Ensemble Learning Research Articles

Traditional deep learning models such as convolutional neural networks (CNNs), which capture localized features, and long short-term memory networks (LSTMs), which focus on long-term dependencies, often face challenges in achieving higher accuracy for time series prediction tasks. To address this limitation, this study proposes a hybrid deep learning model that integrates CNN, LSTM, the reptile search algorithm (RSA), and eXtreme Gradient Boosting (XGB) for pollutant concentration forecasting. Initially, the raw pollutant concentration data undergoes cleaning and normalization via a Min-Max scaler. The processed sequences are then separately fed into LSTM and CNN models to extract weighted features. RSA is applied to optimize these features, while XGB computes feature importance scores, quantifying the contribution of each selected feature to the predictive performance. The proposed model predicts pollutants such as [Formula: see text], CO, SO[Formula: see text], and NO[Formula: see text] up to ten days in advance for urban Indian settings. Comparative evaluations against benchmark models-including Transformer, CNN, BiLSTM, BiRNN, ANN, and BiGRU-demonstrate that the hybrid approach yields consistently superior accuracy and robustness. The hybrid model achieves substantially lower errors and higher [Formula: see text] scores across all pollutants, validating its reliability for long-horizon air quality forecasting.

In agricultural systems, soil pH and electrical conductivity (EC) are crucial chemical properties that directly affect nutrient availability and microbial activity, but the challenging environment of the Peruvian Andes has limited research on their estimation. This study aimed to develop an ensemble learning method to predict soil pH and EC in Andean agroecosystems using environmental predictors. By using simple and weighted averaging, we developed a heterogeneous ensemble learning approach that integrates machine learning (ML) algorithms, including Support Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The weighted ensemble assigns weights to models based on their predictive accuracy, measured by R² from spatial cross-validation. Spatial patterns are noticeable, and pH displays greater spatial clustering than EC. Elevation was the most important predictor in ML models for both parameters. Ensemble models significantly outperformed individual models, with the weighted ensemble achieving R² >0.93 and reducing RMSE by approximately 72%. Among standalone models, RF and XGBoost performed best for pH, while SVM performed the best for EC. ANN models were the least effective. Uncertainty analysis indicated high confidence in pH predictions but moderate to high uncertainty in EC predictions, suggesting that EC is more challenging to predict. Ensemble models with optimized weighting provide robust and accurate mapping of spatially autocorrelated soil properties. The high-confidence pH maps are reliable for soil management decisions, while EC predictions, though more uncertain, effectively identify priority areas for future sampling and investigation.

Introduction Artificial intelligence (AI) has emerged as a valuable tool in mental health care, with applications in the treatment of psychosis. However, its application to psychosocial functioning in psychosis remains underexplored, despite its critical role towards long-term therapeutic outcomes and recovery. The goal of this systematic review and meta-analysis is to identify, summarize, and evaluate the current evidence on AI applications in psychosocial functioning in psychosis. Methods A literature search was conducted using the PubMed, Scopus, and ACM Digital Library databases for articles published between January 2010 and March 2025, in accordance with the PRISMA guidelines. Quality of studies was assessed using the Prediction model Risk Of Bias Assessment Tool (PROBAST), Newcastle-Ottawa Scale (NOS), and the Cochrane Risk of Bias Tool (RoB2.0). Meta-analyses synthesized commonly used performance metrics using random-effects models, with subgroup, sensitivity and publication bias analyses. Results A total of 14 studies were included in this review. Various AI techniques were employed, with supervised machine learning being the most predominant. Psychosocial domains, including social function, occupational function, social cognition and quality of life, were targeted. Meta-analysis revealed moderate discriminative and predictive accuracies of AI models: pooled AUC of 0.70 (95% CI: 0.63–0.76) and RMSE of 8.15 (95% CI: 7.32–8.98). Subgroup analyses indicated higher predictive accuracy for social cognition (AUC=0.77) and clinical symptom-based predictors (RMSE=7.1), with substantial heterogeneity mainly attributed to methodological variability. Conclusions This review discovered the current application of AI in psychosocial functioning in psychosis, including the techniques usage, modeling approaches, targeted domains, and model performance. AI showed promise for early identification, continuous monitoring, and personalized interventions, driven by methodological advances such as ensemble learning with feature selection. Nevertheless, limitations in methodological consistency, data quality, model design, and ethical issues underscore that the field remains in its early stages. Overall, AI should complement clinical expertise, rather than replace it, in delivering psychosocial care in psychosis. Systematic review registration https://www.crd.york.ac.uk/prospero/ , identifier CRD420251051952.

This paper presents a wavelet-based framework for 3-D sign language recognition that outperforms conventional ZCA approaches by preserving critical spatiotemporal relationships in hand kinematics. We formulate hand dynamics as a wavelet exponent energy problem using Daubechies transforms, where joint coordinates, mother coefficients, and window sizes are jointly optimized, avoiding PCA’s disruptive axis rotation while maintaining interpretable feature orientations. The proposed method combines this wavelet energy representation with an ensemble deep learning feature selection layer using binary masking ( λ 1 = 0.10 , λ 2 = 0.01 ) and regularization to eliminate redundancy, achieving 94.09% accuracy (vs. 92.75% for PCC+ZCA (SVD)) with 4.3 × faster inference on Turkish SL datasets. Statistical validation (ANOVA, p < 0.001 ) confirms significant improvements over ZCA-derived features ( t = 24.15 ), with weight analysis revealing proximal arm kinematics as dominant recognition factors. The framework’s computational efficiency (0.12ms inference) and preservation of kinematic relationships demonstrate its superiority for real-life SLR applications compared to both PCC and ZCA approaches.

Predicting Customer Demand for MSE Loan Products: A Bilevel Dynamic Selective Ensemble Learning Method

Timely identification of plant diseases plays a vital role in protecting crop yield and supporting effective decision-making in precision agriculture. Conventional computer vision models achieve high recognition accuracy but often require substantial computing power, making them impractical for low-cost edge hardware widely used in rural areas. In this work, a compact deep learning ensemble is presented, combining three lightweight convolutional neural networks—MobileNetV3-Small, EfficientNet-B0, and ShuffleNetV2—with a Vision Transformer (ViT-B/16). The models operate in parallel, and their outputs are merged using a weighted late-fusion approach, with fusion weights determined through systematic grid search to achieve the best trade-off between predictive performance and processing speed. The Plant Village dataset, consisting of 54,303 images from 38 healthy and diseased leaf categories, was used for evaluation. To improve robustness, the training data were augmented through geometric transformations, contrast adjustment, and controlled noise addition. When tested on a Raspberry Pi 4 device, the ensemble reached an accuracy of 97.85%, precision of 97.67%, recall of 97.92%, and F1-score of 97.79%, with an average inference time of 20.5 ms and a total size of 14.6 MB. These results surpassed those of all individual models and conventional machine-learning baselines. Statistical testing using McNemar’s method confirmed the significance of the improvement (p 0.05). Precision–Recall analysis indicated strong resistance to false positives, while accuracy–latency assessment confirmed suitability for real-time field operation. The proposed system offers a practical, resource-efficient framework for on-site plant disease diagnosis in areas with limited connectivity and computing resources. Further development will focus on adaptation to field-captured imagery, hardware-aware model compression, and the integration of additional sensing modalities such as hyperspectral and thermal imaging.

Letter to the editor "Predicting intraoperative blood loss risk in severe lumbar disc herniation patients undergoing PLIF: a multicenter cohort study using ensemble learning".

Introduction: Right ventricular failure (RVF) is a major adverse event following left ventricular assist device (LVAD) implantation. The complex mechanisms involved make it challenging to accurately predict RVF. Although supervised machine learning is useful for predicting complex outcomes, it is often difficult to identify specific factors that increase a patient's risk. This study aimed to assess the risk of RVF in individual patients and identify their unique risk factors using supervised machine learning. Methods: Between June 2010 and January 2024, 482 consecutive patients underwent continuous-flow LVAD implantation at Osaka University Hospital or the National Cerebral and Cardiovascular Center. Of them, 326 who underwent preoperative right heart catheterization and echocardiography were included in the analysis. Important features for predicting the risk of RVF were selected using the χ2 or Mann-Whitney U test, the Gini index in a random forest algorithm, and a literature review. The optimal classification algorithm for this analysis was selected from among the random forest, eXtreme Gradient Boosting, support vector machine, logistic regression, and ensemble learning algorithms by comparison of the area under the curve, accuracy, F1 score, and sensitivity through five-fold cross-validation of the test data. The SHapley Additive exPlanations (SHAP) value was used to assess the individual risk factors for RVF. Results: Thirteen important features (sex, age, non-ischemic cardiomyopathy, body surface area, aspartate aminotransferase level, blood urea nitrogen level, left ventricular end-diastolic dimension, left ventricular ejection fraction, right ventricular stroke work index, central venous pressure, pulmonary capillary wedge pressure, pulmonary pulsatility index, and Interagency Registry for Mechanically Assisted Circulatory Support profile) were selected. Ensemble learning was the most reliable classification algorithm. The area under the curve, accuracy, F1 score, and sensitivity were 0.87, 0.89, 0.77, and 0.80, respectively. The SHAP analysis revealed that impaired right ventricular function assessed by right heart catheterization, poor preoperative condition, and a good ejection fraction were associated with an increased risk in most cases. Conclusions: Supervised machine learning enables the accurate prediction of RVF after LVAD implantation, while SHAP values visualize individual risk factors and may optimize preoperative conditions.

Background: While numerous AI models leverage ECGs to predict specific cardiac abnormalities, few have been validated at scale across diverse populations and echocardiographic pathologies. This study evaluates the external performance of a deep learning ensemble that predicts a composite of major echocardiographic abnormalities —including reduced ejection fraction (≤35%), valvular disease (mild to severe stenosis and/or moderate to severe regurgitation of Mitral, Aortic and/or Tricuspid valves), and elevated pulmonary artery pressure—from standard 12-lead ECG images. Methods: We retrospectively analyzed 44,403 ECGs from 43,346 patients (≥15 years) across 14 centers between July and September 2024. The data was split into two cohorts: Co1 included 28,509 ECGs paired with same-day echocardiograms; Co2 included 15,894 ECGs (from 15,401 patients) without same-day echocardiograms. The ensemble model—comprising two InceptionNetV3 and one ResNet50 networks trained on DICOM ECG images—classified ECGs as positive or negative for any one of eight echocardiographic dysfunctions based on the Youden index threshold (0.27) calculated during model development. Results: On Co1, the model achieved an ROCAUC of 84% , PR-AUC of 45% , sensitivity 74% , specificity 79% , PPV 34% , NPV 96% , and overall accuracy of 80% . A robust performance was observed in all subsets (Fig. 1). In the subset of false positives (5,127 patients), 932 patients who had follow-up echocardiograms within a median of 54 days revealed that 28% eventually showed echocardiographic dysfunction. In Co2, out of the 15,401 patients, 2,185 patients underwent at least one echocardiogram within 6 months post-ECG. Among those, of the 830 patients who were previously flagged positive for dysfunction by the model, 49% were found to have dysfunction, with a median interval of 28 days between the initial ECG and echo. Conclusion: This AI-enhanced ECG model demonstrates strong generalizability and clinical utility for early identification of echocardiographic abnormalities, offering a scalable solution to triage echocardiography referrals across varied clinical settings, requiring about 2.9 echos to confirm 1 case flagged as dysfunction by the model. Prospective validation is underway to further evaluate its impact on diagnostic workflows and patient outcomes.

Background: Accurate assessment of aortic stenosis (AS) severity is critical to reduce associated morbidity and mortality. Current evaluation relies on expert interpretation of B-mode and Doppler echocardiography across multiple views. This study aimed to develop and validate a deep learning model for automated AS severity assessment using multi-view B-mode and color Doppler echocardiographic videos. Methods: We designed a two-stage framework for automated AS assessment. First, six video-based convolutional neural networks were trained to classify AS severity from distinct echocardiographic views: B-mode [parasternal long-axis (PLAX), parasternal short-axis (PSAX), apical three-chamber and five-chamber (AP)] and color Doppler [PLAX-color, PSAX-color, and AP-color]. Next, outputs from these models were integrated using a machine learning ensemble (HistGradientBoostingClassifier) to produce a study-level AS severity classification. Performance was assessed using area under the receiver operating characteristic curve (AUC) on held-out test set from Kaiser Permanente (KP) and Stanford Health Care (SHC). Results: The models were trained on 213,814 videos from 16,076 studies at KP, and evaluated on internal test data (KP; 23,492 videos from 1,789 studies) as well as external test data (SHC; 13,278 videos from 1,238 studies). The multi-view ensemble model demonstrated strong performance, achieving a macro-AUC of 0.891 (95% CI: 0.881–0.901; Figure 1) on the KP test dataset. Generalizability was confirmed on the external SHC cohort with an AUC of 0.929 (95% CI: 0.918–0.939; Figure 2). Conclusion: In this study, we applied a deep learning framework integrating multi-view echocardiographic videos to assess AS severity. The model demonstrated accurate, generalizable performance and highlights the potential of AI-powered decision support tools in echocardiographic evaluation of AS.

Background: Accurate assessment of mitral stenosis (MS) severity is critical to guide timely clinical management. Current evaluation relies on expert interpretation of B-mode and Doppler echocardiography, requiring integration of multiple views and skilled Doppler imaging. This study aimed to develop and validate a deep learning model for automated MS severity assessment using multi-view B-mode and color Doppler echocardiographic videos. Methods: We developed a two-stage framework for automated MS assessment. First, four video-based convolutional neural networks were trained to classify MS severity from distinct echocardiographic views: B-mode [parasternal long-axis (PLAX), apical three-chamber and five-chamber (AP)] and color Doppler [PLAX-color, and AP-color]. Next, outputs from the four models were integrated using a machine learning ensemble (HistGradientBoostingClassifier) to produce a study-level MS severity classification. Performance was assessed using area under the receiver operating characteristic curve (AUC) on held-out test set from Kaiser Permanente (KP) and Stanford Health Care (SHC). Results: The models were trained on 66,714 videos from 3,921 studies at KP, and evaluated on internal held-out test data (KP; 7,344 videos, 438 studies) and external test data (SHC; 29,181 videos, 1,988 studies). The multi-view ensemble model demonstrated strong performance, achieving a macro-AUC of 0.853 (95% CI: 0.828–0.877; Figure 1) on the KP test dataset. Generalizability was confirmed on the external SHC cohort with an AUC of 0.974 (95% CI: 0.968–0.981; Figure 2). Conclusion: This study confirmed the ability for multi-view deep learning models to assess MS severity. The model demonstrated accurate, generalizable performance and highlights the potential of AI-powered decision support tools in echocardiographic evaluation of MS.

Globally, heart disease (HD) persists as a major contributor to mortality rates, requiring accurate and efficient diagnostic models. While machine learning has shown promise in early detection, challenges such as missing data, class imbalance, suboptimal feature selection, and inefficient hyperparameter tuning hinder predictive accuracy and reliability. Many existing models fail to effectively preprocess medical datasets, leading to biased and computationally expensive predictions. To address these issues, this study proposes a strong hybrid framework for HD prediction. The Balanced Imputation-Normalization Framework incorporates K-Nearest Neighbors (KNN) imputation, StandardScaler normalization, and the Synthetic Minority Oversampling Technique (SMOTE). KNN imputation effectively handles missing data, ensuring reliable representation, while StandardScaler normalization standardizes feature values to enhance model stability. SMOTE is applied to address class imbalance, synthetic samples are generated to augment the minority class. Feature selection is optimized using the Hungarian algorithm, which systematically selects the most relevant attributes while reducing redundancy. Additionally, Bayesian optimization fine-tunes hyperparameters to improve classification performance. For prediction, an ensemble learning approach combines Random Forest (RF), Decision Tree (DT), K-Nearest Neighbors (KNN), Naïve Bayes (NB), and Extreme Gradient Boosting (XGBoost). The Voting Ensemble aggregates predictions using hard and soft voting mechanisms, improving robustness and generalization. Experimental results on benchmark heart disease datasets demonstrate that XGBoost attained a peak accuracy of 96.43%, with subsequent results from the Voting Ensemble at 95.66%, significantly outperforming traditional models and demonstrating that ensemble learning effectively improves accuracy and reduces computational complexity.

Predicting crack propagation rate in reinforced concrete structures using classical and ensemble machine learning with SHAP-based interpretability

The multidimensional urban built environment (BE) in high-density cities has been shown to be closely related to the urban vitality (UV) of residents’ travelling. However, existing research lacks consideration of the differences in this relationship over a week, so this paper proposes an ensemble machine learning approach that simultaneously considers different time periods of the week. This study reveals the impacts of four dimensions of BE variables on UV at different time periods at the scale of the community life circle. The four well-performing base models are integrated to reveal the mechanism of differential effects of BE variables on UV under different time periods in the old city of Nanjing through Shapley addition explanation. The findings reveal that (1) the seven most important built environment variables existed in different time periods of the week: floor area ratio, service POI density, remote sensing ecological index, POI mixability, average building height, fractional vegetation cover, and maximum building area; (2) The nonlinear and threshold effects of the built environment factors differed across time periods of the week; (3) There is a dominant interaction between built environment variables at different time periods of the week. This study can provide guidance for the refined management of complex urban systems.

Adverse drug reactions (ADRs) remain a major global challenge, contributing substantially to patient morbidity, mortality, and healthcare costs. Traditional pharmacovigilance approaches—spontaneous reporting and post-marketing surveillance—are hampered by underreporting, delays, and limited contextual data. The growing availability of electronic health records (EHRs), which capture longitudinal structured and unstructured patient information, presents an unprecedented opportunity to advance ADR prediction. This narrative review synthesizes recent progress in developing and validating predictive models that leverage EHRs, highlighting methodological approaches, challenges, and future directions. Predictive strategies range from traditional regression models to advanced machine learning and deep learning architectures, with multimodal frameworks increasingly integrating structured fields (demographics, labs, prescriptions) and unstructured clinical text through natural language processing. While ensemble and deep learning methods demonstrate superior performance, issues of data quality, missingness, bias, and interpretability persist. Robust validation frameworks—spanning internal cross-validation to multi-center external testing—are critical to ensure generalizability and clinical trustworthiness. Ethical considerations, including fairness, privacy, and transparency, remain central to safe deployment. Looking forward, promising avenues include federated learning across institutions, integration of multi-omics and pharmacogenomic data, explainable AI tailored for clinical use, and real-time monitoring through digital twin frameworks. These trajectories, combined with robust governance and clinician–data scientist collaboration, have the potential to transform ADR detection from a reactive process to proactive, personalized prevention. By synthesizing the existing evidence, this review provides insights into the development of more effective predictive models for ADRs and informs strategies for improving pharmacovigilance. This study will contribute to the ongoing efforts to leverage EHRs and predictive models for improving patient outcomes and reducing the burden of ADRs.

Abstract As the primary contributor to the error budget in microwave-dependent geodetic techniques, tropospheric delay has emerged as a bottleneck in achieving precise positioning of Global Navigation Satellite Systems (GNSSs). Owing to the dynamic nature of water vapor, effective global zenith total delay (ZTD) modeling is essential, yet challenging, underscoring the need for robust solutions. Artificial intelligence (AI) methods provide an effective approach for capturing the spatiotemporal characteristics of ZTD for modeling. In this contribution, we employ stacking ensemble learning to develop the machine-learning-based global ZTD (MGZTD) model, which integrates four decision tree models and artificial neural networks, leveraging the global forecast system (GFS) to implement real-time nonlinear compensation for the empirical ZTD model. Validation against ERA5-derived ZTD shows that the model's optimization achieves a 73.2% accuracy improvement over the baseline without correction. When benchmarked against GNSS post-processed ZTD, the stacking algorithm outperforms base learner schemes, particularly in tropical coastal environments. Furthermore, the MGZTD model achieves a global mean RMS of 18.9 mm, yielding improvements of 52.5%, 51.3%, and 51.6% over state-of-the-art GPT3, PVoxel, and IGPZWD models, respectively. The proposed model can effectively simulate the subtle random time-varying and spatially complex distribution of ZTD, delivering accurate and stable prediction information for users. These indicate the promising potential of the MGZTD model in enhancing real-time GNSS wide-area positioning service and water vapor remote sensing.

Mental health problems are increasingly prevalent among the younger generation, particularly those active on social media, yet early detection efforts often remain limited. Previous studies have explored text-based approaches for identifying mental health issues, but many are constrained by low accuracy in differentiating multiple psychological states or lack integration into accessible tools for end-users. This study addresses these gaps by proposing a hybrid machine learning model for early detection of mental health conditions through social media text analysis. Five algorithms were evaluated, and a soft voting ensemble combining Logistic Regression and Support Vector Machine (SVM) was developed to improve classification across five mental states (Anxiety, Depression, Stress, Emotional Exhaustion, and Healthy) and three risk levels (Low, Medium, High). To ensure practical utility, the model was deployed in an Android-based application, SmartRisk, which allows users to input free text and receive automated assessments. The findings show that the proposed hybrid approach significantly improves detection performance, particularly in identifying depression and high-risk cases, while maintaining high usability in real-world application. The novelty of this study lies in combining hybrid ensemble learning with mobile deployment for practical, text-based early detection of mental health, offering both methodological advancement and societal impact.

Retraction Note: Enhancing Financial Risk Prediction for Listed Companies: A Catboost-Based Ensemble Learning Approach

AI-Driven Ensemble Learning for Tropical Cyclone-Induced Storm Surge Prediction along the Guangdong and Hong Kong Coast

Accurate crop yield prediction remains a cornerstone of sustainable agriculture and food security, particularly in regions vulnerable to climate fluctuations such as Sub-Saharan Africa. This study develops an adaptive hybrid ensemble learning model that integrates climatic and soil parameters to improve crop yield prediction accuracy. The proposed framework combines Decision Tree Regressor and Ridge Regression as base learners, while Linear Regression serves as a meta-model to optimize ensemble predictions. A dataset spanning 1990–2020 was analyzed and preprocessed using normalization and feature selection techniques based on agronomic significance. Model optimization was performed using GridSearchCV to fine-tune hyperparameters. Experimental results revealed that the stacking ensemble achieved superior performance, with an RMSE of 0.1318, MAE of 0.0804, and R² of 0.9766, outperforming individual models. The findings underscore the effectiveness of hybrid ensemble methods in modeling nonlinear agricultural systems and demonstrate the potential of machine learning to support data-driven agricultural decision-making. Future work will explore dynamic adaptation to real-time environmental data and regional transferability across diverse agricultural ecosystems.