IoT Projects – ElysiumPro

In today’s digital age, ensuring students' online safety has become a critical concern for educational institutions and parents alike. This project, "Web Monitoring with parental notification Using Machine Learning," aims to develop an automated system that monitors and analyses the browsing activities of school and college students. The system is designed to detect harmful content accessed by students and notify their parents in real-time. For school students, browsing history is collected in an Excel sheet, and age verification ensures that the system targets only students under 18 years of age. In contrast, for college students, all browsing history is stored, and harmful content is identified using machine learning algorithms. The machine learning pipeline involves several key steps, including data collection from harmful message datasets, pre-processing of data (such as handling missing values and label encoding), text pre-processing using Natural Language Processing (NLP) techniques, and the classification of harmful content using Random Forest classifiers. The model’s performance is evaluated using metrics like accuracy, precision, recall, and F1-score.This system improves the overall monitoring process by providing real-time notifications and email alerts to parents when harmful content is detected.

Cardiovascular Diseases (CVDs) remain a leading cause of mortality, necessitating timely and accurate diagnosis. This study proposes two deep learning models, EnsCVDD-Net and BlCVDD-Net, designed to predict CVDs by incorporating patients’ health and socioeconomic factors while addressing data imbalance using Adaptive Synthetic Sampling. EnsCVDD-Net ensembles LeNet and GRU models, whereas BlCVDD-Net blends LeNet, GRU, and Multilayer Perceptron architectures. Feature selection is performed using the Point Biserial Correlation Coefficient, and model interpretability is enhanced via SHapley Additive exPlanations (SHAP). Experimental results demonstrate that BlCVDD-Net achieves superior performance with 91% accuracy and F1-score, 96% precision, and 86% recall, outperforming baseline and state-of-the-art models. EnsCVDD-Net also delivers competitive results with 88% accuracy and F1-score. Both models are validated through 10-fold cross-validation, offering robust, efficient, and explainable tools for early CVD detection and improved clinical decision-making.

The Heart Disease Prediction Application addresses the growing need for accurate and timely cardiovascular risk assessment to improve patient outcomes and healthcare efficiency. By leveraging advanced machine learning and deep learning techniques, it offers a powerful tool for clinicians and public health officials to make informed decisions based on comprehensive data analysis. Designed to integrate diverse patient records seamlessly, the system enhances predictive accuracy and adaptability to emerging health trends. Its intuitive interface and robust analytics support better resource planning and personalized care, making it an essential solution for advancing heart disease prevention and management in modern healthcare settings.

The Heart Disease Prediction Application addresses the growing need for accurate and timely cardiovascular risk assessment to improve patient outcomes and healthcare efficiency. By leveraging advanced machine learning and deep learning techniques, it offers a powerful tool for clinicians and public health officials to make informed decisions based on comprehensive data analysis. Designed to integrate diverse patient records seamlessly, the system enhances predictive accuracy and adaptability to emerging health trends. Its intuitive interface and robust analytics support better resource planning and personalized care, making it an essential solution for advancing heart disease prevention and management in modern healthcare settings.

This study proposes a novel credit card fraud detection approach that integrates advanced machine learning techniques to address data imbalance and improve classification accuracy. The method employs a hybrid sampling strategy based on the Mahalanobis distance to effectively balance the dataset by identifying and synthesizing representative minority class samples. Subsequently, a Random Forest classifier is utilized for the detection task, leveraging its ensemble learning capability to enhance predictive performance and robustness against noisy data. The combination of hybrid sampling with Random Forest demonstrates superior fraud detection accuracy and reduced false positives compared to traditional methods, highlighting its practical applicability in real-world financial security systems.

This paper presents a comprehensive review and experimental analysis of various machine learning algorithms applied to text classification tasks. It covers traditional methods such as Naive Bayes, Support Vector Machines (SVM), and Decision Trees, alongside more recent approaches including Random Forest, Gradient Boosting, and neural network models like Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). The study further evaluates these algorithms through observational analysis and controlled experiments on multiple benchmark text datasets. Insights on their performance, scalability, and suitability for different text classification scenarios are discussed, providing a holistic understanding of current techniques and guiding future research in natural language processing.

This study explores the integration of artificial intelligence (AI) technologies within legal systems, focusing on the deployment of machine learning algorithms to enhance judicial decision-making, case outcome prediction, and legal document analysis. Key AI techniques examined include natural language processing (NLP) models such as transformers and BERT for understanding and classifying legal texts, supervised learning algorithms like Random Forest and Support Vector Machines for case outcome prediction, and unsupervised methods for legal topic modeling and anomaly detection. Through empirical evaluation, the research assesses AI’s impact on efficiency, fairness, and transparency in justice administration, highlighting both the opportunities and challenges posed by AI-driven legal tools.

This research investigates the detection of undesired behaviors in Alexa skills by analyzing user reviews through machine learning techniques. Natural language processing (NLP) methods are employed to preprocess and represent textual data, including tokenization and vectorization approaches such as TF-IDF and word embeddings. Supervised learning algorithms, including Support Vector Machines (SVM), Random Forest, and Logistic Regression, are applied to classify reviews that indicate problematic or undesirable skill behaviors. Additionally, sentiment analysis and topic modeling techniques are utilized to uncover common issues reported by users. The study demonstrates how machine learning-driven user review analysis can effectively identify and categorize faults in voice assistant applications, guiding improvements in skill quality and user experience.

This study explores the transformative role of big data analytics capabilities in enhancing healthcare service quality through the utilization of advanced technological resources. It highlights the deployment of machine learning algorithms such as decision trees, support vector machines, and deep learning models to analyze large-scale healthcare data for predictive diagnostics, patient risk stratification, and personalized treatment recommendations. The integration of real-time data processing and predictive analytics enables improved decision-making, operational efficiency, and patient outcomes. By leveraging big data analytics, healthcare providers can identify patterns, reduce errors, and optimize resource allocation, thereby significantly elevating the overall quality and effectiveness of healthcare services.

This paper investigates strategies for adapting food supply chains to the challenges posed by climate change through the use of advanced simulation models. It integrates machine learning techniques with system dynamics and agent-based simulation to model complex interactions within supply chains under varying climate scenarios. Predictive models, including regression analysis and reinforcement learning, are employed to forecast supply disruptions and optimize resource allocation. The combined approach allows stakeholders to evaluate adaptation strategies, improve resilience, and enhance decision-making processes. This study highlights the effectiveness of simulation-driven machine learning frameworks in supporting sustainable food supply chain management amid climate uncertainties.

This study explores the use of artificial intelligence (AI) techniques to analyze and understand mobile phone addiction behaviors. Machine learning algorithms such as Support Vector Machines (SVM), Random Forest, and neural networks are applied to classify user behavior patterns based on data collected from smartphone usage logs and self-reported questionnaires. Natural language processing (NLP) methods are also used to analyze textual data from user feedback and social media interactions. By leveraging these AI-driven approaches, the research aims to identify key indicators of addiction, predict risk levels, and provide personalized intervention strategies. The findings demonstrate the potential of AI to support mental health initiatives and improve awareness of mobile phone addiction.

This paper presents an approach for detecting anomalies in cloud network environments through the application of machine learning algorithms. Supervised and unsupervised learning techniques, including Support Vector Machines (SVM), Random Forest, K-Nearest Neighbors (KNN), and clustering methods such as K-Means and DBSCAN, are explored to identify abnormal network traffic patterns indicative of security threats. Feature engineering and dimensionality reduction methods are employed to enhance detection accuracy and reduce computational overhead. Experimental results demonstrate that combining multiple machine learning models improves the detection of anomalies with high precision and low false-positive rates, providing a robust framework for securing cloud infrastructures against evolving cyber-attacks.

This study reviews current and emerging approaches to sustainable packaging for milk and dairy products, focusing on eco-friendly materials, innovative design, and process optimization technologies. It explores the integration of advanced technologies such as biodegradable polymers, smart packaging with embedded sensors, and recycling-friendly materials to minimize environmental impact. Additionally, data-driven approaches including machine learning algorithms are discussed for optimizing supply chain logistics and packaging processes to reduce waste and energy consumption. The paper highlights how combining material science innovations with technological solutions can lead to more sustainable packaging systems, enhancing both environmental sustainability and product safety in the dairy industry.

This paper presents an intelligent ticket assignment system designed to improve customer support efficiency by leveraging deep machine learning techniques. The system employs natural language processing (NLP) models, including deep neural networks and transformer-based architectures, to analyze and classify incoming support tickets based on their content and urgency. Using deep learning-driven feature extraction and multi-class classification, tickets are automatically routed to the most appropriate support agents, optimizing response times and resource allocation. Experimental results demonstrate that the proposed approach significantly outperforms traditional rule-based and shallow learning methods, providing scalable and accurate ticket assignment in dynamic customer service environments.

This project focuses on building a machine learning-based system for detecting malware in educational IoT systems using the IoEd-Net dataset. The process begins with data collection and preprocessing, including handling missing data, feature engineering, normalization, and addressing imbalances. The data is split into training and test sets for model evaluation. Various machine learning algorithms, such as Convolutional Neural Networks (CNN) and K-Nearest Neighbors (KNN), are used to train and predict system behaviors, identifying benign or malicious activities. Model performance is assessed through accuracy, precision, recall, and F1 score, with visualizations like confusion matrices and ROC curves. The system also generates real-time alerts and provides a dashboard for monitoring detected threats. The backend is implemented in Python with Flask for the front-end, MySQL for the database, and Anaconda Navigator for development.

The Smart Security Surveillance System leverages advanced technologies to predict and assess the security status of a given environment, determining whether it is secured or not secured. Utilizing real-time data from sensors, cameras, and intelligent algorithms, the system analyzes potential threats, unusual activities, and security breaches with high accuracy. By integrating machine learning models and automated decision-making processes, the system provides proactive alerts and comprehensive security evaluations, ensuring enhanced safety and rapid response to potential risks. This predictive approach optimizes security management, offering users peace of mind through continuous and intelligent surveillance.

This study focuses on applying machine learning algorithms to detect anomalies in transactional data for preventing fraud and financial misconduct. It utilizes supervised learning models such as Random Forest, Support Vector Machines (SVM), and Gradient Boosting, along with unsupervised methods including Isolation Forest and Autoencoders. Feature engineering and dimensionality reduction techniques are employed to enhance model accuracy and efficiency. The proposed system demonstrates high detection rates combined with low false positive rates, making it an effective tool for real-time fraud detection. This approach improves transaction security and assists financial institutions in quickly identifying suspicious activities to reduce potential losses.

This study investigates the application of advanced deep learning models for the classification of bio-signals and the prediction of various diseases. It utilizes powerful techniques such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks to effectively analyze complex physiological data, including ECG, EEG, and EMG signals. These deep learning models facilitate automated feature extraction and detailed temporal analysis, significantly improving the accuracy and reliability of anomaly detection and disease outcome prediction. Experimental results demonstrate that deep learning techniques consistently outperform traditional machine learning methods, offering a highly robust and scalable framework for non-invasive diagnosis. This approach holds great promise for enabling personalized healthcare solutions and enhancing clinical decision-making to improve patient outcomes.

This paper introduces a novel large language model specifically designed to enhance attribute extraction for e-commerce product categorization. By leveraging advanced transformer architectures, the model achieves context-aware understanding and fine-grained extraction of product attributes from unstructured textual data such as product titles and descriptions. The approach incorporates transfer learning and fine-tuning on extensive domain-specific datasets, which significantly improves both accuracy and generalization across a wide variety of product categories. Experimental results demonstrate that this model outperforms traditional rule-based systems and classical machine learning methods in precision and efficiency. This advancement enables more accurate and efficient product classification, ultimately enhancing search relevance, improving customer experience, and supporting better inventory management for e-commerce platforms.

The rapid proliferation of online threats has heightened the need for effective content detection systems, particularly in identifying phishing websites. Phishing attacks often deceive users by mimicking legitimate websites, leading to data theft and security breaches. This study explores the application of machine learning (ML) techniques for suspicious content detection in websites, focusing on two popular algorithms: Naive Bayes and Random Forest. Using the Phishing Websites Dataset Cleaned, which consists of various features, we evaluate the performance of these algorithms in classifying websites as either legitimate or phishing. Naive Bayes, a probabilistic classifier, and Random Forest, an ensemble learning method, are applied to train models on the dataset. The results demonstrate the potential of both algorithms in detecting phishing websites, with Random Forest showing superior accuracy due to its ability to handle complex, non-linear patterns in the data. This research highlights the effectiveness of ML-based approaches in cyber security, providing a foundation for building automated systems to safeguard users from phishing threats.

The persistence of SMS spam remains a significant challenge, highlighting the need for research aimed at developing systems capable of effectively handling the evasive strategies used by spammers. Such research efforts are important for safeguarding the general public from the detrimental impact of SMS spam. In this study, we aim to highlight the challenges encountered in the current landscape of SMS spam detection and filtering. To address these challenges, we present a new SMS dataset comprising more than 68K SMS messages with 61% legitimate (ham) SMS and 39% spam messages. Notably, this dataset, we release for further research, represents the largest publicly available SMS spam dataset to date. To characterize the dataset, we perform a longitudinal analysis of spam evolution. We then extract semantic and syntactic features to evaluate and compare the performance of well-known machine learning based SMS spam detection methods, ranging from shallow machine learning approaches to advanced deep neural networks. We investigate the robustness of existing SMS spam detection models and popular anti-spam services against spammers’ evasion techniques. Our findings reveal that the majority of shallow machine learning based techniques and anti-spam services exhibit inadequate performance when it comes to accurately classifying SMS spam messages.

Customer churn prediction is vital for streaming services to reduce costly subscriber losses. This study proposes a big data-driven hybrid model that combines deep learning and machine learning techniques to enhance churn forecasting. The model integrates Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to capture temporal usage patterns, alongside Light Gradient Boosting Machine (LightGBM) to utilize both sequential and original features. Feature selection is performed using Chi-squared testing and Sequential Feature Selection to optimize input variables. The hybrid model’s performance is compared with several individual deep learning and traditional machine learning models. Explainability methods, including Shapley Additive Explanations (SHAP) and Explainable Boosting Machine (EBM), provide insights into influential factors driving churn. Experimental results demonstrate that the hybrid approach achieves superior predictive accuracy, with an AUC of 95.60% and an F1 score of 90.09%, offering a powerful and interpretable solution for improving customer retention strategies in streaming services.

Parkinson’s Disease (PD) is a progressive neurological disorder affecting motor functions and cognition, making early and accurate diagnosis essential for effective patient care. This study provides a comprehensive review of PD recognition systems utilizing diverse data modalities, including MRI, gait analysis, handwriting, speech, EEG, and multimodal fusion techniques. Drawing on over 347 articles from top scientific databases, the review explores data collection methods, feature representations, and system performance, emphasizing recognition accuracy and robustness. Unlike prior surveys that focus on single modalities, this work highlights the benefits of multimodal approaches combined with advanced machine learning and deep learning techniques. The survey aims to guide researchers in developing more accurate and reliable PD diagnostic tools by leveraging varied data sources. Ultimately, this study contributes to advancing PD diagnostics and improving patient outcomes through innovative, data-driven, and multimodal strategies.

This study presents an optimized machine learning approach for predicting heart failure survival using Gradient Boosting Machine (GBM) enhanced by Adaptive Inertia Weight Particle Swarm Optimization (AIW-PSO). Utilizing a Kaggle dataset of 299 patients with clinical features such as age, ejection fraction, and serum creatinine, the model addresses class imbalance with SMOTE and applies SelectKBest and Chi-square methods for feature selection. AIW-PSO optimizes GBM hyperparameters by adaptively balancing exploration and exploitation, while model selection incorporates Akaike and Bayesian Information Criteria to balance accuracy and complexity. The optimized GBM achieved a test accuracy of 94%, outperforming traditional models. This work demonstrates the effectiveness of metaheuristic hyperparameter tuning in clinical prediction and highlights AIW-PSO’s potential to improve model reliability and interpretability. The proposed method offers a valuable tool for timely and accurate heart failure survival prediction, supporting clinical decision-making and patient management.

This study improves photovoltaic (PV) power forecasting to support better integration of solar energy into power grids. It employs lightweight Long Short-Term Memory (LSTM) models enhanced with attention mechanisms, optimized using a modified Particle Swarm Optimization (PSO) algorithm for efficient hyperparameter tuning. Evaluated on datasets from PV plants in India and Serbia, the proposed approach outperforms other metaheuristic methods, achieving low mean squared error (MSE) scores of 0.0073 and 0.0077 for the Indian datasets. These results demonstrate the effectiveness of combining lightweight LSTM architectures with advanced optimization techniques to enhance forecasting accuracy. The research contributes to more reliable and cost-efficient solar power predictions, which are essential for improving grid management and advancing the sustainable integration of renewable energy sources.

This study addresses the challenge of accurate metro passenger flow prediction by proposing the Spatio-Temporal Fusion Graph Convolutional Network (STFGCN). The model replaces the traditional Fourier Transform with Discrete Cosine Transform (DCT) to avoid the Gibbs phenomenon, enhancing periodic trend representation through a novel periodic trend-enhancing attention mechanism combined with channel attention. Additionally, a trend similarity-aware attention module captures the evolving temporal patterns, while a dynamic correlation graph convolutional network adapts spatial correlations across different time periods. Evaluated on inbound and outbound passenger flow data from the Hangzhou Metro, STFGCN outperforms existing baseline models significantly. Compared to the CorrSTN model, it achieves improvements of 22.15%, 16.9%, and 0.6% in MAE, RMSE, and MAPE, respectively. These results demonstrate STFGCN’s effectiveness in enhancing prediction accuracy, contributing to more intelligent and efficient subway transportation systems.

This study provides a comprehensive review of solar irradiance forecasting methods, which are vital for optimizing solar energy systems and reducing reliance on traditional energy sources. It examines various approaches, including machine learning models, numerical weather prediction, and hybrid techniques, assessing their accuracy, strengths, and limitations. By analyzing how factors like sunlight availability, cloud cover, and panel orientation impact solar irradiance, the study highlights the role of AI and ML in capturing complex patterns from historical weather data for improved prediction. The research underscores the need for interdisciplinary collaboration and the integration of emerging technologies to enhance forecast reliability. This work serves as a valuable foundation for future research aimed at advancing medium- and long-term solar irradiance prediction, contributing to the sustainable integration of solar power into electricity grids.

This study pioneers the investigation of Bitcoin Ordinals-related data as key features for predicting Bitcoin transaction fee rates and prices. We construct comprehensive datasets combining Bitcoin chain data, Ordinals index, and market data, revealing a strong correlation between Ordinals inscriptions and market activity. Our analysis suggests that SegWit and Taproot upgrades facilitated the rise of Bitcoin Ordinals, influencing blockchain dynamics and price movements. Using the TemporalFusionTransformer model as a baseline, experiments demonstrate that including Ordinals data significantly improves prediction accuracy across MAE, RMSE, and MAPE metrics. Additionally, we introduce a fine-tuned Chronos model that matches or surpasses baseline performance in short-term forecasts, offering fast, accurate, and easily deployable predictions with minimal data complexity. This work provides valuable insights for investors and market participants to enhance decision-making and capitalize on arbitrage opportunities within the Bitcoin Ordinals ecosystem.

This project presents a web application developed using the Flask framework, focusing on predictive maintenance through machine learning. Utilizing a dataset sourced from a repository, the application processes various attributes, including air temperature, process temperature, rotational speed, torque, tool wear, and failure types. The data undergoes rigorous pre-processing to handle missing values, perform label encoding, and eliminate unnecessary columns. Following this, the dataset is split into training (80%) and testing (20%) subsets. Various classification algorithms, including Decision Tree, SVM, and Random Forest, are implemented to predict failure occurrences. The project's objective is to accurately classify the test data into failure and non-failure categories. Finally, the performance of the different algorithms is compared using metrics such as accuracy, loss, confusion matrix, and classification reports, highlighting the most efficient method for predictive maintenance

This study presents a comprehensive global analysis of BRCA1 and BRCA2 germline mutations using data from 29,700 families, emphasizing population-specific genetic variations. By meticulously preprocessing and standardizing mutation data, the research categorizes mutations into pathogenic and benign groups, stratifying families by gene type, geographic region, and ethnicity. Employing frequency and cluster analyses, PCA, hierarchical clustering, and logistic regression, the study uncovers significant regional differences and founder effects in mutation patterns, identifying over 1,600 unique BRCA1 and 1,700 BRCA2 pathogenic variants. Novel common mutations were discovered in underrepresented populations, enabling the proposal of targeted gene panels tailored to specific ethnic groups, which may enhance the efficiency and accessibility of genetic screening. The findings advocate for expanded oncogenetic testing across diverse populations and provide insights for predicting high-likelihood mutations in underserved regions, highlighting the critical need for population-aware genetic diagnostics in global cancer risk management.

This study develops an interpretable machine learning model to predict mortality in Sepsis-Associated Acute Kidney Injury (SA-AKI) patients using the MIMIC-IV database, with external validation on the eICU dataset. From 9,474 SA-AKI patients, 24 key features—including lab results, vital signs, and comorbidities—were selected using statistical methods and expert input. An Extreme Gradient Boosting (XGBoost) model, optimized via GridSearch, achieved a strong internal AUROC of 0.878. Model interpretability was enhanced using SHAP and LIME, identifying critical mortality predictors such as SOFA score, serum lactate, respiratory rate, APACHE II score, and urine output. This approach offers accurate and explainable mortality risk predictions, supporting early identification of high-risk SA-AKI patients in intensive care. The study highlights the potential of advanced machine learning techniques to improve clinical decision-making, with future work aimed at increasing model adaptability and real-world implementation.

This study addresses the challenges of credit risk assessment in complex supply chain finance by evaluating interpretable machine learning models. We applied Extreme Gradient Boosting (XGBoost), Random Forest (RF), Least Squares Support Vector Machine (LSSVM), and Convolutional Neural Network (CNN) models to assess credit risk, complemented by ablation experiments. Using Shapley Additive Explanation (SHAP), we identified key risk factors such as asset-liability ratio, cash ratio, and quick ratio as major contributors to credit risk. Results demonstrate that XGBoost not only delivers superior predictive performance but also offers enhanced interpretability through SHAP, enabling clear insights into risk drivers. This study clarifies the strengths and limitations of different models and provides practical guidance for companies and financial institutions to make informed credit decisions, promoting more sustainable and efficient financial resource allocation in supply chain finance.

This study presents a novel driver risk assessment method that dynamically calculates weekly insurance premiums by integrating Advanced Driver Assistance Systems (ADAS) risk indicators and contextualized geolocation data. Using a naturalistic dataset from 354 commercial drivers over one year, we modeled the relationship between past claims and driving behavior. Claims frequency was analyzed with Poisson regression, while claims occurrence probability was predicted using machine learning models, including XGBoost and TabNet, with interpretability provided by SHAP. Weekly aggregated profiles included driving behavior, ADAS events, and contextual factors. Results show that XGBoost achieved improved prediction accuracy with all attributes, reducing Log Loss from 0.59 to 0.51, while claims frequency models showed no significant difference. Incorporating ADAS and contextual data enables detailed, dynamic premium calculations, supporting more personalized and accurate risk-based pricing. This approach advances insurance pricing by leveraging emerging technologies and contextual driver information.

This paper proposes a hybrid deep learning approach combining Bidirectional Long Short-Term Memory (BiLSTM) networks and an attention mechanism to extract nuanced sentiment and thematic content from literary texts. The BiLSTM processes text from forward and backward directions to capture context, while the attention mechanism enables the model to focus on the most significant passages. Model efficiency is enhanced through hyperparameter optimization using the Improved Particle Swarm Optimization (IPSO) algorithm.A case study on a dataset of 500 English novels demonstrated the model's effectiveness, achieving high accuracy and F1 scores that surpassed traditional methods like CNNs. The analysis successfully identified key emotional themes (joy, fear, sorrow) and thematic content (love, betrayal, revenge). The results highlight the potential of deep learning to provide deeper, multi-layered insights for literary analysis, paving the way for future applications in the humanities.

Effective disaster prediction is essential for disaster management and mitigation. This study addresses a multi-classification problem and proposes the Neural-XGBoost disasterprediction model (NXGB), a hybrid model that combines neural networks (NN) for feature extraction with XGBoost for classification. The NN component extracts high-level features, while XGBoost uses gradient-boosted decision trees for accurate predictions, combining the strengths of deep learning and boosting techniques for improved accuracy. The N-XGB model achieves an accuracy of 94.8% and an average F1 score of 0.95 on a real-world dataset that includes wildfires, floods and earthquakes, significantly outperforming baseline models such as random forest, Support vector machine and logistic regression 85% accuracy. The balanced F1 scores for wildfires 0.96, floods 0.93, and earthquakes 0.96 demonstrate the model’s robustness in multi-class classification. The Synthetic Minority Oversampling Technique (SMOTE) balances datasets and improves model efficiency and capability. The proposed N-XGB model provides a reliable and accurate solution for predicting disasters and contributes to improving preparedness, resource allocation and risk management strategies.

This research explores the integration of causal inference with machine learning through a novel Double Machine Learning (DML) approach for gesture identification. Unlike traditional feature selection methods that focus on correlations, DML prioritizes selecting variables with causal relationships to the final gesture outcome. By identifying causally significant variables, the method enhances both model interpretability and classification performance. Comparative experiments demonstrate that features chosen via DML outperform those selected by conventional techniques such as Variance Threshold, PCA, LASSO, and neural network-based methods across various classifiers. This approach offers a fresh perspective for feature selection, enabling more reliable and insightful models by uncovering underlying causal mechanisms in complex datasets. The findings suggest that leveraging causal relationships through Double Machine Learning can significantly improve prediction accuracy and robustness in gesture recognition tasks, paving the way for broader applications in machine learning that require causal understanding.

This systematic review thoroughly explores a wide range of machine learning and deep learning approaches applied to the detection of fake news. It highlights key methods for feature extraction, representation learning, and classification that enable effective identification of misinformation across diverse and complex datasets. The study also discusses recent advancements in utilizing textual, contextual, and multimodal data sources to significantly improve detection accuracy and robustness. Furthermore, it addresses common challenges faced in this domain, such as data imbalance, constantly evolving news patterns, and the need for model interpretability to build user trust. Overall, the review provides valuable insights into how these AI-driven techniques contribute to enhancing the reliability, efficiency, and timeliness of fake news identification in today’s rapidly changing information environments.

This study proposes a novel approach for kidney tumor detection by integrating fuzzy logic with advanced machine learning techniques. The method combines a Twin Transferable Network to extract robust features with a weighted ensemble of machine learning classifiers to enhance diagnostic accuracy. By leveraging the strengths of multiple classifiers, the ensemble model improves prediction robustness and handles data uncertainty effectively. The fusion strategy, supported by fuzzy enhancement, allows better discrimination of tumor regions in medical images, leading to improved detection performance. Experimental results demonstrate that this integrated approach outperforms traditional models, offering a promising tool for accurate and reliable kidney tumor diagnosis.

This study addresses the prediction of maternal health risks (MHR) using real-world datasets from maternity hospitals and clinics. We propose a Quad-Ensemble Machine Learning framework for Maternal Health Risk Classification (QEML-MHRC) that integrates multiple ML models with ensemble techniques to improve prediction accuracy. Nineteen training and testing experiments were conducted, identifying key risk factors such as high and low blood pressure and elevated blood sugar levels. The model achieved outstanding performance across risk classes, with the high-risk (HR) category predicted correctly 90% of the time. Gradient Boosting Tree (GBT) combined with ensemble stacking yielded the best overall results, with an average evaluation score of 0.86 across classes. By focusing on class-wise performance, the approach enables precise differentiation of risk levels, facilitating early detection of high-risk pregnancies. This method supports medical experts in timely intervention, aiming to reduce maternal complications and improve pregnancy outcomes.

This study proposes a filter-based feature selection combined with a stacking ensemble machine learning framework for thyroid disease detection. Addressing limitations of single-model approaches such as data imbalance and bias, the framework integrates multiple base models to enhance predictive accuracy and robustness. By utilizing a reduced set of clinical attributes, the method also aims to lower screening time and costs. Experiments on a clinical thyroid dataset demonstrate the framework’s superior performance, achieving a ROC-AUC score of 99.9%. The ensemble approach leverages the complementary strengths of individual models, significantly improving detection efficacy over standalone models. These results highlight the promise of ensemble learning in medical diagnosis, advocating for a shift towards collaborative predictive modeling in thyroid disease detection.

Quantum Machine Learning (QML) has shown promise in diverse applications such as environmental monitoring, healthcare diagnostics, and financial modeling. However, its practical implementation faces challenges, including limited quantum hardware and the complexity of integrating quantum algorithms with classical systems. One critical challenge is handling imbalanced datasets, where rare events are often misclassified due to skewed data distributions. Quantum Bayesian Networks (QBNs) address this issue by enhancing feature extraction and improving the classification of rare events such as oil spills. This paper introduces a Bayesian approach utilizing QBNs to classify satellite-derived imbalanced datasets, distinguishing ``oil-spill'' from ``non-spill'' regions. QBNs leverage probabilistic reasoning and quantum state preparation to integrate quantum enhancements into classical machine learning architectures. Our approach achieves a 0.99 AUC score, demonstrating its efficacy in anomaly detection and advancing precise environmental monitoring and management. While integration enhances classification performance, dataset-specific challenges require further optimization.

This paper presents a novel framework for automated software defect prediction targeting eight common defect types: SIGFPE, NZEC, LOGICAL, SYNTAX, SIGSEGV, SIGABRT, SEMANTIC, and LINKER, alongside an error-free class. Utilizing a specialized nine-class dataset, the framework employs a CodeBERT-based model optimized through hyperparameter tuning to enhance defect detection in code snippets. Comparative experiments against models such as RoBERTa, Microsoft CodeBERT, and GPT-2 demonstrate significant improvements, achieving up to 20% and 7% accuracy gains in binary and multi-class prediction tasks, respectively. The results confirm the efficacy of advanced neural language models like CodeBERT in early defect identification, contributing to more reliable software testing and development. This study highlights the promising role of machine learning and natural language models in automating defect prediction and improving software reliability.

This paper addresses the challenge of accurate long-term blood glucose prediction in middle-aged and elderly patients, aiming to reduce reliance on error-prone daily activity inputs such as meals and insulin injections. Using continuous glucose monitoring (CGM) data from 14,733 patients and incorporating three assistance factors, multiple machine learning models were evaluated, including Artificial Neural Networks (ANN), Binary Decision Trees (BDT), Support Vector Machines (SVM), and Gaussian Process Regression (GPR). The Binary Decision Tree achieved the highest classification accuracy of 92.58%, followed by Boosting Regression Tree Ensemble (92.04%) and GPR (88.59%). In terms of prediction error, GPR, BSTE, and ANN reported average root mean square errors of 1.64, 1.67, and 1.69 mg/dL, respectively, for the prediction horizon. The findings demonstrate that BDT and ensemble methods offer promising accuracy and robustness for glucose level prediction without requiring detailed patient activity logs.

This paper presents a Quantum-Trained Convolutional Neural Network (QT-CNN) framework for enhanced deepfake audio detection, leveraging quantum machine learning (QML) to improve efficiency and reduce model complexity. By integrating Quantum Neural Networks (QNNs) with classical CNN architectures, the QT-CNN achieves up to 70% reduction in trainable parameters through a novel quantum-to-classical parameter mapping without sacrificing accuracy. Essential audio features were extracted and preprocessed for robust model training and evaluation. Experimental results demonstrate that QT-CNN maintains comparable accuracy to traditional CNNs across different QNN configurations, while significantly lowering computational overhead. This hybrid quantum-classical approach shows promise for resource-constrained environments and real-world applications in multimedia security. The study highlights the potential of quantum computing integration within AI frameworks to advance deepfake detection, offering a scalable, efficient, and effective solution to growing challenges in audio content authenticity.

This paper introduces a novel emotion detection classifier for EEG signals based on Robert’s similarity measure. Unlike traditional classifiers such as KNN, SVM, and Random Forest, which face challenges with nonlinear patterns and high-dimensional data, the proposed method segments EEG signals into small, medium, and large blocks, showing superior performance with medium and large sizes. Particle Swarm Optimization (PSO) is integrated for feature selection, using Robert’s similarity as the fitness function, enhancing accuracy and computational efficiency. Evaluated on an EEG brainwave dataset, the approach achieved 98.75% accuracy with feature selection, outperforming the 94.04% accuracy without it. The classifier shows promise for applications in healthcare, education, customer service, smart environments, and industrial settings by enabling accurate emotional state detection and adaptive responses. This method offers a versatile and efficient solution for emotionally-aware systems across diverse domains.

— Vital signs have been essential clinical measures. Among these, body temperature (BT) and heart rate(HR) are particularly significant, and numerous studies explored their association in hospitalized adults and children.However, a lack of in-depth research persists in childrenadmitted to the pediatric intensive care unit (PICU) despite their critical condition requiring particular attention. Objective: In this study, we explore the relationship between HRand BT in children from 0 to 18 years old admitted to thePICU of CHU Sainte-Justine (CHUSJ) Hospital. Methods:We applied Machine learning (ML) techniques to unravelsubtle patterns and dependencies within our dataset toachieve this objective. Each algorithm undergoes meticulous hyperparameter tuning to optimize the model performance. Results: On a large database of 4006 childrenadmitted in the PICU, our findings align with prior research,revealing a consistent trend of decreasing HR with increasing patient age, confirming the inverse correlation. Furthermore, a thorough analysis identifies Gradient Boosting Machines (GBM) implemented with Quantile regression(QR) as the most fitting model, effectively capturing thenon-linear relationship between HR, BT, and age. Throughtesting the HR prediction model based on age and BT,the predictive model between the 5th and 95th percentilesaccurately demonstrates the declining trend of HR withage, while HR increases with BT. Based on that, we havedeveloped a user-friendly interface tailored to generateHR predictions at different percentiles based on three keyinput parameters: current HR, current BT, and patient’sage.

Electricity theft poses significant challenges for PT PLN (Persero), especially with 27 million postpaid customers using traditional meters lacking communication capabilities. This study develops a machine learning model to optimize the Target Operation (TO) process, which identifies customers for on-site verification of suspected theft, focusing on the 450 VA subsidized household segment. The model reduces dependency on subjective manual observations by analyzing monthly electricity usage. Multiple classifiers were evaluated, with Random Forest achieving the best performance. A novel sequential evaluation method applies layered filtering using three-theft, two-theft, and one-theft models, enhancing detection accuracy. The combined use of Random Forest and K-Nearest Neighbors yielded the highest metrics: accuracy of 0.89, precision of 0.83, recall of 0.98, F1-score of 0.90, and AUC of 0.89. This approach provides PLN with reliable, efficient, and objective TO recommendations, improving subsidy allocation and minimizing human error.

Click fraud in online advertising significantly undermines industry integrity by misappropriating advertising budgets. This study addresses the challenge by implementing comprehensive feature engineering to capture subtle behavioral patterns distinguishing fraudulent from legitimate clicks. Nine machine learning (ML) and deep learning (DL) models were evaluated after Recursive Feature Elimination (RFE). Tree-based ML models like Decision Tree (DT) and Random Forest (RF) achieved over 98.99% accuracy, while Gradient Boosting (GB), LightGBM, and XGBoost exceeded 98.90%. Deep learning models, including CNN, DNN, and RNN, also showed strong results, with RNN reaching 97.34% accuracy. Precision scores for models such as Artificial Neural Networks (ANN) surpassed 98%, indicating reliable fraud detection. These findings highlight the effectiveness of tree-based and advanced ML/DL algorithms in detecting click fraud and provide valuable insights to develop robust anti-fraud strategies for online advertising.

Accurate classification of sleep disorders like obstructive sleep apnea and insomnia is vital for enhancing human health and quality of life. Traditional expert-based sleep stage analysis is labor-intensive and error-prone, motivating the use of machine learning algorithms (MLAs) for automated diagnosis. This study compares 15 ML classifiers on the Sleep Health and Lifestyle Dataset to classify None, Sleep Apnea, and Insomnia. Two experiments were conducted using feature selection via Gradient Boosting Regressor with Mean Decrease Impurity. In the first experiment, Gradient Boosting, Voting, Catboost, and Stacking classifiers achieved a top accuracy of 97.33%, with Gradient Boosting showing the highest AUC (0.9953) and superior computational efficiency. In the second, Decision Tree attained 96% accuracy with significant speed improvements on engineered features. Results identify Gradient Boosting as the most effective and efficient classifier for sleep disorder detection.

The rising prevalence of depression calls for effective early detection methods. Social media provides a rich source of user-generated data to identify depressive symptoms using machine learning (ML). This systematic review analyzes 47 studies on ML-based depression detection on social media, focusing on biases and methodological challenges. Key issues include an overreliance on Twitter and English-language data, limited geographic diversity, and predominant use of non-probability sampling, which affect model generalizability. Only a minority of studies addressed linguistic nuances or performed proper hyperparameter tuning. Data partitioning and evaluation methods were often inconsistent, increasing risks of overfitting and biased results. Reporting transparency also varied widely. These findings underscore the need for more diverse datasets, standardized preprocessing, rigorous model validation, and improved reporting. Addressing these gaps will enhance the robustness and applicability of ML models for depression detection, supporting better mental health monitoring through social media analysis.