Taxi-out time prediction of departure flights based on Stacking and SHAP

The increasing frequency of extreme events poses a significant threat to engineering safety and socio-economic stability. Traditional prediction and assessment models exhibit notable limitations when dealing with complex and dynamic environments, whereas machine learning methods, leveraging their powerful data processing capabilities and adaptability, demonstrate broad application prospects. However, existing research still lacks comprehensive solutions when applying machine learning to complex engineering systems, particularly in the risk assessment and prediction of extreme events. In light of this, this study focuses on flood risk assessment and constructs an integrated prediction model based on multiple machine learning algorithms. Through correlation analysis, cluster analysis, ensemble learning, and neural network enhancement, this study identifies key influencing factors of flood occurrence and establishes flood risk evaluation and probability prediction models. Shapley Additive exPlanations (SHAP) value analysis reveals that policy factors and coastal vulnerability contribute most significantly to the model output, with SHAP importance values exceeding 35 and approaching 30, respectively, indicating their critical impact on flood risk prediction. Furthermore, the CatBoost-based early warning evaluation model achieves an accuracy of 0.9346 and an F1 score of 0.9246 on the test set, while the random forest model achieves an MSE of 0.000153 and an MAE of 0.00172. Additionally, the neural network-enhanced random forest model maintains high prediction accuracy even when using only five key indicators. These results demonstrate the significant advantages of machine learning methods in flood risk assessment and prediction, providing reliable support for engineering safety and disaster prevention and mitigation.

Tree-based machine learning (ML) algorithms, such as Extra Trees (ET), Random Forest (RF), Gradient Boosting Machine (GBM), and XGBoost (XGB) are among the most widely used in early drug discovery, given their versatility and performance. However, models based on these algorithms often suffer from misclassification and reduced interpretability issues, which limit their applicability in practice. To address these challenges, several approaches have been proposed, including the use of SHapley Additive Explanations (SHAP). While SHAP values are commonly used to elucidate the importance of features driving models' predictions, they can also be employed in strategies to improve their prediction performance. Building on these premises, we propose a novel approach that integrates SHAP and features value analyses to reduce misclassification in model predictions. Specifically, we benchmarked classifiers based on ET, RF, GBM, and XGB algorithms using data sets of compounds with known antiproliferative activity against three prostate cancer (PC) cell lines (i.e., PC3, LNCaP, and DU-145). The best-performing models, based on RDKit and ECFP4 descriptors with GBM and XGB algorithms, achieved MCC values above 0.58 and F1-score above 0.8 across all data sets, demonstrating satisfactory accuracy and precision. Analyses of SHAP values revealed that many misclassified compounds possess feature values that fall within the range typically associated with the opposite class. Based on these findings, we developed a misclassification-detection framework using four filtering rules, which we termed "RAW", SHAP, "RAW OR SHAP", and "RAW AND SHAP". These filtering rules successfully identified several potentially misclassified predictions, with the "RAW OR SHAP" rule retrieving up to 21%, 23%, and 63% of misclassified compounds in the PC3, DU-145, and LNCaP test sets, respectively. The developed flagging rules enable the systematic exclusion of likely misclassified compounds, even across progressively higher prediction confidence levels, thus providing a valuable approach to improve classifier performance in virtual screening applications.

Detecting new or worsening hypomobility in acute ischemic cerebrovascular patients is challenging, especially when asleep or unattended. This study used a wearable movement acceleration monitoring system to identify changes in these patients, aiming to improve early detection. Continuous bilateral upper limb acceleration data, clinical characteristics, specific treatments, stroke etiology, and in-hospital outcomes were collected from patients. The primary outcome was newly emerging or worsening hypomobility during monitoring. An XGBoost model, trained with synthetic minority oversampling to address class imbalance and validated via 5-fold cross-validation, analyzed movement acceleration features to diagnose hypomobility timing. Model performance was evaluated through AUC and feature importance metrics. From April 2023 to February 2025, 85 patients with acute ischemic cerebrovascular events were enrolled; three were excluded due to data errors. A total of 82 patients were included in the analysis, comprising 76 (92.7%) with ischemic stroke and 6 (7.3%) with transient ischemic attack. Among them, 26 (31.7%) were women. The median monitoring duration was 26.9 hours (IQR: 24.6–46.5), with 15 patients (18.3%) developing hypomobility. The XGBoost model achieved an AUC of 0.975 (95% CI: 0.965–0.985) and a mean AUC of 0.975 (SD 0.003) across folds. Optimized with a learning rate of 0.1, maximum depth of 6, and 200 boosting rounds, the model, at a cutoff of 0.587, recorded an average sensitivity of 0.969 and specificity of 0.900, accurately detecting 96.9% of the hypomobility cases. The overall metrics included a sensitivity of 0.966, specificity of 0.900, positive predictive value of 0.896, negative predictive value of 0.968, and F1-score of 0.930. The SHAP (SHapley Additive exPlanations) analysis revealed the significant contributions of the interaction terms (mean |SHAP| = 3.475) and slope features for movement changes (e.g., 1-min RSMA and LSMA slopes), while elevating the importance of the ‘Likely weak side’ predictor (mean |SHAP| = 2.053) in orienting asymmetry. This wearable movement acceleration monitoring system, by continuously tracking upper limb acceleration data, effectively detects the onset of hypomobility in acute ischemic cerebrovascular patients, highlighting its substantial potential for clinical application in enabling timely interventions and improving patient outcomes.

BackgroundTraumatic rib fractures can lead to respiratory complications necessitating unplanned intubation, but predictors have been inadequately delineated. We used interpretable machine learning to predict unplanned intubations in rib fracture patients while identifying predictors.MethodsTQIP 2017-2022 was queried for adult patients admitted to the hospital following a rib fracture injury. An XGBoost model was developed to predict unplanned intubation using variables that can be known on admission. A 70/10/20 train/validation/test split was used. SHapley Additive exPlanations (SHAP) were used for interpretation. SHAP allows individualized interpretation of predictors for each patient.ResultsThe cohort had 905615 patients; 2.3% had unplanned intubations. Model metrics at the F1 maximizing threshold (0.78) included AUROC = 0.83, F1 score = 0.17, accuracy = 0.94, precision = 0.12, recall = 0.29, specificity = 0.95, and Brier score = 0.17. The most influential variables, as determined by mean absolute SHAP values, were admission location (0.62), Injury Severity Score (0.40), age (0.37), absence of comorbidities (0.18), pulse rate (0.14), pneumothorax (0.13), oxygen saturation (0.15), chronic obstructive pulmonary disease (0.11), respiratory rate (0.10), and sex (0.10). ICU admission was the location most influential in predicting an unplanned intubation. SHAP dependency plots determined the directional relationship between variables' values and SHAP values.DiscussionPatients above the F1 maximizing threshold had a 7.4-fold increase in unplanned intubations compared to those below. Nearly 30% of all unplanned intubations were captured at this threshold. Our model's identification of these high-risk patients and influential factors not previously considered in the literature could guide closer monitoring and early interventions.

Class imbalance in vehicle payment method classification can cause predictive models to become biased toward the majority. This study aims to build a classification model for automotive consumer payment methods using Extreme Gradient Boosting (XGBoost), with class balancing handled through the Synthetic Minority Over-sampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN), and model interpretability performed using SHAP (SHapley Additive Explanations). The dataset consisted of 11,011 records and 13 attributes derived from Toyota vehicle delivery order transactions. Results show that the XGBoost model without balancing achieved 67.37% accuracy but only 0.24 recall for the Cash class. After applying SMOTE, the recall for the Cash class improved to 0.58, while ADASYN produced a similar improvement at 0.59, with overall accuracy maintained at around 61–62% and a stable ROC-AUC of 0.65. Feature importance and SHAP analysis identified c_vehicle_model and c_city as the most influential factors in predicting the payment method. From a business perspective, the improved ability to detect cash customers reduces the risk of misclassification and enables dealers to better segment customer payment preference. This supports more effective marketing campaigns, sales strategies, and financing risk management. The combination of XGBoost, SMOTE, ADASYN, and SHAP has proven effective in handling imbalanced data while offering transparent interpretability of predictions, making it a practical foundation for data-driven decision-making in the automotive industry.

Cardiometabolic multimorbidity (CMM) is a prevalent syndrome among middle-aged and older adults, significantly impairing quality of life and imposing substantial health and economic burdens on China's aging healthcare system. The development of timely predictive models is crucial for enabling early intervention. This study aimed to integrate multidimensional data from the China Health and Retirement Longitudinal Study (CHARLS) to develop an effective predictive model for assessing the five-year risk of CMM onset, thereby facilitating early intervention and management for individuals at risk in China. We analyzed data from the 2015 to 2020 CHARLS surveys, involving 5,388 middle-aged and older adults initially free of CMM. The dataset was randomly split into a training set (70%) and a validation set (30%). Key predictors were identified from 31 potential variables using LASSO regression with 10-fold cross-validation. Selected predictors underwent correlation analysis and were used to construct both Extreme Gradient Boosting (XGBoost) and Logistic Regression (LR) models. Variable contributions in the XGBoost model were interpreted using SHapley Additive exPlanations (SHAP) values, while the LR model was visualized via a nomogram. The performance of the superior model was evaluated using receiver operating characteristic curves, calibration curves, and decision curve analysis. LASSO regression screened significant variables from the initial 31 candidates. Correlation analysis ultimately identified nine key predictors: systolic blood pressure, BMI, fasting blood glucose, total cholesterol, triglycerides, uric acid, age, comorbidities, and pain. The logistic regression model demonstrated superior and more stable predictive performance on the validation set, with an area under the curve (AUC) of 0.732 (95% CI: 0.703-0.761). Calibration curves indicated reliable predictive accuracy, and decision curve analysis established the clinical net benefit across a range of risk thresholds. This study developed and validated a reliable, clinically applicable logistic regression model, supplemented by a nomogram, for assessing the five-year risk of CMM in Chinese middle-aged and older adults. The model effectively identifies high-risk individuals, supporting targeted early intervention and management strategies to alleviate the health and economic burdens associated with CMM in this population.

The application of Artificial Intelligence (AI) and Machine Learning (ML) models is increasingly leveraged to automate and optimize Data Centre (DC) operations. However, the interpretability and transparency of these complex models pose critical challenges. Hence, this paper explores the Shapley Additive exPlanations (SHAP) values model explainability method for addressing and enhancing the critical interpretability and transparency challenges of predictive maintenance models. This method computes and assigns Shapley values for each feature, then quantifies and assesses their impact on the model’s output. By quantifying the contribution of each feature, SHAP values can assist DC operators in understanding the underlying reasoning behind the model’s output in order to make proactive decisions. As DC operations are dynamically changing, we additionally investigate how SHAP can capture the temporal behaviors of feature importance in the dynamic DC environment over time. We validate our approach with selected predictive models using an actual dataset from a High-Performance Computing (HPC) DC sourced from the Enea CRESCO6 cluster in Italy. The experimental analyses are formalized using summary, waterfall, force, and dependency explanations. We delve into temporal feature importance analysis to capture the features’ impact on model output over time. The results demonstrate that model explainability can improve model transparency and facilitate collaboration between DC operators and AI systems, which can enhance the operational efficiency and reliability of DCs by providing a quantitative assessment of each feature’s impact on the model’s output.

Flowering time is a fundamental trait that determines crop adaptation and yield stability. To accurately predict flowering time and identify key regulatory factors, it is necessary to extract biologically meaningful signals from high-dimensional and multi-omics datasets. Although machine learning has been increasingly applied in plant genomics, there is still limited research on how feature selection (FS) methods and genomic prediction (GP) models affect the prediction of flowering time and gene discovery, particularly regarding the combination of different FS and GP approaches and the interpretability of prediction models. To address this gap, we conducted a large-scale benchmarking study that jointly evaluated seven feature selection methods and six prediction models, resulting in 42 FS-GP combinations. By integrating SNP and transcriptomic data, we assessed predictive performance and further interpreted model outputs using SHAP (SHapley Additive exPlanations) within a random forest (RF) framework to quantify feature contributions. This strategy successfully identified known flowering time regulators in maize, including ZmMADS69 and ZmRap2.7, and revealed additional candidate genes potentially involved in the flowering regulatory network. Overall, this study offers valuable insights into the genetic regulation of flowering time in maize and provides an effective framework for discovering candidate genes from multi-omics data for crop improvement.

The objective of this study is to predict the occurrence of postpartum hemorrhage in women with placenta previa based on machine learning. This retrospective study enrolled 845 singleton pregnant patients with placenta previa from two hospitals. They were allocated into a training cohort (n = 403), a testing cohort (n = 174), and the external validation cohort (n = 268). Univariate and multivariate regression analyses were employed to select clinical variables (p < 0.05), which were subsequently utilized to develop 11 machine learning prediction models. The area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), accuracy (ACC), sensitivity (SEN), and specificity (SPE) were used to evaluate the performance of the models. Besides, SHapley Additive exPlanations (SHAP) was used to interpret the role and effectiveness of variables in the predictive model. Three machine learning models with the best predictive performance were combined into a Prediction Ensemble Classifier through voting. The Gradient Boosting Machine demonstrated the best predictive performance. In the validation cohort, AUC of the Gradient Boosting Machine model is 0.810(95% CI 0.754-0.865), ACC was 0.765(95% CI 0.716-0.813), SEN was 0.613(95% CI 0.513-0.723), while these values of the Prediction Ensemble Classifier were 0.813(0.756-0.871), 0.806(0.757-0.854), and 0.480(0.375-0.597), respectively. The importance of SHAP variables in the model, ranked from high to low, is as follows: d-dimer, ultrasound diagnosis of placenta accreta spectrum, neutrophils, prothrombin time, and platelets. The Gradient Boosting Machine model demonstrated excellent performance in predicting postpartum hemorrhage in cases of placenta previa. Furthermore, SHAP analysis enabled interpretation of the variables in the model.

IntroductionThis study aimed to develop a diabetic retinopathy (DR) Prediction model using various machine learning algorithms incorporating the novel predictor Triglyceride-glucose index (TyG). Furthermore, the model was interpreted using the SHapley Additive exPlanations (SHAP) method.MethodReal-world data were collected from a general hospital in a major city and a county clinic, then divided into the DR Group (1392) and non-DR group (2358). Baseline data were collected, and variables were selected using Recursive Feature Elimination with Cross-Validation (RFECV). The performance of five machine learning algorithms, including Logistic Regression model (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and XGBoost (XGB), was assessed based on accuracy, sensitivity, specificity, and Area Under the Curve (AUC) of the Receiver Operating characteristic Curve (ROC). The optimal model was interpreted using SHAP.ResultLVM and LR demonstrated superior performance in both the test set and training set (ROC, 0.85 and 0.82, respectively). The top five predictors identified by SHAP analysis included TyG, Insulin therapy, HbA1c, Diabetes Course, HDL. HDL was identified as a protective factor, while the remaining factors were associated with retinopathy.ConclusionLR and SVM demonstrated the best performance. To our knowledge, this is the first machine learning-based DR prediction model integrating the triglyceride-glucose index (TyG) as a core predictor, overcoming limitations of insulin resistance (IR) assessment in resource-limited settings. TyG provides a cost-effective alternative to conventional IR biomarkers (e.g., HOMA-IR), enabling practical DR risk stratification in primary care.

Endoscopic disease activity monitoring is important for the long-term management of patients with ulcerative colitis (UC), there is currently no widely accepted non-invasive method that can effectively predict endoscopic disease activity. We aimed to develop and validate machine learning (ML) models for predicting it, which are desired to reduce the frequency of endoscopic examinations and related costs. The patients with a diagnosis of UC in two hospitals from January 2016 to January 2021 were enrolled in this study. Thirty nine clinical and laboratory variables were collected. All patients were divided into four groups based on MES or UCEIS scores. Logistic regression (LR) and four ML algorithms were applied to construct the prediction models. The performance of models was evaluated in terms of accuracy, sensitivity, precision, F1 score, and area under the receiver-operating characteristic curve (AUC). Then Shapley additive explanations (SHAP) was applied to determine the importance of the selected variables and interpret the ML models. A total of 420 patients were entered into the study. Twenty four variables showed statistical differences among the groups. After synthetic minority oversampling technique (SMOTE) oversampling and RFE variables selection, the random forests (RF) model with 23 variables in MES and the extreme gradient boosting (XGBoost) model with 21 variables in USEIS, had the greatest discriminatory ability (AUC = 0.8192 in MES and 0.8006 in UCEIS in the test set). The results obtained from SHAP showed that albumin, rectal bleeding, and CRP/ALB contributed the most to the overall model. In addition, the above three variables had a more balanced contribution to each classification under the MES than the UCEIS according to the SHAP values. This proof-of-concept study demonstrated that the ML model could serve as an effective non-invasive approach to predicting endoscopic disease activity for patients with UC. RF and XGBoost, which were first introduced into data-based endoscopic disease activity prediction, are suitable for the present prediction modeling.

The main goal of this research is the interpretability of deep learning (DL) model output (e.g., CNN and LSTM) used to map land susceptibility to subsidence hazard by means of different techniques. For this purpose, an inventory map of land subsidence (LS) is prepared based on fieldwork and a record of LS presence points, and 16 features controlling LS were mapped. Thereafter, 11 effective features controlling LS were identified by means of a particle swarm optimization (PSO) algorithm, which was then used as input in the CNN and LSTM predictive models. To address the inherent black box nature of DL models, six interpretation methods (feature interaction, permutation importance plot (PFIM), bar plot, SHapley Additive exPlanations (SHAP) main plot, heatmap plot, and waterfall plot) were used to interpret the predictive model outputs. Both models (CNN and LSTM) had AUC > 90 and therefore provided excellent accuracy for mapping LS hazard. According to the most accurate model-the CNN predictive model-the range from very low to very high hazard classes occupied 20%, 20%, 25%, 16.3%, and 18.7% of the study area, respectively. According to three plots (bar plot, SHAP main plot, and heatmap plot), which were constructed based on the SHAP value, distance from the well, GDR and DEM were identifiedas the three most important features with the highest impact on the DL model output. The results of the waterfall plot indicate two effective features consisting of distance from the well and coarse fragment, and two effective features comprising landuse and DEM, contributed negatively and positively to LS, respectively. Overall, these explanation techniques can address critical concerns with respect to the interpretability of sophisticated DL predictive models.

Coronary heart disease (CHD) and carotid stenosis share similar pathogenic mechanisms such as chronic inflammation, metabolic dysregulation, oxidative stress, and immune imbalance. These conditions are often comorbid. This study aimed to develop and validate an effective machine learning model that uses routine blood biomarkers to predict CHD risk in patients with carotid stenosis. Clinical data from 723 patients diagnosed with carotid artery stenosis between January 2019 and December 2024 were retrospectively collected, including demographic characteristics, hematological and biochemical laboratory parameters, and their derived composite indices. We used six feature selection methods and 11 machine learning (ML) algorithms to construct predictive models, and we systematically compared their performance. To interpret the established predictive model, we applied SHapley Additive exPlanations (SHAP) analysis to visualize and elucidate the risk prediction framework. We identified the intersection of six feature selection methods, screened a total of nine potential predictors related to CHD, and used them to construct a prediction model. Cross-validation results demonstrated that the predictive model based on the random forest (RF) algorithm achieved the best performance among all evaluated algorithms (AUC = 0.800; sensitivity = 0.729; specificity = 0.792; F1 score = 0.733). The SHAP plots for RF indicate that estimating residual cholesterol (RC), fasting plasma glucose (FPG), and monocyte-to-lymphocyte ratio (MLR) are the three most important features for predicting CHD. Significant differences were observed in SHAP values for key biomarkers (including RC, MLR) across varying degrees of carotid artery stenosis. The RF model achieved the best predictive performance in predicting CHD in patients with carotid stenosis. Elevated metabolic and immunoinflammatory markers significantly enhanced the predictive power of the model. Prospective multicenter validation is warranted to confirm its generalizability across diverse populations.

The issue of weak interpretability in geological disaster susceptibility assessments using machine learning models has been a long-standing concern. Although SHAP (Shapley Additive Explanations) models have been extensively used in recent years to interpret the decision-making details of models, the specialized skills required and the non-intuitiveness of SHAP plots make their application challenging in practical decision-making environments. In response, our study introduces a map-based SHAP visualization framework to enhance the interpretability of susceptibility assessment results. Utilizing Optuna for hyperparameter tuning, we developed a high-performance XGBoost model to assess the susceptibility of the most impactful disaster in Hangzhou: urban road collapses. In addition to interpreting the contributions of evaluation factors through traditional SHAP summaries and bar plots, we displayed the SHAP values for each evaluation factor using map visualizations, and discussed the model’s sensitivity to different values. To validate the alignment between model predictions and physical collapse mechanisms, our study selected typical collapse cases, interpreted these cases combining map visualizations, SHAP force plots at collapse points, and the physical mechanisms of collapse. Our research improves the interpretability of susceptibility assessments with machine learning by using map visualizations, providing new insights into spatial effects and robust support for urban decision-making applications.

Landslide susceptibility assessment (LSA) plays a pivotal role in regional disaster prevention, particularly in southeastern Tibet, where frequent landslides pose significant threats to human safety and critical infrastructure. However, current LSA approaches face two key challenges: (1) the absence of standardized guidelines for selecting appropriate slope unit scales, which may result in over−smoothing or excessive noise in spatial patterns; and (2) the limited interpretability of machine learning models, which hampers understanding of factor contributions. This study investigates the scale effects of slope unit delineation on LSA in the Yuqu River Basin. Using the r.slopeunits method, six datasets at varying scales were generated to capture terrain heterogeneity. An XGBoost−based framework was applied for susceptibility modeling, with SHAP (Shapley Additive Explanations) used to enhance model interpretability. Results indicate that slope unit scale substantially affects sample distribution, feature representation, and model performance. At the smallest scale (c = 0.05), excessive data redundancy and imbalanced class ratios reduced accuracy (AUC = 0.824). At the largest scale (c = 0.5), spatial heterogeneity was over−smoothed, also impairing performance (AUC = 0.832). The intermediate scale (c = 0.3) performed best, yielding a balanced representation and a mean AUC of 0.856. SHAP analysis highlighted freezing index, relative height, and rainfall as the most influential factors. Notably, susceptibility increased significantly when the freezing index ranged between 1500 and 3000 °C·d and relative height between 500 and 1500 m. Additionally, interactions—such as between the freezing index and slope gradient or fault density—further intensified landslide risk, underscoring the need to consider nonlinear dependencies. By integrating multi−scale modeling with SHAP−based interpretation, this study enhances both the predictive accuracy and transparency of LSA.