ABSTRACT Accurate maps of rice paddy and cropping intensity at the high spatial resolution are crucial for rice production estimates and food security, yet are inadequate for the entire East and Southeast Asia. Here, we proposed a novel algorithm to map rice paddies and cropping intensity, integrating phenology and machine learning approaches with multi-source remote sensing data. Specifically, by generating random training samples via a buffer approach within X-Means clustering of Sentinel-2 time series, we identified rice paddies and cropping intensity using flooding-transplanting and tillering-heading signals. Multiple Random Forest classifiers were then combined to produce 10-m resolution rice paddy and cropping intensity maps for East and Southeast Asia in 2023. Our rice paddy and cropping intensity maps achieved an overall accuracy of 95% and 91%, respectively, based on 102,075 validation samples collected through field surveys, visual interpretation, and multi-source rice datasets. The mapped rice paddy areas exhibited significant linear correlations with official statistics, with correlation coefficients (r) of 0.95 at both national and provincial scales. Compared to existing rice paddy maps, our method yields superior reliability and accuracy in large-scale rice paddy extraction, effectively reducing commission errors associated with dense, small water bodies.

To the best of our knowledge, this study provides the first systematic comparison between classical regression and advanced machine learning models for predicting the profitability of Moroccan firms listed on the Casablanca Stock Exchange. While prior research has largely focused on developed markets, profitability prediction in emerging economies such as Morocco remains underexplored, despite the market’s structural particularities (sectoral concentration, reliance on bank financing, and limited disclosure practices). This article provides the first systematic comparative analysis between regression and machine learning approaches applied to Moroccan listed companies, highlighting the advantages and limitations of each method in capturing complex and non-linear financial dynamics. Using a dataset covering ten years of financial statements, we evaluate multiple models, including OLS, Ridge regression, Random Forest, Gradient Boosting, Support Vector Regression, KNN, and XGBoost. Results show that machine learning models consistently outperform regression in predictive accuracy, while regression retains value in interpretability. Findings contribute to academic research by extending profitability forecasting studies to an under-explored emerging market, and to practice by offering investors, policymakers, and managers tools that improve risk assessment, capital allocation, and decision-making under conditions of uncertainty. These implications are particularly relevant for emerging economies, where informational asymmetries and structural heterogeneity complicate financial forecasting.

Although there are numerous studies exploring predictors of clinical trial failure, no comprehensive review of their methodological specificities and findings exists. We performed a scoping review with the aim of exploring the methodological approaches and findings of studies analysing predictors of clinical trial failure. The Ovid Medline and Embase databases were systematically searched from inception to December 13, 2024, for studies employing frequentist statistics or machine learning (ML) approaches to assess predictors of trial failure across multiple clinical trials. A generalized linear model (GLM) was employed to assess the impact of certain methodological factors (failure and non-failure definitions, study types included and trial phases included) on reported failure proportions. To estimate the effects of the predictors included in the model on failure proportions, odds ratios (OR) with 95% confidence interval (95% CI) were calculated from model coefficients. The literature search identified 17,961 records, 81 of which were included in the review. Most of the studies used Clinicaltrials.gov data (73 studies, 90.1%). Frequentist statistics were used to analyze predictors of trial failure in 73 studies (90.1%), and remaining 8 studies employed ML techniques (9.9%). The GLM showed a 27.5% deviance reduction, indicating that certain methodological factors substantially contribute to observed differences in failure proportions. Studies including trials with both completed and ongoing statuses when calculating failure proportions had lower odds of failure compared to those just including completed statuses (OR = 0.44, 95% CI: 0.29-0.67, p < 0.001). There has been a recent expansion of ML approaches, potentially signaling the beginning of a paradigm shift. Methodological variations substantially influence reported failure proportions, implicating the need for adoption of standardized definitions of failure and calculation approach. We recommend categorizing terminated and withdrawn studies as failed and completed ones as non-failed.

Accurate prediction of real estate prices remains a major challenge due to dynamic market conditions and the limitations of traditional valuation methods. Empirical studies that directly compare human experts, machine learning (ML) models, and hybrid approaches are rare. This study examines the predictive accuracy and efficiency of an XGBoost-based ML model, real estate experts, and a hybrid human-machine approach. A model was trained using 21,736 real estate transactions from Vienna (2018-2022). We then conducted an experimental procedure with 13 experts who evaluated newly built apartments sold in 2023 under three conditions: limited information, state-of-the-art expert methods, and collaboration between experts and ML model. The results show that the ML model achieves accuracy comparable to that of experts while significantly reducing the time required for the task. Within the hybrid approach, experts were able to achieve the highest accuracy in comparison to other methods. These results underscore the potential of human-ML collaboration.

Developing efficient catalysts for proton exchange membrane fuel cells (PEMFCs) requires minimising platinum (Pt) usage while enhancing oxygen reduction reaction (ORR) activity. While density functional theory (DFT) calculations provide valuable structural insights, understanding complex structure-property relationships often requires extensive manual analysis. We apply the Self-Consistent Attention Neural Network (SCANN) to analyze DFT data for Pt clusters (Pt 4 , Pt 13 ) on single-wall carbon nanotubes (SWNTs) with adsorbed O 2 . SCANN accurately predicts deformation energies (MAE < 1 meV/atom) and frontier orbital energies (MAE < 12 meV), while its attention scores suggest that p–d hybridisation between Pt and SWNT atoms may influence adsorption behaviour. Our minimum energy path analysis examines O–O bond dissociation, with attention scores revealing potential electronic redistribution patterns that could influence this critical catalytic step. These computational findings demonstrate how interpretable machine learning approaches may advance the understanding and design of Pt–SWNT catalytic systems for sustainable energy applications.

There has been a notable increase in artificial intelligence (AI) studies in dentistry. However, the inadequate use of proper validation methods has led to overly optimistic performance metrics of machine learning (ML) models. External validation provides evidence of a ML model's performance with independent datasets and is crucial for generalizability. We developed Extreme Gradient Boosting (XGBoost) models to detect dental caries using easy-to-collect questionnaire data. ML model training was conducted using cross-validation nested resampling with a holdout test set, utilizing NHANES datasets (n = 6070). Performance of the trained model was tested using external data from the Northern Finland Birth Cohorts (NFBC1966 and NFBC1986; n = 3616). To enhance interpretability, beeswarm plots were constructed to visualize variable importance. The ML model demonstrated acceptable performance in predicting dental caries on the internal dataset, with an area under the operating characteristics curve (AUC) of 0.785 (95% CI 0.756-0.813). However, the model encountered difficulties in identifying participants with dental caries, as shown by its poor sensitivity of 0.391, despite achieving a high specificity of 0.919. When applied to the external dataset, the ML model encountered significant challenges, with the AUC dropping to 0.550 (95% CI 0.532-0.569), sensitivity decreasing to 0.053, and specificity slightly improving to 0.974. Important variables identified by the model were self-rated condition of teeth and gums, presence of missing teeth, financial status, and time since last dental visit. The performance of our ML model during external validation degraded notably compared to the internal validation. However, the XAI methodology exhibited great potential to be used in the future for individualized dental caries risk assessment.

The long-term performance of concrete-filled steel tubes (CFSTs), particularly creep at the member level and the structural-level creep effect, poses significant challenges to full-life cycle design. Conventional finite-element methods (FEMs) entail high computational costs and exhibit strong parameter dependency. For improvement, we propose a deep learning-based model for predicting CFST creep by leveraging the capabilities of a long short-term memory (LSTM) neural network and its improved versions. The predictions from the machine learning model were compared with experimental results and those obtained from FEM based on the Kelvin chain viscoelastic model. By comparing the performance of various machine learning approaches and FEM in predicting CFST creep, a reliable and efficient method is proposed to accurately predict the long-term creep behavior of CFSTs. Some suggestions are obtained: (1) The hyperparameters of all models were obtained by optimization algorithm. The improved LSTM model outperforms traditional machine learning algorithms and FEM in predicting CFST creep, and the CNN-LSTM-Attention model achieves the highest accuracy, with an R 2 of 0.92. (2) The prediction accuracy of the CNN-LSTM-Attention model was significantly improved by increasing the data acquisition frequency and sample size. Compared to smaller datasets, when the sample size was increased to 12,960, the R 2 of this model was raised from 0.92 to 0.96. (3) The future trend of CFST creep was predicted using the optimal CNN-LSTM-Attention model, and the prediction shows that the creep deformation rate gradually decreased, and the creep values tend to stabilize over the following 60 days.

Plant-pollinator interactions span a continuum from strict specialisation to generalisation and most flowers are visited by more than a single functional group of pollinators. However, one functional group might be more efficient than the others and thus exert stronger selective pressure on floral traits. In this study we aim at identifying the evolutionary drivers of floral syndromes in the genus Aquilegia. We analyse floral syndromes using multivariate statistics, morphospace analyses, as well as a machine learning approach (random forests), testing for the association between floral traits and documented pollinators for 28 Aquilegia species. In particular, we test whether pollen-collecting pollinators (small bees, large bees, syrphid flies) and nectar-collecting pollinators (large bees, hummingbirds, hawkmoths) are associated with specific floral traits. Furthermore, we test whether mixed pollination systems are reflected in floral syndrome properties. Our results indicate that floral syndromes in Aquilegia are mainly shaped by nectar-collecting pollinators (and not by pollen-collecting pollinators). Flowers pollinated by large bees are mostly pendent and short-spurred; hummingbird flowers are red, with constricted spurs and short petal blades; and hawkmoth flowers are erect with long and slender spurs. Flowers pollinated by two groups of nectar-collecting pollinators show syndromes corresponding to only one of their pollinator groups. Despite their ubiquity, we did not find cues for selection by any of the pollen-collecting pollinators. Nevertheless, selection for traits associated with pollen-collecting pollinators, such as openly accessible stamens and a contrasted yellow floral centre (almost always present in Aquilegia), cannot be ruled out. In conclusion, floral syndromes in Aquilegia are associated with nectar-collecting pollinators only, maybe because they are more efficient at pollinating, which remains to be tested in field experiments.

Spontaneous premature birth (SPTB) is a common pregnancy complication; however, few studies have explored cell senescence-related markers in SPTB. Bioinformatics and machine learning approaches were used to predict potential biomarkers associated with SPTB. Normal and SPTB gene expression profiles were obtained from the Gene Expression Omnibus (GEO) database, and cell senescence-associated genes from the Human Aging Genomic Resources (HAGR) database. Functional enrichment analysis and protein-protein interaction (PPI) network analysis of differentially expressed senescence-related genes in SPTB were conducted using Gene Ontology (GO), Kyoto Encyclopedia of Genes and Genomes (KEGG), and STRING databases. The infiltration of 22 types of immune cells in SPTB was calculated using the CIBERSORT deconvolution algorithm. Machine learning methods were employed to identify hub differentially expressed genes (DEGs). Datasets GSE174415 and GSE118442 were extracted for validation to determine the final hub genes. Additionally, receiver operating characteristic (ROC) curves were constructed to assess the diagnostic potential of the hub genes, and significant pathways associated with the final hub gene were explored by Gene Set Enrichment Analysis (GSEA). Finally, real-time quantitative polymerase chain reaction (RT-qPCR) was performed to validate the hub gene in clinical specimens. A total of 923 DEGs were identified, including 525 upregulated and 398 downregulated in the SPTB group. These 923 genes were intersected with 866 cell senescence-related genes, yielding 48 intersection genes. Functional enrichment analysis indicated that these intersection genes were primarily associated with cytokine–cytokine receptor interactions and the PI3K-Akt signaling pathway. The expression of activated dendritic cells and follicular helper T cells was significantly lower in the SPTB group compared to the full-term pregnancy group. A total of six hub genes, LGALS3, ESR1, PLA2G2A, TWIST1, CBS, and PLA2R1, were identified by machine learning. According to dataset validation, TWIST1 was identified as the final hub gene. TWIST1 was downregulated in placental tissues of the SPTB group and demonstrated high diagnostic value for SPTB. Thus, TWIST1 may be a novel molecular target for predicting and diagnosing SPTB, providing diagnostic value and novel insights into this condition.

Abstract Event log data frequently contain incomplete and incorrect entries due to errors or hardware malfunctions during the logging process. Poor event log quality hinders the application of business process analysis methods. Traditional approaches to handling business process event logs involve imputing the affected values i.e. replacing them with reliable estimations. To this end, machine learning approaches have been proposed, mainly deep neural network-based ones. Interestingly, data imputation for arbitrary time series encompasses a broader range of techniques, categorized into matrix-based, profile-based, and neural network-based methods. However, such techniques are not readily applicable to business process logs due to the fact that the latter consist of timestamped sequences of varying length and semantics, e.g., the same position in the trace may correspond to a different activity. Our contribution is firstly to show that dealing with the varying length problem using state-of-the-art zero padding based autoencoders is suboptimal and then to propose a pre-processing grouping framework, where imputation of missing values is based solely on the members of the same group. We introduce three variants of groupings, with one of them being counter-intuitive in the sense that it leverages logs of different sequential patterns, but is particularly effective. More importantly, the manner in which we approach the problem gives rise to a novel trade-off, namely accuracy vs. coverage, which has not been investigated before. The presented solutions are evaluated against a baseline imputation method and a state-of-the-art methodology, and the decrease in errors reached 3 to 10 times.

Genomic selection (GS) is a promising strategy for accelerating genetic gains of complex traits in breeding programs. Despite the recent advancements in high-throughput genotyping technologies, the selection of the type of marker systems needed for GS remains challenging in breeding programs. In this study, we explored 3K array single nucleotide polymorphisms (SNPs) and genotyping by sequencing (GBS) SNP markers for genomic prediction of oat biomass yield using different statistical and machine learning approaches. An oat panel consisting of 420 lines was phenotyped for biomass-related traits for 3 years and genotyped using two different marker platforms (3K array and GBS). Our results showed similar performance of both the 3K array and GBS-based SNPs in terms of training population optimization, forward prediction, and univariate and multivariate genomic prediction of forage yield. The genomic best linear unbiased prediction (GBLUP), Bayes-B, and random forest models gave similar predictive ability for dry matter yield (DMY) in different harvest-year combinations and for both marker platforms. The multivariate models involving various combinations of secondary traits (simple breeders' field notes and data) resulted in more than twofold increases in predictive abilities compared to the univariate models. Comparison of the 25% top-performing observed and predicted genotypes showed a higher overlap percentage (30.10%-66.99%) for multivariate GBLUP models compared to the univariate models (27.18%-51.46%). This further elucidates the great potential of multivariate GS models incorporating the more robust and easily reproducible 3K array SNP markers for improving the genetic gains of DMY in breeding programs.

Full-field displacement reconstruction in structural health monitoring using machine learning approach: case study with experimental validation

ObjectiveThis study aimed to identify diagnostic biomarkers for keloid and explore potential therapeutic agents from traditional Chinese medicine (TCM) by integrating network pharmacology approaches. Specifically, we sought to uncover key molecular targets for Aloe vera and validate their roles in keloid pathogenesis.MethodsWe integrated keloid transcriptome datasets (GSE218007 and GSE237752) by merging GEO data, and identifying differentially expressed genes (DEGs). Functional enrichment analysis (GO, GSEA) and machine learning approaches were applied to select diagnostic biomarkers. Candidate genes were validated via Receiver Operating Characteristic (ROC) curves in training and independent cohorts (GSE44270). PPI networks and Cytohubba algorithms identified hub genes, while TCMSP-screened compounds from Aloe vera were docked with targets using molecular docking.Results91 Identified DEGs enriched in fibrosis-related pathways. Machine learning prioritized two diagnostic biomarkers: AMPH and TNFRSF9 (AUC > 0.85 in training/testing). PPI analysis revealed IL6 as a hub gene. Aloe vera-derived quercetin targeted HAS2 and IL6 (both P < 0.05 in validation), with molecular docking confirming stable binding (binding energy <−7 kcal/mol). IL6 emerged as both a key network hub and a therapeutic target, linking keloid and TCM mechanisms.ConclusionAMPH and TNFRSF9 are promising diagnostic biomarkers for keloid, while quercetin from Aloe vera targets HAS2 and IL6, offering therapeutic potential. The dual role of IL6 underscores its centrality in keloid pathogenesis, connecting bioinformatics predictions with TCM pharmacology. This study provides a foundation for clinical prediction and targeted treatment strategies.

Abstract The abundance of large datasets has driven machine learning (ML) model performance and scalability breakthroughs. However, many domains and practical applications must contend with the limitations imposed by small and very small datasets. This survey thoroughly examines state-of-the-art methodologies and challenges in ML approaches tailored for scenarios where data scarcity is a fundamental constraint. We begin by outlining the theoretical foundations that govern learning from small data. Then, we discuss recent advancements in data-related frameworks (i.e., training and evaluation methods, etc.) and algorithmic architectures (meta and transfer learning). We also explore the trade-offs and related issues inherent in designing models for small data, such as overfitting, generalization error, and the bias-variance dilemma, as well as identify minimal interventions that can overcome such issues. Further, this survey covers the role of synthetic data generation and simulation-based approaches to enlarge data availability while critically assessing the implications of these techniques on model performance. Finally, in synthesizing open literature, we shed light on emerging trends/research directions that aim to overcome challenges arising from limited data, such as incorporating domain knowledge and causal principles to guide the learning process and integrating symbolic reasoning with statistical learning.

Federated learning (FL) is an efficient approach for large-scale distributed machine learning that promises data privacy by keeping training data on client devices. However, recent research has uncovered vulnerabilities in FL, impacting both security and privacy through poisoning attacks and the potential disclosure of sensitive information in individual model updates as well as the aggregated global model. This paper explores the inadequacies of existing FL protection measures when applied independently, and the challenges of creating effective compositions.Addressing these issues, we propose WW-FL, an innovative framework that combines secure multi-party computation (MPC) with hierarchical FL to guarantee data and global model privacy. One notable feature of WW-FL is its capability to prevent malicious clients from directly poisoning model parameters, confining them to less destructive data poisoning attacks. We furthermore provide a PyTorch-based FL implementation integrated with Meta’s CrypTen MPC framework to systematically measure the performance and robustness of WW-FL. Our extensive evaluation demonstrates that WW-FL is a promising solution for secure and private large-scale federated learning.

The rapid growth of mobile applications in Ethiopia has generated a wealth of user-generated content in the form of app reviews and ratings. These reviews provide critical insights into user satisfaction, app performance, and feature demands. However, systematic analysis of such unstructured and multilingual feedback remains limited in Ethiopia due to the absence of automated tools and localized natural language processing (NLP) resources. This study introduces a multi-task review mining framework that integrates sentiment classification, feedback categorization, and rating prediction. A dataset of 10,200 Ethiopian mobile app reviews collected from the Google Play Store was preprocessed, annotated, and analyzed using both machine learning and deep learning models. Experimental results indicate that convolutional neural networks (CNNs) outperformed other models, achieving 98.7% accuracy for sentiment classification, 96.6% for feedback categorization, and an R2 of 0.40 for rating prediction. Among traditional models, XGBoost demonstrated strong performance, particularly in classification tasks. The findings highlight the effectiveness of CNN-based models in extracting actionable insights from multilingual reviews, offering developers and policymakers data-driven tools to improve app quality and enhance user satisfaction. This study contributes to the growing field of opinion mining in low-resource contexts and aligns with Ethiopia’s Digital Transformation 2025 agenda.

Gestational diabetes mellitus (GDM) is characterized by glucose intolerance during pregnancy, resulting from insulin resistance, and is associated with increased maternal and neonatal risks. Inflammasomes play a critical role in GDM pathophysiology by driving immune dysregulation and chronic inflammation. This study aimed to identify inflammasome-related genes, which may serve as potential diagnostic markers and contribute to GDM pathogenesis. RNA sequencing datasets from the GEO were merged and analysed to identify differentially expressed genes (DEGs). The key genes identified by LASSO logistic regression, random forest, Boruta, and SVM-RFE algorithms were superimposed with genes related to inflammasomes. In addition, the CIBERSORT algorithm was used to analyse the immune characteristics of GDM. The expression of inflammasome-related genes was validated in the clinical samples by qPCR and correlated with clinical parameters. PANX1 was identified as a key inflammasome-associated gene, with its diagnostic potential confirmed by ROC curve analysis with an AUC of 93.75%. CIBERSORT and correlation analyses confirmed a significant association between PANX1 expression and immune cell infiltration. qPCR validation using placental samples from GDM and control subjects revealed elevated PANX1 expression in GDM. Furthermore, a significant correlation was observed between PANX1 expression and levels of glucose, HbA1c, and HOMA-IR, suggesting that elevated PANX1 may be associated with metabolic dysregulation in GDM. This study highlights PANX1 as a novel diagnostic biomarker and a therapeutic target in GDM pathophysiology.

Background The usual assessment for axillary lymph node (ALN) status in breast cancer (BC) in current clinical practice is based on an invasive procedure that has a low efficiency rate and frequently results in operative-associated problems for patients. Therefore, our goal was to create an effective preoperative ultrasound (US) radiomics evaluation method for ALN status in patients with clinical stages T1–2 invasive BC using machine learning (ML) approaches. Methods Between January 2020 and January 2024, we retrospectively analyzed the medical records of 671 patients with histologically proven malignant breast tumors in our hospital.The data set was divided into model training group and validation testing group with a 75/25 split.There were two categories for ALN tumor burden: low (1–2 metastatic ALNs) and high (≥ 3 metastatic ALNs). The PyRadiomics package was used to obtain radiomic features (RF), and a support vector machine (SVM) with the LASSO approach was used to create a radiomic signature (RS).The training group’s multivariate logistic regression results were used to create a nomogram that combined the BC US radiomics score with a clinical parameter.Additionally, the area under the operating characteristic curve (AUC) was used to evaluate their prediction performance. Results With an AUC of 0.920 (95% CI: 0.901, 0.943) in the test cohort, clinical parameter coupled RS provides the greatest diagnostic performance in predicting ALN status between disease-free axilla and any axillary metastases.In the testing cohort, the sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) were 90%, 82%, 83%, 89%, and 86%, respectively. With an AUC of 0.939 (95% CI: 0.892, 0.970) in the test cohort, this clinical measure paired with RS can also distinguish between a low and a substantial metastatic burden of axillary illness. Conclusions For patients with early-stage BC, our work provides a noninvasive imaging biomarker to forecast the extent of ALN metastases.The imaging biomarker demonstrated strong predictive value and the potential for extended application to customize surgical care.

ABSTRACT In additive manufacturing (AM) processes, in-situ monitoring combined with machine learning (ML) approaches plays a crucial role in ensuring consistent product quality and preventing defects. However, existing ML methods for anomaly detection predominantly rely on correlation-based models that lack interpretability and fail to capture underlying spatiotemporal and causal dynamics. This study proposes an anomaly detection framework that integrates spatiotemporal dependency learning (STL) and Granger causality learning (GCL) through graph attention network mechanisms. The STL module enforces spatial consistency and temporal smoothness in learned feature representations, while the GCL module identifies causal relationships between historical process signatures and both historical and current parameters, and current states through attention-based causal aggregation and disentanglement techniques. By combining these complementary modules, our method achieves superior anomaly detection performance while providing interpretable insights through spatial-temporal dependency interpretation, causal disentanglement analysis, and causal attribution analysis. Experimental validation demonstrates improved detection accuracy compared to existing baselines, with attention-based mechanisms enabling the identification of specific process parameters and spatial regions contributing to anomalous behaviour. This framework facilitates proactive quality control in AM processes by bridging the gap between high-accuracy anomaly detection and practical interpretability requirements in manufacturing applications.

Climate change increasingly threatens global food security through disrupted precipitation patterns and extreme weather events, requiring resilient systems for assessing agricultural vulnerability. This study developed and compared machine learning approaches for predicting cropland vulnerability using Earth Observation data, operationalized through NDVI anomalies as a defensible biophysical metric. We employed both a dual-stream deep learning architecture and a Random Forest model to predict 2023 NDVI anomalies across Bangladesh croplands using a 22-year time series (2001–2023) of MODIS vegetation indices, ERA5 climate variables, and static environmental covariates. A spatially aware block cross-validation strategy ensured rigorous, independent performance evaluation. Results demonstrated that the Random Forest model (R2 = 0.70, RMSE = 197.03) substantially outperformed the deep learning architecture (R2 = 0.02, RMSE = 357.57), explaining 70% of cropland stress variance and enabling early detection of vulnerable areas three months before harvest. Feature importance analysis identified recent climate variables, March precipitation, February NDVI, and vapor pressure deficit as primary vulnerability drivers. Spatial analysis revealed distinct vulnerability patterns, with Natore and Magura districts exhibiting elevated stress consistent with 2023 drought conditions, threatening the productivity of the region’s critical irrigation-dependent rice cultivation. These findings demonstrate that simpler, interpretable models can sometimes outperform complex architectures while providing useful information for early warning systems and precision targeting of climate adaptation interventions.