Browse Items (3095 total)
Sort by:
-
Edge and Fog Computing in Cyber-Physical Systems
The benefits of cyber-physical system advances include low latency and high bandwidth data processing in areas such as automotive, healthcare, and business automation. Traditional environments are often located in centralized and remote locations and cannot meet the demand. Edge computing and cloud computing have become fundamental concepts that will bring computing closer to the center of the data. Edge computing can reduce latency and bandwidth consumption by processing data on or near IoT devices. Fog computing adds another layer to this by distributing work and storage across multiple nodes, thus providing a scalable and flexible infrastructure. This article discusses the principles, benefits, and challenges of integrating edge and cloud computing into a CPS environment. It leverages the power of proximity-based edge computing and the centralized capabilities of cloud computing to provide scalable, instantaneous responses to CPS applications or time to optimize services. The demonstration shows a variety of things from smart cities to the use of IoT in healthcare in CPS. The article also covers some specific security and privacy issues and future directions in distributed computing, including the role of AI and 5G, which are supposed to offer additional resources in various applications. 2025 IEEE. -
Premium Unlocked AI for Medical Document Decoding
As healthcare systems evolve to become more digital, an enormous volume of medical data is available in various formats, including unstructured data, scanned documents, handwritten prescriptions, diagnostic images, audio transcriptions, and clinical video recordings. The complexity and unorganised form of data continue to pose serious challenges with regard to automation, accuracy, and consistency in healthcare and insurance businesses. This study introduces an AI-based multimodal framework that incorporates the use of Optical Character Recognition (OCR), the MiniCPMV-4.5 model, and Retrieval-Augmented Generation (RAG) with Large Language Models (LLMs) to enhance the intelligent processing and contextual comprehension of intricate medical data, thus overcoming these limitations. It applies OCR to scanned images and handwritten documents to precisely recover the textual information from them and uses domain-specific named entity recognition (NER) to recognize significant medical information, e.g., patient information, diagnoses, procedures, and financial information. The extracted information is then converted to vector embeddings and stored in a powerful vector database, Milvus, that enables fast and efficient semantic search as well as context-sensitive reasoning. The proposed framework, along with the visual and auditory inputs, video understanding, multilingual capacity, and the S2S (speech-to-speech) and TTS (text-to-speech) translation, makes it more accessible and engaging to the user. This system reduces the level of human involvement and provides real-time insights quickly and more precisely so that more efficient decisions and operations can be made in the fields of healthcare and insurance. 2025 IEEE. -
Early Sepsis Prediction using Hybrid LightGBM and LSTM Model
Sepsis is a critical organ malfunction that results from an abnormal response of the body to infection and might be lethal. The early detection of sepsis is essential for the patient's life. However, the traditional clinical diagnostic systems are not capable of analyzing the complicated changes in the patient's vitals over time. Therefore, a hybrid predictive framework that merges Light Gradient Boosting Machine (LightGBM) and Long Short-Term Memory (LSTM) networks for fast and accurate sepsis detection in real-time using freely accessible MIMIC-III data, has been proposed in this research. Using LightGBM, the nonlinear relationships among the features are learnt very fast and efficient, while the LSTM gives the temporal dependencies in the sequence of the patient vital signs. The combined output of the two models is said to be more sensitive and robust than that of the single models. A Streamlit-based clinical dashboard is being provided, allowing for real-time predictions and visualization for healthcare professionals. The proposed system has shown a considerable increase in the accuracy of early sepsis detection and offers a non-restricted method for AI-assisted ICU monitoring. 2025 IEEE. -
Privacy Risk Prediction from Social Media Metadata using Feature Selection Approaches
Millions of new people sign up to online social networks (OSNs) every year, which contributes to the growing spread of Personally Identifiable Information. This often ends up occurring unconsciously, either due to the low stakes involved or because the user doesn't understand or underestimates what can go wrong. This trend indicates the need for a trustworthy means to quantify the privacy danger of sharing information online. The volume of OSN data can simply be too staggering for any degree of meaningful manual review, given both the time and man-hours this would entail. This research presents a two-step, unsupervised, and efficient method to estimate privacy risks at the post level. The first step involves using the most advanced reasoning-based Large Language Model, Gemini 2.5 Pro, to generate a comprehensive 'vulnerability score', which is used as a reference for model training. The next step involves comparing the two most used machine learning feature selection techniques, Recursive Feature Elimination (RFE) and Correlation-Based Selection, to select the best features for predicting this score from metadata alone. The results indicate that Correlation-Based Selection produces better results for both the regression and classification-based models, and the top-performing regression model achieves an R-squared of 0.86. Through this, a practical and scalable method to identify privacy-sensitive content effectively on large datasets has been presented in this study. 2025 IEEE. -
Unmasking the Masked: A Classical Machine Learning Pipeline for Detecting Forged Receipts
The abundance of digital and paper document forgery requires strong automated detection tools against financial fraud. This research provides a classical machine learning method for forged receipt detection using multimodal features from image and text modalities. The approach entailed designing a feature set to obtain textural and statistical attributes from receipt images via Local Binary Patterns (LBP) and Canny edge detection, along with structural features obtained from the associated text files. Another demanding issue in this area is the excessive class imbalance between genuine and forged documents. To overcome this issue, Synthetic Minority Over-sampling Technique (SMOTE) is used to create a balanced training dataset. The models are assessed using the macro F1-score, precision, recall, PR AUC and ROC AUC to address class imbalance. The enhanced detection of the minority class is achieved using SMOTE, while hyperparameter tuning leads to the improvements in performance. The final Tuned Support Vector Machine model achieves a macro F1-score of 0.5429, and it has the highest recall on forged receipts, demonstrating that it detects more histories of tampered documents effectively. This research sets a good baseline for receipt forgery detection and emphasizes that class imbalance solving is a key towards creating a working system. 2025 IEEE. -
Customer Evaluation of Internet Banking Services: Analysing South Indian Bank's Digital Banking Experience
This research intends to assess the valuation of Christ University's customers on internet banking services provided by South Indian Bank. The research objectives are to evaluate the impact of efficiency, security, ease of use, reliability and social influence on customer satisfaction in the context of internet banking. Consumers of Christ University were selected to answer in the survey to give their view and experience toward the internet banking services offered by the South Indian Bank. The strategies used in data analysis incorporated regression analysis to realize correlation between such factors and extent of customer satisfaction. The research offers huge information on how it is possible to enhance the levels of customer satisfaction specifically in internet banking services of South Indian Bank. Therefore, the study underpins an importance of enhancing digital banking platforms to fulfil the need and expectation of customers. The study is also beneficial for other financial institution with the aim at improving the level of customer satisfaction and ability to retain the clients in the realm of digital banking. 2025 IEEE. -
Real-Time Football Match Analysis with Region-Independent Player Tracking Using Deep Learning
This project explores the application of AI in analyzing football games by tracking players across the entire video frame. Unlike traditional methods that focus on limited areas, the system here uses YOLO for detecting players everywhere in the frame and ByteTrack to follow them throughout the match. The goal is to get a clearer picture of each player's movement, particularly their speed and distance covered. Manual methods or GPS-based tools often fall short in providing quick, reliable data, especially in real-time scenarios. This study compensates for camera motion and adjusts for different viewpoints to get more accurate tracking results. As a way to test player identity consistency, the system randomly assigns popular player names to different tracking IDs. Experiments on public match videos show that the system can keep track of players even during zoom-ins, crowding, or partial visibility. Code snippets show how the model works in practice. Our results show that using full-frame AI tracking gives coaches more detailed tactical insights and helps them develop more effective strategies. 2025 IEEE. -
AI-Driven Early Diagnosis of Acute Liver Failure: A Machine Learning Perspective
The liver performs a valuable role in operating proper metabolism. This organ in the human body is responsible for maintaining and preserving overall health and well-being. However, when it fails to function optimally, it can cause severe and significant health complications. Liver diseases are multifactorial conditions that can be challenging to diagnose and treat. Early detection of any disease is beneficial for effective treatment and diagnosis of patients' conditions. Machine Learning algorithms create a great platform for analyzing medical data that helps improve disease detection procedures. This paper aims to get a better understanding of ML algorithms for detecting diseases associated with the liver. The paper tries to explore various machine learning techniques for predicting accurate liver diseases. It uses various parameters as symptoms and calculates ALF (Acute Liver Failure) based on the parameters and ALF predicts in-case the person is suffering from a Liver disease or not. Accuracy was calculated with various ML techniques i.e. Logistic Regression Classification, KNN Classification, Decision Tree, Random Forest and Support Vector Machine. Among all these, Logistic Regression was found to be most effective in identifying and predicting the outcome of the dataset compared to other algorithms. SVM has a higher cross-validation score but Accuracy, precision and recall are very low thus, cannot select this model. 2025 IEEE. -
FINNET: A Hierarchical Graph Learning Framework for Adaptive Cross-Market Financial Risk Prediction
Systemic financial risk emerges from complex multi-scale interactions among entities, sectors, and markets. We introduce FINNET, a hierarchical graph neural network framework that models these vertical dependencies through volatility-aware adaptive pooling. Our approach features: (1) a tri-scale graph structure capturing entity, sector, and market dynamics; (2) dynamic embeddings combining static features with time-varying signals; (3) transfer learning for emerging markets; and (4) transparent risk decomposition for regulatory compliance. Validated on 58,432 financial entities across three continents, FINNET achieves 0.891 AUC with only 3.8% performance degradation during crises, while providing early warnings 15 days before failures. 2026 IEEE. -
Posture Classification Using a Hybrid Deep Learning Model
Automated posture detection is a critical task in ergonomics and healthcare, yet it presents significant challenges for standard computer vision models, particularly in handling class imbalance and understanding geometric constraints. This paper proposes an enhanced hybrid deep learning model that synergizes the feature extraction power of a pre-trained ResNet50 architecture with engineered geometric features derived from the Radon Transform and pre-calculated joint angles. Our approach utilizes a dual-balancing strategy, combining data upsampling with a custom weighted loss function, to effectively address the problem of underrepresented classes. By processing visual and geometric data streams in parallel and fusing them within a deep architecture, our model achieves a holistic understanding of the subject's posture. The fine-tuned model demonstrates strong performance on an unseen test set, achieving a final accuracy of 92% for wrist posture and 92% for neck posture. Crucially, it attains a robust F1-score of 0.74 for the challenging 'Bad Wrist Posture' minority class, a significant improvement compared to the ResNet50-only baseline (F1=0.24) and achieves excellent ROC-AUC scores of 0.9859 for wrist and 0.9838 for neck, proving the efficacy of our hybrid, dual-balancing methodology for realworld application. 2026 IEEE. -
It-Driven Governance Models for Social Impact and Community Development
Local development planning in India is impaired by fragmented administrative datasets, limited citizen participation, and opaque decision processes despite the availability of India Stack primitives and Open Government Data (OGD) such as NFHS and MGNREGA. This study aimed to design, implement, and evaluate Federated Neuro-Symbolic Governance (FNSG), a privacy-preserving cross-silo framework that couples neural concept learners with symbolic rule synthesis, to deliver interpretable district-level decision support. FNSG operates in a hierarchical federated topology (district clients, state aggregator) using NFHS-5, MGNREGA, census/SECC proxies, and geospatial layers. Local clients train neural concept encoders, extract symbolic rule candidates, and transmit clipped, DP-noised updates plus anonymized rule evidence for secure aggregation and global rule distillation. In pilot experiments (K=20 districts, geographic holdout) FNSG achieved mean test AUC 0.84 versus centralized 0.88 and FedAvg 0.80, Precision@100 0.69, and produced concise global rules with fidelity ? 0.72. Privacy - utility sweeps indicate acceptable utility at ? ? 2 (AUC ? 0.82), with degradation at ? ? 1; convergence required ? 110 rounds and per-epoch client time ? 18 s on CPU (? 8 s on GPU). FNSG balances utility, interpretability, and data sovereignty. Policy implications include staged pilots, capacity building for district data centers, mandatory audit logs, and India Stack-mediated consent and integration. 2026 IEEE. -
Corporate Credit Rating Assessment for Financial Risk and Regulatory Compliance
Accurate corporate credit rating is crucial to financial risk management and regulation but the current models tend to use narrow data modalities, fail to consider time and relational relationships and have weak probabilistic calibration. These constraints make them less effective in detecting the risk of default and under pinning decision-making that is in line with the regulator. The objective of this study was to formulate and test a multimodal model with a time-dependent credit rating system to incorporate financial, textual, market and relational information. The publicly available corporate financial statements, market time series data, text disclosures and inter-firm relational information were used to conduct an experimental study. Baseline logistic regression, a hybrid XGBoost with FinBERT embeddings model, and a proposed Temporal Heterogeneous Graph Transformer with cross-modal fusion were implemented and compared using discrimination, calibration, and computational efficiency metrics. The model proposed had the best predictive performance up to a ROC-AUC of 0.903 and PR-AUC of 0.482 which is better than the baseline (0.761) and hybrid (0.842) models. Calibration analysis revealed more correspondence with observed default frequencies, and confusion matrices revealed that the number of true default detection improved as 64 (baseline) to 158. Ablation and Pareto analysis was used to verify that multimodal fusion and temporal graph modelling were the major sources of performance improvements. These findings indicate that the combination of multimodal, temporal, and relational data has a significant positive effect on the accuracy and reliability of credit ratings and provides an institutional and supervisory-appropriate credit risk evaluation framework to the regulator. 2026 IEEE. -
Advancements in EEG and EMG Signals for Motor Imagery Classification and Artifact Removal: A Comprehensive Review and Analysis
An essential noninvasive method for assessing brain electrical activity and gaining important knowledge about how the brain functions is electroencephalography (EEG). Understanding the brain's reactions to particular sensory, cognitive, or motor events requires understanding event-related potentials (ERPs), which are derived from EEG. By displaying variations in frequency content across time, time- frequency analysis improves ERP interpretation. Each of the five EEG frequency bands - delta, theta, alpha, beta, and gamma - has a unique clinical significance and is linked to different physiological and cognitive processes. In order to improve motor control and rehabilitation, this work focuses on the development of NeuroMotor Fusion approaches, which integrate EEG and Electromyography (EMG) signals for motor imagery classification. It looks at new developments in the classification of motor imagery and investigates cutting edge methods such as VR motor priming and brain-computer interfaces (BCIs). The study also discusses the difficulties in removing artifacts from EEG and EMG signals, using hybrid techniques to reduce ocular and muscular artifacts. The study produced a 96.2% accuracy rate in motor function enhancement using the ShallowFBCSPNet model architecture and the MOABBDataset "BNCI2014-001". These findings show that NeuroMotor Fusion has a great deal of promise for use in neurological disease support, individualized motor skill training, and rehabilitation. 2025 IEEE. -
Attainments of Mission Statements and Programme Educational Objectives in Outcome Based Education for a Degree Programme
The outcome-based education (OBE) assesses the outcomes of the students in the form of attainments. The attainments are of a programme which can be undergraduate, postgraduate or diploma. The OBE are assessed in the form of attainments of programme outcomes (POs), programme educational objectives (PEOs) and attainment of mission statements. The attainments are calculated on the basis of direct and indirect tools. The outcomes of theory and lab courses, projects, and placements are considered direct tools. Indirect tools include several surveys like alumni, curriculum, exit, etc. The attainments are finally calculated by combining direct and indirect tools. The attainments of different batches are compared and assessed. The entire attainment process is covered in this research article, and the attainment of a certain batch is shown by the attainments of PEOs and POs. 2025 IEEE. -
Prioritized QoS Enforcement in Smart Healthcare IoT Using Adaptive Deep Q-Network-Based Traffic Decision System
Healthcare IoT systems have been plagued with significant challenges with regard to maintaining an optimum QoS due to the dynamic conditions of the network, diverse device capabilities, and stringent real-time constraints imposed by patient monitoring-type applications. Traditional QoS mechanisms are basically static; they do not take into account changes within the network. Hence, service delivery experiences degradation, with attendant risk to patients' safety. As a solution, this research proposes an adaptive QoS approach employing Deep Q-Network (DQN) reinforcement learning algorithms to dynamically control resource allocation and traffic prioritization in healthcare IoT networks. This system involves multi-agent reinforcement learning architecture where continuous state-action space mapping is utilized for adjusting bandwidth allocation, latency management, and packet prioritization automatically based on network conditions and the criticality levels of applications in real-time. Experimentally, the solution has attained an accuracy of 94.7 percent in QoS prediction, an 87.3 percent reduction in average latency to critical healthcare applications, 91.2 percent improvement in network throughput utilization, and an 89.6 percent success rate in adhering to service level agreements in peak traffic conditions. Through reinforcement learning-based decision making, the adaptive QoS mechanism dynamically accommodates the requirements of healthcare IoT, ensuring reliable service delivery while optimizing the usage of network resources. 2025 IEEE. -
Bridging the Rural Digital Divide: Machine-Learning-Driven Predictive Modeling of Digital Literacy Program Outcomes
The research project performed multiple regression model evaluations to assess how effective digital literacy schemes are in rural education settings. Training program achievements relied on predicted educational proficiency scores while program evaluation relied on both comprehensive participant demographic details and process training statistics. Our study examined numerous regression approaches from basic Linear Regression forms through advanced Random Forest and Gradient Boosting models and concluded with complex methods including Stacking and XGBoost. The research analyzed prediction accuracy and model explanatory power using Mean Squared Error (MSE) and Rsquared (R2) values during the evaluation process. Multiple feature applications were the best fit for the deterministic ensemble techniques which exhibited superior performance but alternatively different analytical models displayed stable prediction results. This research proposes educational method advancement through machine learning approaches capable of creating custom solutions targeting rural user requirements. This study delivers key information to stakeholders in its combined study of digital education enhancements and sophisticated learning evaluation data analysis techniques. 2025 IEEE. -
Bridging the Rural Digital Divide: Machine-Learning-Driven Predictive Modeling of Digital Literacy Program Outcomes
The research project performed multiple regression model evaluations to assess how effective digital literacy schemes are in rural education settings. Training program achievements relied on predicted educational proficiency scores while program evaluation relied on both comprehensive participant demographic details and process training statistics. Our study examined numerous regression approaches from basic Linear Regression forms through advanced Random Forest and Gradient Boosting models and concluded with complex methods including Stacking and XGBoost. The research analyzed prediction accuracy and model explanatory power using Mean Squared Error (MSE) and Rsquared (R2) values during the evaluation process. Multiple feature applications were the best fit for the deterministic ensemble techniques which exhibited superior performance but alternatively different analytical models displayed stable prediction results. This research proposes educational method advancement through machine learning approaches capable of creating custom solutions targeting rural user requirements. This study delivers key information to stakeholders in its combined study of digital education enhancements and sophisticated learning evaluation data analysis techniques. 2025 IEEE. -
Leveraging Machine Learning to Predict Revenue-Generating Sessions in E-Commerce Platforms
Due to the rapid growth of e commerce, develops effective predictive models of online shopper behavior has become important. The goal of this study is to use dataset of online shopping sessions to predict purchase intentions based on session characteristics, user behavior and site metrics. This research aims to apply machine learning and deep learning models to predict online purchasing intentions to assist businesses to improve their strategies of maximizing conversion rates. Using a dataset having numerical and categorical features, features like page views, session duration, bounce rates etc., and the presence of some special days near the user session, we used. We evaluated nine models, including the traditional methods: Logistic Regression, Decision Tree, Naive Bayes, ensemble methods: Random Forest, Gradient Boosting, XGBoost, and more advanced ones like Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), and Neural Networks. Then, key metrics including Accuracy, Precision, Recall, F1 Score and ROC AUC were used to asses each model. We find that ensemble models perform best (ROC AUC = 0.9245) with Gradient Boosting performing best, with XGBoost and Random Forest close behind. With a competitive ROC AUC of 0.9000, neural networks showed strong potential, but fell slightly behind in recall compared with ensemble methods. Logistic Regression and Decision Tree were simpler models that did not achieve as strongly in predictive accuracy as more complex model; however they provided a baseline insight. Through this analysis, ensemble models and deep learning showed to be very efficient to predict online purchase intentions and provide actionable insights to optimize e-commerce platforms. 2025 IEEE. -
An Intelligent Approach for Breast Cancer Diagnosis Using Fuzzy Logic and Extreme Learning Machine
The long-term prognosis and mortality rates can be improved with early identification of breast cancer. The time-consuming and expensive procedures of mammography, MRI, ultrasound, CT, PT, and biopsy have been the subject of much research; nevertheless, these approaches are not suitable for younger women and can be rather expensive. This study employed cutting-edge image processing to improve early breast cancer detection. The researchers utilised anisotropic filtering to reduce background noise in medical images after picking mammograms at random from the Digital Database for Screening Mammography. The use of morphology-based feature extraction allowed for autonomous and accurate categorisation after mass segmentation using a genetic algorithm with recurrent thresholding. By merging a KF with an ELM enhanced with an AV, a new model named KF-av-elm improves diagnostic accuracy. Medical imaging noise and estimating errors are both significantly reduced by the combination method. Their accuracy rating of 98.28% allowed them to outperform other approaches. The KF-av-elm model appears to be a reliable, efficient, and effective diagnostic tool; its adoption may lead to better identification and outcomes for breast cancer patients. 2025 IEEE. -
EEG Emotion Recognition Using PSO-Based Feature Selection and Convolutional Neural Networks
EEG signals have become a promising source for emotion recognition due to their ability to capture the brain's electrical activity connected with different emotional conditions. In this work, a novel approach is proposed that integrates Particle Swarm Optimization (PSO)-based feature selection with Convolutional Neural Networks (CNNs) for improved EEG emotion classification. The method with the preprocessing of a notch filter to eliminate noise and enhance the quality of the EEG signals. Key features, including Magnitude Squared Coherence Estimate (MSCE) and Power Spectral Density (PSD), are extracted to capture essential frequency-domain information. PSO is employed to optimize the selection of features, reducing dimensionality while preserving the most relevant and informative attributes for emotion recognition. The optimized feature was subsequently passed to a CNN classifier, which improves the model's capability to accurately differentiate between different emotional states. This study is implemented using Python software to analyze emotion, and the effectiveness of the proposed approach is assessed using the EEG Brainwave dataset. Experimental results demonstrate that the proposed approach delivers an accuracy of 92.6% and a precision of 91%, highlighting its effectiveness in real-time, high-precision emotion recognition from EEG data. 2025 IEEE.