Infertility has emerged as a significant public health concern, with assisted reproductive technology (ART) is a last-resort treatment option. However, ART's efficacy is limited by significant financial cost and physical discomfort. The aim of this study is to build Machine learning (ML) decision-support models to predict the optimal range of embryo numbers to transfer, using data from infertile couples identified through literature reviews. Binary classification models were developed to classify cases into two groups: those transferring two or fewer embryos and those transferring three or four. Four popular ML algorithms were used, including random forest (RF), logistic regression (LR), support vector machine (SVM), and artificial neural network (ANN), considering seven criteria: the woman's age, sperm origin, the developmental qualities of four potential embryos, infertility duration, assessment of the woman, morphological qualities of the four best embryos on the day of transfer, and number of oocytes extracted. The stratified 3-fold cross-validation results show that the SVM model obtained the highest average accuracy (95.83%) and demonstrated the best overall performance, closely followed by the ANN and LR models with an average accuracy equal to 91.67%. The RF model achieved a slightly lower average accuracy (88.89%), which demonstrated the lowest variability. Testing on a new dataset revealed all models performed well, with ANN and SVM models classified all test set instances correctly, while the RF and LR models achieved 91.68% accuracy. These results highlight the superior generalization and effectiveness of the ANN and SVM models in guiding ART decisions.

Investigating neural dynamics through EEG signals offers valuable insights into brain activity, especially for automated seizure detection. The identification of epileptogenic zones is crucial for effective epilepsy treatment, particularly in surgical planning. This work introduces a novel method for seizure detection using EEG signals, designed to benefit clinicians by integrating spectral features with Long Short-Term Memory (LSTM) networks, enhanced by brain region-specific analysis. This research work captures critical frequency domain characteristics by extracting pivotal spectral features from EEG data, thereby improving the signal representation for LSTM networks. Additionally, this proposed work has employed the Synthetic Minority Over-sampling Technique (SMOTE) to handle the class imbalance problem. Furthermore, a comprehensive spatial analysis of EEG signals is performed to evaluate performance variations across distinct brain regions, enabling targeted region-wise analysis. This strategy effectively reduces the number of channels required, minimizing the need to process all 22 channels specified in the CHB-MIT dataset, thus significantly decreasing computational complexity while preserving high seizure detection performance. This work has obtained a mean value of accuracy of 95.43%, precision of 95.46%, sensitivity of 95.59%, F1-score of 95.48%, and specificity of 95.25% for the brain region providing the best performance for seizure discrimination. The results demonstrate that integrating spectral features and LSTM, augmented by spatial insights, enhances seizure detection performance and hence assists in identifying epileptogenic regions. This tool enhances clinical applications by improving diagnostic precision, personalized treatment strategies, and supporting precise surgical planning for epilepsy, ensuring safer resection and better outcomes.

This study investigates the adaptation of DeepInsight for cancer subtype classification using high-dimensional gene expression data. Originally designed for non-image data, DeepInsight has been adapted for cancer classification.

We evaluated DeepInsight's performance against several models, including support vector machines, LightGBM, neural networks, and decision trees, with and without the application of the Synthetic Minority Oversampling Technique. The study utilized gene expression data from breast, lung, and colon cancers. A novel multi-class feature selection method was introduced, using modified aggregated class activation maps to identify key genes across different cancer subtypes. These critical genes were further analyzed through Gene Ontology to explore their roles in significant biological processes.

DeepInsight consistently outperformed traditional models in terms of F1 score across breast, lung, and colon cancer datasets, effectively addressing multi-class classification challenges. Notably, several top genes were identified as significant across multiple methods. Furthermore, we conducted a Gene Ontology analysis on the critical genes, including the top genes identified by DeepInsight and the common genes recognized through multiple methods.

The adaptation of DeepInsight provides an approach to cancer subtype classification by transforming high-dimensional gene expression data into image representations. Utilizing aggregated class activation maps, it effectively identifies critical pixels within these images, enabling the discovery of distinct genes that may not be highlighted by other methods. DeepInsight demonstrates potential as a valuable tool for classifying cancer subtypes and critical genes.

Work-related injuries from overexertion, particularly lifting, are a major concern in occupational safety. Traditional assessment tools, such as the Revised NIOSH Lifting Equation (RNLE), require significant training and practice for deployment. This study presents an approach that integrates tactile gloves with computer vision (CV) to enhance the assessment of lifting-related injury risks, addressing the limitations of existing single-modality methods. Thirty-one participants performed 2747 lifting tasks across three lifting risk categories (LI < 1, 1 ≤ LI ≤ 2, LI > 2). Features including hand pressure measured by tactile gloves during each lift and 3D body poses estimated using CV algorithms from video recordings were combined and used to develop prediction models. The Convolutional Neural Network (CNN) model achieved an overall accuracy of 89 % in predicting the three lifting risk categories. The results highlight the potential for a real-time, non-intrusive risk assessment tool to assist ergonomic practitioners in mitigating musculoskeletal injury risks in workplace environments.

This study presents a simple approach for detecting honey adulteration by integrating calorimetric data from differential scanning calorimetry (DSC) with machine learning classification (MLC) techniques, specifically using convolutional neural network (CNN) model alongside the Synthetic Minority Over-sampling TEchnique (SMOTE) for data augmentation. The thermal profiles of different honey varieties, sugar adulterants, and adulterated samples were acquired using DSC. Shifts in glass transition temperatures were observed in adulterated honey. The DSC data were analyzed using principal component analysis and MLC workflow. CNN model applied to original dataset reported accuracy of 24-67 %. However, integrating CNN model with SMOTE algorithm resulted in a significant accuracy improvement to 60-91 %. The integration of DSC with MLC provides a rapid and accurate method for detecting honey adulteration, demonstrating strong generalization capability. The proposed approach could facilitate the development of a framework to detect fraudulent practices, safeguarding honey industry and consumers from sugar-based adulterations.

Lottery gambling is a relatively benign form of gambling. Nonetheless, individuals with gambling problems may engage in lottery play and/or play the lottery exclusively. Lottery loyalty programs have data that could be used to screen for problem gambling, as they collect information on demographics and ticket purchases from players who sign up to receive incentives. The current study evaluates the feasibility of machine learning to identify individuals who have gambling problems using data collected from a state lottery loyalty program.

Data from ticket uploads was merged with an online survey sent to loyalty program participants (N = 5903). The Problem Gambling Severity Index (PGSI) was used to screen for problem gambling, with a five or greater denoting problem gambling (n = 809; 14%). Other survey items queried frequency of other gambling (e.g., casino slot machine) as well as amounts spent. Random forests analysis, a predictive modeling technique, was used to predict individuals who have gambling problems.

Problem gambling was more common among loyalty program players than typical in population samples. The random forest algorithm performed fairly well overall, but sensitivity was poor, indicating that the model did not identify individuals with problem gambling effectively. Lottery loyalty programs may be a promising setting for screening and secondary prevention efforts because of relatively high prevalence of problem gambling, but random forests may not be the best approach for detecting those at risk.

This study explores the application of machine learning (ML) in identifying transcriptomic changes associated with pulmonary pathologies induced by titanium dioxide nanoparticles (TiO2-NPs). Such an approach significantly contributes to understanding the underlying mode-of-action of TiO2-NP inhalation and follows the European Chemicals Agency's recommendations on applying Novel Approach Methodologies designed for reducing animal studies. The lung gene expression profiles from mice exposed via single intratracheal instillations to TiO2-NPs with varying physicochemical properties on day 1, and day 28 post-exposure were analyzed to develop computational models for predicting the lung pathologies of rutile TiO2-NPs. More than 600 random forest models were generated and rigorously validated, leading to the identification of 17 high-quality models with an average accuracy of 0.95. These models link nanoparticle-deposited surface area, charge, and post-exposure sampling time with dysregulation in key genes, including serum amyloid Saa1 (59.7-fold increase), Saa3 (253.7-fold increase), and the cytokine Ccl2 (3.4-fold increase). These genes are strongly associated with lung inflammation and fibrosis, key pathological responses to nanomaterial exposure. The study highlights critical nanoparticle features that drive transcriptomic changes. Hierarchical clustering confirmed the mechanistic links between nanoparticle properties and transcriptomic changes. This study demonstrates ML's potential to integrate omics data for nanosafety, offering a robust framework for early detection of adverse effects. The models enable the prediction of gene expression changes based on nanoparticle features, aiding in potential Safe and Sustainable-by-design of nanomaterials.

Chronic kidney disease (CKD) poses a significant risk for diabetes patients, often leading to severe complications. Early and accurate CKD stage detection is crucial for timely intervention. However, it remains challenging due to its asymptomatic progression, the oversight of routine CKD tests during diabetes checkups, and limited access to nephrologists. This study aimed to address these challenges by developing a multiclass CKD stage prediction model for diabetes patients using longitudinal data from the Chronic Renal Insufficiency Cohort (CRIC) study. A novel iterative backward feature selection strategy was employed to determine key predictors of the CKD stage. TabNet, an attention-based deep learning architecture, was used to build classification models in complete and simplified categories. The complete model used 31 features, including complex kidney biomarkers, while the simplified model used 15 features readily available from routine checkups. The performance of TabNet was compared against traditional tree-based ensemble methods (XGBoost, random forest, AdaBoost) and a multi-layer perceptron. Model-specific and model-agnostic explainable AI (XAI) techniques were applied to interpret model decisions, enhancing the transparency and clinical applicability of the proposed approach. The TabNet models demonstrated superior performance, achieving 94.06 % and 92.71 % accuracy in cross-validation for the complete and simplified models, respectively, and 91.00 % and 88.00 % accuracy on test sets. XAI analysis identified serum creatinine, cystatin C, sex, and age as the most influential factors in CKD stage classification. The proposed TabNet models offer a robust approach for early CKD severity detection in diabetes patients, potentially improving clinical decision-making and patient outcomes.

Determine the efficacy of commonly used approaches to handling missing and/or imbalanced Electronic Health Record (EHR) data on the performance of predictive models targeting risk of admission, intensive care unit (ICU) use, or prolonged length of stay (PLOS) among presenting febrile pediatric emergency department (ED) patients.

Historical ED EHR data was used to train a series of XGBoost (XGB) and logistic regression (LR) classifiers. Data handling strategies included imputation methods (multiple imputation (MI), median imputation, complete case (CC) analysis), and imbalanced data corrections (minority oversampling, stratified sub-group analysis). Model performance was evaluated using discriminative (AUC, AUPRC) and calibration metrics (Brier score, Z-scores, p-values).

Among the study population, 34 % were admitted, 2 % utilized the ICU, and 7 % had a PLOS. Significant data missingness was observed and determined to be not at random (MNAR). In predicting admissions using data recorded within the first two hours of presentation, LR trained using full cohort with median imputation was comparable to MI yielding well-calibrated admissions models with an AUC/AUPRC of 0.82/0.73 while CC analysis yielded an AUC/AUPRC of 0.76/0.78. XGB, trained with unimputed data, produced a well-calibrated admissions classifier with an AUC/AUPRC of 0.85/0.78. In contrast, imbalanced data correction techniques, including synthetic minority oversampling (SMOTE), risk stratification, or the use of XGB did not significantly improve the poor AUPRC and calibration performance of LR models predicting ICU and PLOS.

Both XGB and LR with median imputation demonstrated robust performance in predicting admissions in the presence of missing data. However, deriving clinically useful models for rare outcomes, such as ICU use or PLOS, remains a challenge due to poor precision/recall and calibration performance. Further research is needed to improve the prediction of rare outcomes in this population.

The openness of Internet stimulates a large number of fraud behaviors which have become a huge threat. Graph-based fraud detectors have attracted extensive interest since the abundant structure information of graph data has proved effective. Conventional Graph Neural Network (GNN) approaches reveal fraudsters based on the homophily assumption. But fraudsters typically generate heterophilous connections and label-imbalanced neighborhood. Such behaviors deteriorate the performance of GNNs in fraud detection tasks due to the low homophily in graphs. Though some recent works have noticed the challenges, they either treat the heterophilous connections as homophilous ones or tend to reduce heterophily, which roughly ignore the benefits from heterophily. In this work, an integrated two-strategy framework HeteGAD is proposed to balance both homophily and heterophily information from neighbors. The key lies in explicitly shrinking intra-class distance and increasing inter-class segregation. Specifically, the Heterophily-aware Aggregation Strategy tease out the feature disparity on heterophilous neighbors and augment the disparity between representations with different labels. And the Homophily-aware Aggregation Strategy are devised to capture the homophilous information in global text and augment the representation similarity with the same label. Finally, two corresponding inter-relational attention mechanisms are incorporated to refine the procedure of modeling the interaction of multiple relations. Experiments are conducted to evaluate the proposed method with two real-world datasets, and demonstrate that the HeteGAD outperforms 11 state-of-the-art baselines for fraud detection.

Extrachromosomal circular DNA (eccDNA) has emerged as a potential biomarker for disease due to its stable closed circular structure. However, the diagnostic utility of eccDNA remains underexplored. In this study, we demonstrate that the characteristics of eccDNA associated with genomic repetitive elements change in breast cancer patient tissues and plasma. These changes can serve as signatures for accurate cancer classification. We profiled eccDNA annotated to repeat elements across the genome in tissues and plasma, aggregating each repeat element to the superfamily and subfamily level. Our findings indicate that eccDNA associated with repetitive elements in cancer exhibits regular patterns of enrichment or depletion in specific elements, particularly at the family level. Additionally, these repeat element changes are present in different subtypes of breast cancer, correlated with varying hormone receptor expression. Although there are differences in the landscapes of eccDNA on repetitive elements between cancer tissues and paired plasma, the unique characteristics of eccDNA associated with repetitive sequences in the plasma of cancer patients facilitate better differentiation from normal individuals. These analyses reveal that changes in eccDNA associated with repeat sequences in human cancers can be used as diagnostic biomarkers for cancer patients.

Distinguishing between productive (wet) and non-productive (dry) cough types is important for evaluating respiratory health, assisting in differential diagnosis, and monitoring disease progression. However, assessing cough type through the perception of cough sounds in clinical settings poses challenges due to its subjectivity. Employing objective cough sound analysis holds promise for aiding diagnostic assessments and guiding the management of respiratory conditions. This systematic review aims to assess and summarize the predictive capabilities of machine learning algorithms in analyzing cough sounds to determine cough type.

A systematic search of the Scopus, Medline, and Embase databases conducted on March 8, 2025, yielded three studies that met the inclusion criteria. The quality assessment of these studies was conducted using the checklist for the assessment of medical artificial intelligence (ChAMAI).

The inter-rater agreement for annotating wet and dry coughs ranged from 0.22 to 0.81 across the three studies. Furthermore, these studies employed diverse inputs for their machine learning algorithms, including different cough sound features and time-frequency representations. The algorithms used ranged from conventional classifiers like logistic regression to neural networks. While the classification accuracy for identifying wet and dry coughs ranged from 78% to 87% across these studies, none of them assessed their algorithms through external validation.

The high variability in inter-rater agreement highlights the subjectivity in manually interpreting cough sounds and underscores the need for objective cough sound analysis methods. The predictive ability of cough-type classification algorithms shows promise in the small number of studies analyzed in this systematic review. However, more studies are needed, particularly those validating their models on independent and external datasets.

Machine learning techniques hold significant potential to support the diagnosis and prognosis of diseases. However, the success of these approaches is heavily dependent on rigorous data acquisition, preprocessing and data organization.

This article reviews the literature to evaluate key factors in dataset construction, focusing on data structure, preprocessing, and data organization, particularly in the context of imaging data.

The main issues with data construction when dealing with medical applications are noise (incorrect or irrelevant data), sparsity/ limited availability, representativeness/variability, and data imbalance (uneven class distribution).While preprocessing steps prepare the data to be suitable for the models, data organization focuses in improving data arranging to increase the model performance. Additionally, the impact of CNN complexity in processing balanced, imbalanced, and complex datasets shows that complex CNNs are not always the optimal choice for every classification problem.

By integrating knowledge from Health Sciences and Biomedical Engineering, we aim to enhance healthcare professionals' understanding of machine learning for image analysis in Oral Medicine and Pathology. This encourages their involvement in patient recruitment and data acquisition, broadening their roles and significantly contributing to the creation of well-characterized datasets for future research and applications.

Buprenorphine for opioid use disorder (B-MOUD) is essential to improving patient outcomes; however, retention is essential.

To develop and validate machine-learning algorithms predicting retention, overdoses, and all-cause mortality among US military veterans initiating B-MOUD.

Veterans initiating B-MOUD from fiscal years 2006-2020 were identified. Veterans' B-MOUD episodes were randomly divided into training (80%;n = 45,238) and testing samples (20%;n = 11,309). Candidate algorithms [multiple logistic regression, least absolute shrinkage and selection operator regression, random forest (RF), gradient boosting machine (GBM), and deep neural network (DNN)] were used to build and validate classification models to predict six binary outcomes: 1) B-MOUD retention, 2) any overdose, 3) opioid-related overdose, 4) overdose death, 5) opioid overdose death, and 6) all-cause mortality. Model performance was assessed using standard classification statistics [e.g., area under the receiver operating characteristic curve (AUC-ROC)].

Episodes in the training sample were 93.0% male, 78.0% White, 72.3% unemployed, and 48.3% had a concurrent drug use disorder. The GBM model slightly outperformed others in predicting B-MOUD retention (AUC-ROC = 0.72). RF models outperformed others in predicting any overdose (AUC-ROC = 0.77) and opioid overdose (AUC-ROC = 0.77). RF and GBM outperformed other models for overdose death (AUC-ROC = 0.74 for both), and RF and DNN outperformed other models for opioid overdose death (RF AUC-ROC = 0.79; DNN AUC-ROC = 0.78). RF and GBM also outperformed other models for all-cause mortality (AUC-ROC = 0.76 for both). No single predictor accounted for >3% of the model's variance.

Machine-learning algorithms can accurately predict OUD-related outcomes with moderate predictive performance; however, prediction of these outcomes is driven by many characteristics.

To examine the discrimination, calibration, and algorithmic fairness of the Epic End of Life Care Index (EOL-CI).

We assessed the EOL-CI's performance by estimating area under the receiver operating characteristic curve (AUC), sensitivity, and positive and negative predictive values in community-dwelling adults ≥65 years of age in a single health system in the Southeastern United States. Algorithmic fairness was examined by comparing the model's performance across sex, race, and ethnicity subgroups. Using a machine learning approach, we also explored local re-calibration of the EOL-CI considering additional information on past hospitalizations and frailty.

Among 215 731 patients (median age = 74 years, 57% female, 12% of Black race), 10% were classified as medium risk (15-44) and 3% as high risk (≥45) by the EOL-CI. The observed 1-year mortality rate was 3%. The EOL-CI had an AUC 0.82 for 1-year mortality, with a positive predictive value of 22%. Predictive performance was generally similar across sex and race subgroups, though the EOL-CI displayed better performance with increasing age and in older adults with 2 or more outpatient encounters in the past 24 months. Local re-calibration of the EOL-CI was required to provide absolute estimates of mortality risk, and calibration was further improved when the EOL-CI was augmented with data on inpatient hospitalizations and frailty.

The EOL-CI demonstrates reasonable discrimination, albeit with better performance in older adults and in those with greater health system contact.

Local refinement and calibration of the EOL-CI score is required to provide direct estimates of prognosis, with the goal of making the EOL-CI a more a valuable tool at the point of care for identifying patients who would benefit from targeted palliative care interventions and proactive care planning.

The prediction of Intensive Care Unit (ICU) readmission has become a crucial area of research due to the increasing demand for ICU resources and the need to provide timely interventions to critically ill patients. In recent years, several studies have explored the use of statistical, machine learning (ML), and deep learning (DL) models to predict ICU readmission. This review paper presents an extensive overview of these studies and discusses the challenges associated with ICU readmission prediction. We categorize the studies based on the type of model used and evaluate their strengths and limitations. We also discuss the performance metrics used to evaluate the models and their potential clinical applications. In addition, this review explores current methodologies, data usage, and recent advances in interpretability and explainable AI for medical applications, offering insights to guide future research and development in this field. Finally, we identify gaps in the current literature and provide recommendations for future research. Recent advances like ML and DL have moderately improved the prediction of the risk of ICU readmission. However, more progress is needed to reach the precision required to build computerized decision support tools.

The AI model, enhanced by SMOTE to balance data classes, accurately predicted visual field deterioration in patients with myopic normal tension glaucoma. Using SHAP analysis, the key variables driving disease progression were identified.

To develop and validate a Synthetic Minority Over-sampling Technique (SMOTE)-enhanced artificial intelligence (AI) model for predicting visual field progression in myopic normal tension glaucoma (NTG) patients.

This retrospective cohort study included 100 eyes from myopic NTG patients with a mean follow-up of 10.3±3.2 years. Baseline parameters included intraocular pressure (IOP), central corneal thickness, axial length, and visual field metrics. A SMOTE-enhanced AI model was created to address class imbalance in progression events. Model performance was evaluated using receiver operating characteristic (ROC) analysis, cross-validation, and calibration plots. Predictive factor importance was evaluated through SHapley Additive exPlanations (SHAP) analysis.

Visual field progression was observed in 28% of patients, with a median progression time of 3.2 years. The AI model achieved an area under the ROC curve (AUC) of 0.83 (95% CI, 0.75-0.91), with promising sensitivity (0.81) and specificity (0.77). SHAP analysis identified baseline mean deviation (MD), age, axial length, baseline IOP, and visual field index (VFI) as key predictors. When patients were stratified based on model-predicted risk scores, those with scores above 0.8 had significantly higher observed progression rates (82.6%) compared with those with lower risk scores. Subgroup analysis revealed strong correlations between progression risks and older age, greater axial length, and worse baseline MD.

The SMOTE-enhanced AI model shows reasonable predictive performance and potential clinical utility for identifying visual field progression in myopic NTG patients, though further validation in larger cohorts is needed. By addressing class imbalance and myopia-specific challenges, this approach enables personalized risk stratification and early intervention.

Smartphone applications (apps) show promise as an effective and scalable intervention modality for disordered eating, yet responsiveness varies considerably. The ability to predict user responses to app-based interventions is currently limited. Machine learning (ML) techniques have shown potential to improve prediction of complex clinical outcomes. We applied ML techniques to predict responsiveness to a dialectical behaviour therapy-based smartphone app for recurrent binge eating.

Data were collected as part of a randomised controlled trial (RCT). The present sample was based on data from 576 participants with recurrent binge eating. 10 common classification and regression approaches were used to predict outcomes that represent key stages of the user experience, including initial intervention uptake, app adherence, study drop-out, and symptom change. Models were developed using 69 self-reported baseline variables (i.e., demographic, clinical, psychological) and several app usage variables (i.e., number of modules completed) as predictors.

All models, using only baseline predictors, performed sub-optimally at predicting engagement (AUCs = 0.48-0.61; R2 = 0.00-0.04) and symptom level change (R2 = 0.00-0.07). Incorporating usage data improved prediction of study dropout (AUC = 0.69-0.76).

ML models were unable to accurately predict responsiveness using self-reported baseline predictors alone. Predicting outcomes with greater precision may require consideration of how predictors change over time and interact with a user's context. Modelling usage pattern data appears to improve prediction of dropout, highlighting the potential value of tracking intervention usage to identify individuals at risk of disengagement.

The formation of autism advocacy organisations led by family members of autistic individuals led to intense criticism from some parts of the autistic community. In response to what was perceived as a misrepresentation of their interests, autistic individuals formed autistic self-advocacy groups, adopting the philosophy that autism advocacy should be led 'by' autistic people 'for' autistic people. However, recent claims that self-advocacy organisations represent only a narrow subset of the autistic community have prompted renewed debate surrounding the role of organisations in autism advocacy. While many individuals and groups have outlined their views, the debate has yet to be studied through computational means. In this study, we apply machine learning and natural language processing techniques to a large-scale collection of Tweets from organisations and individuals in autism advocacy. We conduct a specification curve analysis on the similarity of language across organisations and individuals, and find evidence to support claims of partial representation relevant to both self-advocacy groups and organisations led by non-autistic people. In introducing a novel approach to studying the long-standing conflict between different groups in the autism advocacy community, we hope to provide both organisations and individuals with new tools to help ground discussions of representation in empirical insight.Lay AbstractSome autism advocacy organisations are run by family members of autistic people, and claim to speak on behalf of autistic people. These organisations have been criticised by autistic people, who feel like autism charities do not adequately represent their true interests. In response to these organisations, autistic people have come together to form autistic self-advocacy organisations, or groups in which activists can spread awareness of autism from an autistic point-of-view. However, some people say that autistic self-advocacy organisations do not sufficiently represent the needs of all autistic people. These tensions between organisations and individuals have made it difficult to determine which organisations can make the claim that they represent all autism advocates individuals equally, instead of showing preference to a sub-group within the autism community. In this study, we try to approach this issue using computational tools to see if, in their Twitter posts, both kinds of organisations show a preference for the interests of autistic people or parents of autistic children. We do so by comparing a large body of Tweets by organisations to Tweets by autistic people and parents of autistic children. We find that both kinds of organisations match the interests of one group of autism advocates better than the other. The insight we provide has the potential to inspire new conversations and solutions to a long-standing conflict in autism advocacy.

Crashes involving farm equipment vehicles are a significant safety concern on public roads, particularly in rural and agricultural regions. These vehicles display unique challenges due to their slow-moving operational speed and interactions with faster vehicles, often leading to severe crashes. This study analyzed crashes involving farm equipment vehicles to examine the factors influencing crash severity, with a particular focus on comparing incidents on county roads to those on non-county roads. The dataset included key variables such as road geometry, lighting conditions, and traffic interactions, with preprocessing techniques like Synthetic Minority Over-sampling Technique (SMOTE) applied to address class imbalance. The TabNet model, a tabular deep learning model, was employed to analyze crash dynamics, offering both predictive accuracy and interpretability through feature importance and SHapley Additive exPlanations (SHAP) plots. Findings revealed that crash severity on county roads is primarily influenced by crash speed limit, first harmful event, traffic control, and person age, reflecting the role of road geometry and demographic risk in rural settings. In contrast, non-county roads were more affected by lighting conditions, intersection-related features, and population group, emphasizing the impact of visibility and traffic complexity in urban areas. Speed limit consistently emerged as a critical factor across all road types and severity levels. The study emphasized the need for targeted safety interventions, including visibility enhancements, speed management, and enhanced education campaigns for county and non-county areas.

Seed viability, a key indicator for quality assessment, directly impacts the emergence of field seedlings. The existing nondestructive testing model for maize seed vitality based on naturally aged seeds and predominantly relying on single-modal data like MV and RS, achieves an accuracy of less than 70 %. To elucidate the influence of different data on model accuracy, this study proposes the MSCNSVN model for detecting seed viability by collecting multisensor information from maize seeds using sensors, such as MV, RS, TS, FS, and SS. Our findings indicated that (1) the single-modal FS dataset achieved optimal prediction accuracy, with FS570/600 contributing the most; (2) multimodal data fusion outperformed single-modal data, with an accuracy improvement of 10 %, while the MV + RS + FS dataset achieved the highest accuracy; (3) the MSCNSVN model demonstrated superior performance compared to baseline models; (4) modeling with dual-variety datasets and endosperm surface datasets improved accuracy by 2 %-3 %.

Objective.Accurate localization of the epileptogenic zone (EZ) is crucial for epilepsy surgery, but the class imbalance of epileptogenic vs. non-epileptogenic electrode contacts in intracranial electroencephalography (iEEG) data poses significant challenges for automatic localization methods. This review evaluates methodologies for handling the class imbalance in EZ localization studies that use machine learning (ML).Approach.We systematically reviewed studies employing ML to localize the EZ from iEEG data, focusing on strategies for addressing class imbalance in data handling, algorithm design, and evaluation.Results.Out of 2,128 screened studies, 35 fulfilled the inclusion criteria. Across the studies, the iEEG contacts annotated as epileptogenic prior to automatic localization constituted a median of 18.34% of all contacts. However, many of these studies did not adequately address the class imbalance problem. Techniques such as data resampling and cost-sensitive learning were used to mitigate the class imbalance problem, but the chosen evaluation metrics often failed to account for it.Significance.Class imbalance significantly impacts the reliability of EZ localization models. More comprehensive management and innovative approaches are needed to enhance the robustness and clinical utility of these models. Addressing class imbalance in ML models for EZ localization will improve both the predictive performance and reliability of these models.

Extended power transmission outages caused by weather events can significantly impact the economy, infrastructure, and residents' quality of life in affected regions. One of the challenges is providing early, accurate warnings for these disruptions. To address this challenge, we introduce HMN-RTS, a hierarchical multiplex network designed to predict the duration of a forced transmission outage by leveraging a multi-modal approach. We investigate outage duration prediction over two years at the county level, focusing on the states of the Pacific Northwest region, including Idaho, California, Montana, Washington, and Oregon. The multiplex network layers collect diverse data sources, including information about power outages, weather data, weather forecasts, lightning, land cover, transmission lines, and social media. Our findings demonstrate that this approach enhances the accuracy of predicting power outage duration. The HMN-RTS model improves 3 hours ahead outage predictions, achieving a macro F1 score of 0.79 compared to the best alternative of 0.73 for a five-class classification. The HMN-RTS model provides valuable predictions of outage duration across multiple time horizons and seasons, enabling grid operators to implement timely outage mitigation strategies. Overall, the results underscore the HMN-RTS model's capability to deliver early and practical risk assessments.

Despite emerging evidence on the health impacts of fine particulate matter (PM2.5) from wildland fire smoke, the specific effects of PM2.5 composition on health outcomes remain uncertain. We developed a three-level, chemical transport model-based framework to estimate daily full-coverage concentrations of smoke-derived carbonaceous PM2.5, specifically organic carbon (OC) and elemental carbon (EC), at a 1 × 1 km2 spatial resolution from 2002 to 2019 across the contiguous U.S. (CONUS) and Southern Canada (SC). A 10-fold random cross-validation confirmed robust performance, with daily R2 = 0.77 (OC) and 0.80 (EC) in the smoke-off scenario and 0.67 (OC) and 0.71 (EC) in the smoke-on scenario, and exceeded 0.90 at the monthly scale after residual adjustment. Modeling results indicated that increases in wildland fire smoke have offset approximately one-third of the improvements in background air quality. In recent years, wildland fire smoke has become more frequent and carbonaceous PM2.5 concentrations have intensified, especially in the Western CONUS and Southwestern Canada. Wildfire season is also starting earlier and lengthens throughout the year, leading to more population being exposed. We estimated that long-term exposure to fire smoke carbonaceous PM2.5 is responsible for approximately 7455 and 259 non-accidental deaths annually in the CONUS and SC, respectively, with associated annual monetized damage of 68.3 billion USD for the CONUS and 1.9 billion CAD for SC. The Southeastern CONUS, where prescribed fires are prevalent, contributed most to these health impacts and monetized damages. Our findings offer critical insights to inform policy development and assess future health burdens associated with fire smoke exposure.

Despite extensive research on keratoconus (KC) detection with traditional machine learning models, stacking ensemble learning approaches remain underexplored. This paper presents a stacking ensemble learning method to enhance automated KC screening.

This study utilizes a clinical dataset containing detailed corneal data from 2491 cases classified as non-KC (NKC), subclinical KC (SCKC) and clinical KC (CKC). Each cornea is represented by 79 features extracted from Pentacam imaging. Following extensive pre-processing, key corneal features that are strongly correlated with the target diagnosis are identified. These features are the keratometry of the steepest anterior point, surface variance index, vertical asymmetry index, height decentration index, and height asymmetry index. A novel stacking ensemble model is developed using the selected features to improve corneal classification into NKC, SCKC, and CKC by integrating top tree-based classifiers (random forest, gradient boosting, decision trees) with a support vector machine meta-classifier.

The pre-processing and feature selection techniques reduced the model's parameters to just 6.33% of the original dataset, improving classification performance, and cutting over 85% of the training time. The performance of the developed model was validated and tested on unseen data. Experimental results showed that the model outperforms existing studies, achieving 99.72% accuracy, precision, sensitivity, F1, and F2 scores, with a Matthews correlation coefficient of 0.995. It accurately classified all NKC and CKC cases, with just one misclassification involving an SCKC case. The model also demonstrated consistent performance on 100 additional unseen test cases, underscoring its generalizability and robustness in KC screening.

By combining the strengths of diverse base models and key Pentacam indices, the stacking ensemble approach ensures reliable, accurate KC screening, providing clinicians with an automated tool for early detection and better patient management.

Cushing syndrome (CS) is a rare endocrine disorder characterized by excessive secretion of glucocorticoids, leading to a variety of clinical manifestations, comorbidities, and increased mortality despite treatment. Despite advances in imaging modalities and biochemical testing, the diagnosis and management of CS remains challenging. Several tests are used to confirm the diagnosis of CS, including urinary free cortisol measurements, dexamethasone suppression tests (1 mg, 2 mg, and 8 mg), and nocturnal salivary cortisol measurements. However, each of these tests has some limitations, making the diagnosis of CS.

In this paper, we explore the potential of state-of-the-art machine learning algorithms as a clinical decision support system for analyzing and classifying CS. Our aim is to use advanced machine learning methods to analyze the accuracy rates of diagnostic tests and identify the most sensitive tests for diagnosing CS.

In this study, we performed binary classification based on data from 278 patients with CS (CS+) and 220 healthy patients (CS-). We developed a linear mathematical model with high predictive ability, achieving a classification accuracy of 97.03% and a Kappa value of 94.05%. The correlation graph shows that CS has strong positive relationships with 2 mg (78.8%), 1 mg (76.9%), and mc (72.1%), and moderate positive correlations with 8 mg (45%) and saliva (45.4%). In contrast, gender has almost no correlation with CS, so it was removed from the dataset. As a result, the model achieves an overall classification accuracy of 97.03%. Finally, we converted the linear model into a mobile application for use by specialist doctors in the field of endocrinology.

Traditional diagnostic methods can be time-consuming and require specialized medical expertise. Recently, advances in machine learning and mobile technology have opened new avenues for improving diagnostic accuracy and accessibility. This study explores the integration of machine learning algorithms into a mobile application designed to assist healthcare professionals and patients in the diagnosis of CS.

The time after hospital discharge carries high rates of mortality in neonates and young children in sub-Saharan Africa. Previous work using logistic regression to develop risk assessment tools to identify those at risk for postdischarge mortality has yielded fair discriminatory value. Our objective was to determine if machine learning models would have greater discriminatory value to identify neonates and young children at risk for postdischarge mortality.

We conducted a planned secondary analysis of a prospective observational cohort at Muhimbili National Hospital in Dar es Salaam, Tanzania and John F. Kennedy Medical Center in Monrovia, Liberia. We enrolled neonates and young children near the time of discharge. The outcome was 60-day postdischarge mortality. We collected socioeconomic, demographic, clinical, and anthropometric data during hospital admission and used machine learning (ie, eXtreme Gradient Boosting (XGBoost), Hist-Gradient Boost, Support Vector Machine, Neural Network, and Random Forest) to develop risk assessment tools to identify: (1) neonates and (2) young children at risk for postdischarge mortality.

A total of 2310 neonates and 1933 young children enrolled. Of these, 71 (3.1%) neonates and 67 (3.5%) young children died after hospital discharge. XGBoost, Hist Gradient Boost, and Neural Network models yielded the greatest discriminatory value (area under the receiver operating characteristic curves range: 0.94-0.99) and fewest features, which included six features for neonates and five for young children. Discharge against medical advice, low birth weight, and supplemental oxygen requirement during hospitalisation were predictive of postdischarge mortality in neonates. For young children, discharge against medical advice, pallor, and chronic medical problems were predictive of postdischarge mortality.

Our parsimonious machine learning-based models had excellent discriminatory value to predict postdischarge mortality among neonates and young children. External validation of these tools is warranted to assist in the design of interventions to reduce postdischarge mortality in these vulnerable populations.

Rapid weight gain (RWG) during infancy, defined as an upward crossing of one centile line on a weight growth chart, is highly predictive of subsequent obesity risk. Identification of infant RWG could facilitate obesity risk assessment from infancy.

Leveraging machine learning (ML) algorithms, this study aimed to develop and validate risk prediction models to identify infant RWG by the age of 1 year.

Data from 7 Australian and New Zealand cohorts were pooled for risk model development and validation (n=5233). A total of 8 ML algorithms predicted infant RWG using routinely available prenatal and early postnatal factors, including maternal prepregnancy weight status, maternal smoking during pregnancy, gestational age, parity, infant sex, birth weight, any breastfeeding and timing of solids introduction at the age of 6 months. Pooled data were randomly split into a training dataset (70%) and a test dataset (30%) for model training and validation, respectively. Model consistency was evaluated using 5-fold cross-validation. Model predictive performance was evaluated by area under the receiver operating characteristic (ROC) curve (AUC), accuracy, precision, sensitivity, specificity, and Cohen κ.

The average prevalence of infant RWG was 27%. In the training dataset, all ML algorithms showed acceptable to excellent discrimination with AUCs ranging from 0.75 to 0.86. Accuracy, which indicates the overall correctness of the model, ranged from 0.69 to 0.78. Precision, which measures the model's ability to avoid false positives, ranged from 0.68 to 0.77. The spread of sensitivity, specificity, and Cohen κ of all models was 0.68-0.80, 0.65-0.78, and 0.38-0.56, respectively. Of the 8 algorithms, the Gradient Boosting model showed the most favorable predictive accuracy. Validation of the Gradient Boosting model in the testing dataset exhibited excellent discrimination (AUC 0.3-0.6) and good ability to make accurate predictions, particularly true positive cases (with accuracy and sensitivity>0.75), but modest performance for precision (0.57-0.60) and Cohen κ (0.47-0.52).

This study developed the first set of ML-based risk prediction models to identify infants' risk of experiencing RWG by the age of 1 year with acceptable accuracy. The models could be feasibly integrated into routine child growth monitoring and may facilitate population-wide early obesity risk assessment in primary health care.

Diabetes mellitus, a global health concern with severe complications, demands early detection and precise staging for effective management. Machine learning approaches, combined with bioinformatics, offer promising avenues for enhancing diagnostic accuracy and identifying key biomarkers.

This study employed a multi-class classification framework to classify patients across four health states: healthy, prediabetes, type 2 Diabetes Mellitus (T2DM) without complications, and T2DM with complications. Three models were developed using molecular markers, biochemical markers, and a combined model of both. Five machine learning classifiers were applied: Random Forest (RF), Extra Tree Classifier, Quadratic Discriminant Analysis, Naïve Bayes, and Light Gradient Boosting Machine. To improve the robustness and precision of the classification, Recursive Feature Elimination with Cross-Validation (RFECV) and a fivefold cross-validation were used. The multi-class classification approach enabled effective discrimination between the four diabetes stages.

The top contributing features identified for the combined model through RFECV included three molecular markers-miR342, NFKB1, and miR636-and two biochemical markers the albumin-to-creatinine ratio and HDLc, indicating their strong association with diabetes progression. The Extra Trees Classifier achieved the highest performance across all models, with an AUC value of 0.9985 (95% CI: [0.994-1.000]). This classifier outperformed other models, demonstrating its robustness and applicability for precise diabetes staging.

These findings underscore the value of integrating machine learning with molecular and biochemical markers for the accurate classification of diabetes stages, supporting a potential shift toward more personalized diabetes management.

DNA-encoded chemical libraries (DECLs) have become integral to early-stage drug discovery, yielding active compounds and extensive labeled datasets for machine learning (ML)-based prediction of bioactive molecules. However, the information content of DECL selection data remains scarcely explored. This study systematically investigates for the first time the prevalence of false negatives and the influence of the linker in DECL data. Using a focused DECL targeting the poly-(ADP-ribose) polymerases PARP1/2 and TNKS1/2 as a model system, we found that our DECL selections frequently miss active compounds, with numerous false negatives for each identified hit. The presence of the DNA-conjugation linker emerged as a factor contributing to the underdetection of active molecules. This bias toward false negatives compromises the predictive power of DECL data for prioritizing hits, anticipating target selectivity, and training ML models, as determined by analyzing the effects of undersampling and oversampling techniques in learning the PARP2 data. Conversely, the linker's presence in DECLs offers advantages, such as enabling the identification of target-selective protein engagers, even when the underlying molecules themselves may not be selective. These findings highlight the challenges and opportunities of DECL data, emphasizing the need for best practices in data handling and ML model development in drug discovery.

Lyn is a multifunctional Src-family kinase (SFK) that regulates immune signaling and has been implicated in diverse types of cancer. Unlike other SFKs, its full-length structure and regulatory dynamics remain poorly characterized. In this study, we present the first long-timescale molecular dynamics analysis of full-length Lyn, including the SH3, SH2, and SH1 domains, across wildtype, ligand-bound, and cancer-associated mutant states. Using principal component analysis, dynamic cross-correlation matrices, and network-based methods, we show that ATP binding stabilizes the kinase core and promotes interdomain coordination, while the ATP-competitive inhibitor dasatinib and specific mutations (e.g., E290K, I364N) induce conformational decoupling and weaken long-range communication. We identify integration modules and develop an interface-weighted scoring scheme to rank dynamically central residues. This analysis reveals 44 allosteric hubs spanning SH3, SH2, SH1, and interdomain regions. Finally, a random forest classifier trained on 16 MD-derived features highlights key interdomain descriptors, distinguishing functional states with an AUC of 0.98. Our results offer a dynamic and network-level framework for understanding Lyn regulation and identify potential regulatory hotspots for structure-based drug design. More broadly, our approach demonstrates the value of integrating full-length MD simulations with network and machine learning techniques to probe allosteric control in multidomain kinases.

Vaccination with ChAdOx1 nCoV-19 is an important countermeasure to fight the COVID-19 pandemic. This vaccine enhances human immunoprotection against SARS-CoV-2 by inducing an immune response against the SARS-CoV-2 S protein. However, the immune-related genes induced by vaccination remain to be identified. This study employs feature ranking algorithms, an incremental feature selection method, and classification algorithms to analyze transcriptomic data from an experimental group vaccinated with the ChAdOx1 nCoV-19 vaccine and a control group vaccinated with the MenACWY meningococcal vaccine. According to different time points, vaccination status, and SARS-CoV-2 infection status, the transcriptomic data was divided into five groups, including a pre-vaccination group, ChAdOx1-onset group, MenACWY-onset group, ChAdOx1-7D group, and MenACWY-7D group. Each group contained samples with 13,383 RNA features and 1662 small RNA features. The results identified key genes that could indicate the efficacy of the ChAdOx1 nCoV-19 vaccine, and a classifier was developed to classify samples into the above groups. Additionally, effective classification rules were established to distinguish between different vaccination statuses. It was found that subjects vaccinated with ChAdOx1 nCoV-19 vaccine and infected with SARS-CoV-2 were characterized by up-regulation of HIST1H3G expression and down-regulation of CASP10 expression. In addition, IGHG1, FOXM1, and CASP10 genes were strongly associated with ChAdOx1 nCoV-19 vaccine efficacy. Compared with previous omics-driven studies, the machine learning algorithms used in this study were able to analyze transcriptome data faster and more comprehensively to identify potential markers associated with vaccine effect and investigate ChAdOx1 nCoV-19 vaccine-induced gene expression changes. These observations contribute to an understanding of the immune protection and inflammatory responses induced by the ChAdOx1 nCoV-19 vaccine during symptomatic episodes and provide a rationale for improving vaccine efficacy.

Identification of accelerated aging and its biomarkers can lead to more timely therapeutic interventions and decision-making. Therefore, we sought to predict aging-related slow gait, a known predictor of accelerated aging, and its determinants.

We applied a deep learning neural network (NN) and compared it to conventional logistic regression (LR) analysis. We incorporated 1,363 participants from the Baltimore Longitudinal Study of Aging to predict current and future slow gait at 6-year and 10-year follow-up using two clinically-relevant cut-points.

Our NN achieved a maximum sensitivity (specificity) of 81.2% (87.9%), for a 10-year prediction with 0.8 m/s cut-point. We demonstrated the necessity of class balancing and found the NN to perform comparably to or in some cases, better than, LR which achieved a maximum sensitivity and specificity of 84.5% and 86.3%, respectively. Sobol index analysis identified the strongest determinants to be age, BMI, sleep, and grip strength.

The novel use of a NN for this purpose, and successful benchmarking against conventional techniques, justifies further exploration and expansion of this model.

Malocclusion presents functional and aesthetic challenges, necessitating accurate diagnosis and treatment. However, variability in orthodontic treatment planning persists due to subjective assessments, limiting consistency and objectivity. Electronic dental records (EDRs) contain vast patient data that could address these challenges, but much of the rich clinical information is documented as free text, complicating analysis. This study aims to develop an Orthodontic Natural Language Processing (ONLP) model to extract structured orthodontics-related information from unstructured EDRs and identify critical features influencing malocclusion using machine learning (ML).

Data from 7693 orthodontic patients were analysed to train, test and validate the ONLP and ML models. A gold-standard dataset was created through manual review. The ONLP model utilised supervised (Named Entity Recognition-NER) and unsupervised (K-means clustering) approaches to structure information from free text. Machine learning models, including Logistic Regression, Gaussian Naive Bayes, Random Forest and XGBoost, were subsequently applied to identify feature importance for malocclusion classification.

The ONLP model achieved 89% sensitivity, 92% specificity and 91% accuracy in extracting orthodontics-related information. The supervised model demonstrated 84% accuracy, 82% F1-score and 84% recall, excelling in identifying Classes I and III malocclusions but showing reduced sensitivity for Class II. Machine learning analysis highlighted key features for malocclusion classification: maxillary crowding, overjet and arch perimeter discrepancy for Class I; maxillary spacing and anterior crossbite for Class II; and dental midline deviation and occlusal wear for Class III.

This study demonstrates a novel approach to automating orthodontic data extraction using the ONLP model, enabling advanced big data analytics and enhancing data-driven orthodontic research and care.

Background: Radiation therapy is a primary and cornerstone treatment modality for brain metastasis. However, it can result in complications like necrosis, which may lead to significant neurological deficits. This study aims to develop and validate an ensemble model with radiomics to predict radiation necrosis. Method: This study retrospectively collected and analyzed MRI images and clinical information from 209 stereotactic radiosurgery sessions involving 130 patients with brain metastasis. An ensemble model integrating gradient boosting, random forest, decision tree, and support vector machine was developed and validated using selected radiomic features and clinical factors to predict the likelihood of necrosis. The model performance was evaluated and compared with other machine learning algorithms using metrics, including the area under the curve (AUC), sensitivity, specificity, negative predictive value (NPV), and positive predictive value (PPV). SHapley Additive exPlanations (SHAP) analysis and local interpretable model-agnostic explanations (LIME) analysis were applied to explain the model's prediction. Results: The ensemble model achieved strong performance in the validation cohort, with the highest AUC. Compared to individual models and the stacking ensemble model, it consistently outperformed. The model demonstrated superior accuracy, generalizability, and reliability in predicting radiation necrosis. SHAP and LIME were used to interpret a complex predictive model for radiation necrosis. Both analyses highlighted similar significant factors, enhancing our understanding of prediction dynamics. Conclusions: The ensemble model using radiomic features exhibited high accuracy and robustness in predicting the occurrence of radiation necrosis. It could serve as a novel and valuable tool to facilitate radiotherapy for patients with brain metastasis.

Background/Objectives: Optimization algorithms are acknowledged to be critical in various fields and dynamical systems since they provide facilitation in identifying and retrieving the most possible solutions concerning complex problems besides improving efficiency, cutting down on costs, and boosting performance. Metaheuristic optimization algorithms, on the other hand, are inspired by natural phenomena, providing significant benefits related to the applicable solutions for complex optimization problems. Considering that complex optimization problems emerge across various disciplines, their successful applications are possible to be observed in tasks of classification and feature selection tasks, including diagnostic processes of certain health problems based on bio-inspiration. Sepsis continues to pose a significant threat to patient survival, particularly among individuals admitted to intensive care units from emergency departments. Traditional scoring systems, including qSOFA, SIRS, and NEWS, often fall short of delivering the precision necessary for timely and effective clinical decision-making. Methods: In this study, we introduce a novel, interpretable machine learning framework designed to predict in-hospital mortality in sepsis patients upon intensive care unit admission. Utilizing a retrospective dataset from a tertiary university hospital encompassing patient records from January 2019 to June 2024, we extracted comprehensive clinical and laboratory features. To address class imbalance and missing data, we employed the Synthetic Minority Oversampling Technique and systematic imputation methods, respectively. Our hybrid modeling approach integrates ensemble-based ML algorithms with deep learning architectures, optimized through the Red Piranha Optimization algorithm for feature selection and hyperparameter tuning. The proposed model was validated through internal cross-validation and external testing on the MIMIC-III dataset as well. Results: The proposed model demonstrates superior predictive performance over conventional scoring systems, achieving an area under the receiver operating characteristic curve of 0.96, a Brier score of 0.118, and a recall of 81. Conclusions: These results underscore the potential of AI-driven tools to enhance clinical decision-making processes in sepsis management, enabling early interventions and potentially reducing mortality rates.

Breast cancer remains a leading cause of cancer-related deaths among women worldwide, highlighting the urgent need for early detection. While mammography is the gold standard, it faces cost and accessibility barriers in resource-limited areas. Infrared thermography is a promising cost-effective, non-invasive, painless, and radiation-free alternative that detects tumors by measuring their thermal signatures through thermal infrared radiation. However, challenges persist, including limited clinical validation, lack of Food and Drug Administration (FDA) approval as a primary screening tool, physiological variations among individuals, differing interpretation standards, and a shortage of specialized radiologists. This survey uniquely focuses on integrating texture analysis and machine learning within infrared thermography for breast cancer detection, addressing the existing literature gaps, and noting that this approach achieves high-ranking results. It comprehensively reviews the entire processing pipeline, from image preprocessing and feature extraction to classification and performance assessment. The survey critically analyzes the current limitations, including over-reliance on limited datasets like DMR-IR. By exploring recent advancements, this work aims to reduce radiologists' workload, enhance diagnostic accuracy, and identify key future research directions in this evolving field.

The rising prevalence of nervous system disorders has become a significant global health challenge, with environmental pollutants identified as key contributors. However, the large number of environmental related compounds, combined with the low efficiency of traditional methods, has resulted in substantial gaps in neurotoxicity data. In this study, we developed a robust and interpretable neurotoxicity prediction model using a high-quality data set. To identify the best predictive model, three molecular representation methods (molecular fingerprints, molecular descriptors, and molecular graphs) combined with six traditional machine learning (ML) algorithms and two deep learning (DL) approaches were evaluated. The optimal model, combining molecular fingerprints and descriptors with eXtreme Gradient Boosting (XGBoost), achieved a training accuracy of 0.93 and an area under the curve (AUC) of 0.99, outperforming other ML and DL models, while maintaining interpretability. The model was used to screen 1170 compounds detected in human blood, predicting 1145 successfully. Among 89 compounds with known neurotoxicity data, the model achieved an accuracy of 0.74. It identified 821 potentially neurotoxic compounds, including 36 with high detection concentrations, warranting further study. An online platform (http://www.envwind.site/tools.html) was developed to expand accessibility. This model offers an efficient tool for predicting neurotoxicity and managing environmental health risks.

Per- and polyfluoroalkyl substances (PFAS) are persistent environmental contaminants with strong carbon‑fluorine (CF) bonds that contribute to bioaccumulation and long-term environmental and health risks. Traditional PFAS detection and treatment methods are often time-consuming, costly, and limited in scope. Recently, artificial intelligence (AI)-based technologies, particularly machine learning (ML), have emerged as powerful tools for enhancing PFAS monitoring, source identification, and remediation. ML models such as random forest (RF), gradient boosting decision trees (GBDT), support vector machines (SVM), and artificial neural networks (ANN) have been successfully applied to classify PFAS contamination sources with over 96 % accuracy, predict PFAS concentrations in groundwater with an AUC of 0.90, and optimize removal processes such as nanofiltration and adsorption with R2 values exceeding 0.93. Despite these advancement, challenges remain in ensuring high-quality datasets, addressing data imbalance and improving model interpretability. Future research should focus on expanding public datasets, leveraging Automated ML (AutoML) for optimization, and integrating Al-driven sensors for real-time detection. AI-based approaches present a transformative opportunity to enhance efficiency, accuracy, and cost-effectiveness in PFAS management, aiding regulatory decision-making and environmental protection.

The critical flicker fusion threshold (CFFT) is the frequency at which a flickering light source becomes indistinguishable from continuous light. The CFFT is an important biomarker of health conditions, such as Alzheimer's disease and epilepsy, and is affected by factors as diverse as fatigue, drug consumption, and oxygen pressure, which make CFFT individual- and context-specific. Other causal factors beyond such biophysical processes are still to be uncovered. We investigate the connection between CFFT and specific eye-movements, called microtremors, which are small oscillatory gaze movements during fixation periods. We present evidence that individual differences in CFFT can be accounted by microtremors, and design an experiment, using a high-frequency monitor and recording the participant's eye-movements with an eye-tracker device, which enables to measure the range of frequencies of a specific individual's CFFT. Additionally, we introduce a classifier that can predict if the CFFT of specific participant lies in the range of high or low frequencies, based on the corresponding range of frequencies of eyes' microtremors. Our results show an accuracy of [Formula: see text] for a frequency threshold of 60 Hz and [Formula: see text] for a threshold of 120 Hz.

Technical debt prediction (TDP) is crucial for the long-term maintainability of software. In the literature, many machine-learning based TDP models have been proposed; they used TD-related metrics as input features for machine-learning classifiers to build TDP models. However, their performance is unsatisfactory. Developing and utilizing more effective metrics to build TDP models is considered as a promising approach to enhance the performance of TDP models. Social Network Analysis (SNA) uses a set of metrics (i.e., SNA metrics) to characterize software elements (classes, binaries, etc.) in software from the perspective of software as a whole. SNA metrics are regarded as a compensation of TD-related metrics used in the existing TDP work, and thus are expected to improve the performance of existing TDP models. However, the effectiveness of SNA metrics in the field of TDP has never been explored so far. To fill this gap, in this paper, we propose an improved software technical debt prediction approach. First, we represent software as a Class Dependency Network, based on which we compute the value of a set of SNA metrics. Second, we combine SNA metrics with the TD-related metrics to create a combined metric suite (CMS). Third, we employ CMS as the input features and utilize seven commonly used machine learning classifiers to build TDP models. Empirical results on a publicly available data set show that (i) the combined metric suite (i.e., CMS) can indeed improve the performance of existing TDP models; (ii) XGBoost performs best among the seven classifiers, with an [Formula: see text] value of 0.77, an MI ratio of approximately 0.10, and a recall close to 0.87. Furthermore, we also reveal the relative effectiveness of different metric combinations.