies to optimize resource allocation, improve service quality, and enhance patient satisfaction. Tiered diagnosis and treatment tools play a crucial role in helping doctors evaluate patient conditions and make informed recommendations for appropriate medical care.

Development and validation of an explainable machine learning prediction model of hemorrhagic transformation after intravenous thrombolysis in stroke.

To develop and validate an explainable machine learning (ML) model predicting the risk of hemorrhagic transformation (HT) after intravenous thrombolysis. We retrospectively enrolled patients who received intravenous tissue plasminogen activator (IV-tPA) thrombolysis within 4.5 h after symptom onset to form the original modeling cohort. HT was defined as any hemorrhage on head CT scan completed within 48 h after IV-tPA administration. We utilized the Random Forest (RF), Multilayer Perceptron (MLP), Adaptive Boosting (AdaBoost), and Gaussian Naive Bayes (GauNB) algorithms to develop ML-HT models. The models' predictive performance was evaluated using confusion matrix (including accuracy, precision, recall, and F1 score), and discriminative analysis (area under the receiver-operating-characteristic curve, ROC-AUC) in the original cohort, followed by validation in an independent external cohort. The models' explainability was assessed using SHapley Additive exPlanations (SHAP) global feature plot, SHAP Summary Plot, and Partial Dependence Plot. A total of 1,007 patients were included in the original modeling cohort, with an HT incidence of 8.94%. The RF-based ML-HT model showed metrics of 0.874 (accuracy), 0.972 (precision), 0.890 (recall), 0.929 (F1 score); with ROC-AUC of 0.7847 in the original cohort and 0.7119 in the external validation cohort. The MLP model showed 0.878, 0.967, 0.989, 0.978, 0.7710, and 0.6768, respectively. The AdaBoost model showed 0.907, 0.967, 0.989, 0.978, 0.7798, and 0.6606, respectively. The GauNB model showed 0.848, 0.983, 0.598, 0.716, 0.6953, and 0.6289, respectively. The explainable analysis of the RF-based ML model indicated that the National Institute of Health Stroke Scale (NIHSS) score, age, platelet count, and atrial fibrillation were the primary determinants for HT following IV-tPA thrombolysis. The RF-based explainable ML model demonstrated promising predictive ability for estimating the risk of HT after IV-tPA thrombolysis and may have the potential to assist the clinical decision-making in emergency settings.

Quantification of the Influencing Factors of Stand Productivity of Subtropical Natural Broadleaved Forests in Eastern China Using an Explainable Machine Learning Framework

Natural broadleaf forests (NBFs) are the most abundant zonal vegetation type in subtropical regions. Understanding the mechanisms influencing stand productivity in NBFs is important for developing “nature-based” solutions for climate change mitigation. However, minimal research has captured the effects of nonlinearities and feature interactions that often have nonlinear impacts on stand productivity and influencing factors. To address this research gap, we used continuous forest inventory data, and a machine learning model for stand productivity of NBFs was constructed. Subsequently, through leveraging the interpretable machine learning framework of the SHapley Additive explanation (SHAP) and partial dependence plot, we determined global and local explanations of the influencing factors of stand productivity. Our findings indicate the following: (1) The Autogluon model performed the strongest based on R2, RMSE, and rRMSE metrics. (2) The basal area (BA), neighborhood comparison of diameter at breast height (NC), and stand age (AGE) were the key influencing factors. Stand productivity increased with increasing BA and decreased with increasing NC and AGE. BA was maintained above 15 m2ha−1 and NC was maintained below 0.45, which represent favorable conditions for NBFs to maintain optimal growth. (3) SHAP interaction values were calculated to determine the effects of the five major interactions on stand productivity. Our study provides a reference for the sustainable management of NBFs, thereby highlighting the important role of forests in mitigating climate change.

Identification of optimal portal pressure decrease to control ascites while minimizing HE after TIPS: A multicenter study.

Clinically significant portal hypertension in patients with liver cirrhosis can lead to refractory ascites. A TIPS treats clinically significant portal hypertension but may cause overt hepatic encephalopathy (oHE). Our aim was to determine the optimal reduction of the portal pressure gradient (PPG) through TIPS to control ascites without raising oHE risk. This multicenter study screened 1509 patients from 3 European centers (Hannover, Vienna, and Hamburg) undergoing TIPS implantation between 2000 and 2023. Patients with TIPS indications other than refractory ascites/hepatic hydrothorax, vascular liver disease, HCC, or insufficient PPG data were excluded. PPG was measured before and after TIPS insertion. Outcome data were assessed up to 1 year after TIPS insertion. Analyses were conducted utilizing a modern machine learning model, namely a competing-risk random survival forest, partial dependence plots, and competing risk analyses with liver transplantation/death as competitors. The cohort was divided into a 60% derivation and 40% validation cohort. Overall, 729 patients (median MELD: 13 [IQR 10-16], 66% male, 23% oHE before TIPS) were analyzed. The derivation cohort comprised 438 patients, and the validation cohort comprised 291 patients. The optimal PPG reduction, determined by maximally selected Gray statistic and PDP of the random survival forest, was 60%-80%. In this range, patients showed significantly fewer hepatic decompensations due to ascites (HDA) (subdistribution hazard ratio [sHR]: 0.7 [0.52-0.96]) with similar oHE incidences (sHR: 0.92 [0.67-1.27]). The PPG range was confirmed in the validation cohort (HDA: sHR: 0.66 [0.46-0.96]; oHE: sHR: 0.89 [0.61-1.32]). A targeted PPG reduction of 60%-80% showed significantly reduced HDA without increased oHE risk. Therefore, PPG reduction within this range could be a valid reduction target.

InterDIA: Interpretable Prediction of Drug-induced Autoimmunity through Ensemble Machine Learning Approaches.

Drug-induced autoimmunity (DIA) is a non-IgE immune-related adverse drug reaction that poses substantial challenges in predictive toxicology due to its idiosyncratic nature, complex pathogenesis, and diverse clinical manifestations. To address these challenges, we developed InterDIA, an interpretable machine learning framework for predicting DIA toxicity based on molecular physicochemical properties. Multi-strategy feature selection and advanced ensemble resampling approaches were integrated to enhance prediction accuracy and overcome data imbalance. The optimized Easy Ensemble Classifier achieved robust performance in both 10-fold cross-validation (AUC value of 0.8836 and accuracy of 82.81%) and external validation (AUC value of 0.8930 and accuracy of 85.00%). Paired case studies of hydralazine/phthalazine and procainamide/N-acetylprocainamide demonstrated the model's capacity to discriminate between structurally similar compounds with distinct immunogenic potentials. Mechanistic interpretation through SHAP (SHapley Additive exPlanations) analysis revealed critical physicochemical determinants of DIA, including molecular lipophilicity, partial charge distribution, electronic states, polarizability, and topological features. These molecular signatures were mechanistically linked to key processes in DIA pathogenesis, such as membrane permeability and tissue distribution, metabolic bioactivation susceptibility, immune protein recognition and binding specificity. SHAP dependence plots analysis identified specific threshold values for key molecular features, providing novel insights into structure-toxicity relationships in DIA. To facilitate practical application, we developed an open-access web platform enabling batch prediction with real-time visualization of molecular feature contributions through SHAP waterfall plots. This integrated framework not only advances our mechanistic understanding of DIA pathogenesis from a molecular perspective but also provides a valuable tool for early assessment of autoimmune toxicity risk during drug development.

Predicting the risk of cardiovascular disease in adults exposed to heavy metals: Interpretable machine learning.

Machine learning exhibits excellent performance in terms of predictive power. We aimed to construct an interpretable machine learning model utilizing National Health and Nutrition Examination Survey data to investigate the relationship between heavy metal exposure and cardiovascular disease (CVD). A total of 4600 adults were included in the analysis. The Least Absolute Shrinkage and Selection Operator regression method was employed to select relevant feature variables. Subsequently, six machine learning models were constructed, including random forest, decision tree, gradient boosting decision tree, k-nearest neighbor, support vector machine, and AdaBoost algorithms. Feature importance analysis, partial dependence plot, and shapley additive explanations were integrated to enhance the interpretability of the CVD prediction model. Among all models, the random forest exhibited the best performance, with an accuracy of 90 %, an area under the curve of 0.85, and an F1 score of 0.86. Urine cadmium (Cd), blood lead (Pb), urine thallium (Tl), and urine tungsten (W) were identified as the most significant predictors of CVD, with importance scores of 0.062, 0.057, 0.051, and 0.050, respectively. At the overall level, higher levels of urine Cd, blood Pb, and urine W were associated with an increased risk of CVD, whereas a lower level of urine Tl was linked to a reduced CVD risk. Additionally, the analysis of synergistic effects revealed that Cd was the predominant determinant of CVD risk. The random forest-based CVD prediction model demonstrated excellent predictive power and provided valuable insights for personalized patient care and optimal resource allocation in populations exposed to heavy metals.

Maritime accident prediction in busan port using machine learning: An integrated approach with maritime accident reports and VTS data

This study develops and evaluates Machine Learning (ML) models for predicting maritime accidents, uniquely incorporating both environmental and Vessel Traffic Services (VTS) features. Using data from Busan Port spanning November 2018 to July 2024, we compared XGBoost, Random Forest, Neural Network, and Support Vector Machine models.The dataset comprised 84 maritime accident/near-miss and 1000 non-accident points, with data augmentation applied to improve model generalization. The Random Forest model performed best with 94.50 % accuracy, while all models exceeded 90 %.Feature importance analysis revealed that 3 of the top 10 features were VTS-related, highlighting their significance in accident prediction. VTS operator (VTSO) rank and VTS sectors significantly impacted accident risk, with Class 2 VTSOs associated with higher risk and Sector 3 showing lower risk. Partial Dependence Plot analysis (PDP) visualized each feature's impact on predictions and revealed inter-model differences in feature interpretation.These findings offer valuable insights for maritime safety policies, vessel operations management, and VTS improvements, laying groundwork for real-time maritime accident prediction systems.