		FeedBack \| Contact Us \| Links \| Site Map

Welcome to Journal of Theoretical and Applied Information Technology

Submit Paper / Call for Papers

Journal receives papers in continuous flow and we will consider articles from a wide range of Information Technology disciplines encompassing the most basic research to the most innovative technologies. Please submit your papers electronically to our submission system at http://jatit.org/submit_paper.php in an MSWord, Pdf or compatible format so that they may be evaluated for publication in the upcoming issue. This journal uses a blinded review process; please remember to include all your personal identifiable information in the manuscript before submitting it for review, we will edit the necessary information at our side. Submissions to JATIT should be full research / review papers (properly indicated below main title).

Journal of Theoretical and Applied Information Technology
February 2026 | Vol. 104 No.3

Title:	SATELLITE BASED ASSESSMENT OF SOIL HEAVY METAL CONTAMINATION USING DEEP LEARNING AND SWARM INTELLIGENCE
Author:	NAGA SANTHA KUMARI CHEETI,GOVADA ANURADHA
Abstract:	Coastal–deltaic ecosystem are sensitive interfaces where riverine and marine processes interact, resulting in long-term activity of heavy metal contamination. Heavy metals, such as Cd, Pb, Hg, As and Cr tend to accumulate in soil and sediments that may affect the ecosystems and human health. Traditional point-based monitoring techniques have been widely applied for assessing spatial and temporal burden due to contamination at focused site but may be scale limited and resolution limiting. Despite extensive studies on heavy metal contamination using either geostatistical methods or standalone machine and deep learning models, existing literature lacks an integrated analytical synthesis that systematically evaluates optimization-driven ML–DL frameworks tailored to the non-linear, heterogeneous, and hydrodynamically active nature of coastal–deltaic ecosystems. This review addresses this gap by critically analyzing how swarm-intelligence-optimized learning architectures improve predictive reliability, interpretability, and scalability when fusing satellite, soil, and hydrological data. The study contributes new knowledge by establishing a unified conceptual framework that links algorithmic optimization behavior with delta-specific environmental processes, thereby advancing early-warning and decision-support.Advances in remote sensing, Machine Learning (ML), Deep Learning (DL) and Swarm Intelligence (SI) in recent years have synergetically contributed the development of intelligent data-driven techniques for predictive environmental modeling. Adding optimization techniques like PSO, GWO, and ACO in the frame work enriches Model calibration, Accuracy performance of model and reinterpreting phenomenon. This survey focuses on ML–DL–SI optimization-driven frameworks that integrate satellite imagery for heavy metal prediction. These unified, explainable systems promise improved early detection and sustainable management of contamination in fragile coastal–deltaic environments.
Keywords:	Coastal–Deltaic Ecosystems;Heavy Metal Contamination;Machine Learning (ML);Deep Learning (DL);Swarm Intelligence (SI);Optimization Algorithms.
DOI:	https://doi.org/10.5281/zenodo.18666451
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	QUANTIFYING THE UNCERTAINTY OF FINANCIAL DISTRESS SEARCH BEHAVIOR: A HYBRID LSTM AND MONTE CARLO DROPOUT APPROACH FOR THE PAWNING INDUSTRY
Author:	WANDA WANDOKO, ROBERTUS NUGROHO PERWIRO ATMOJO, IGNATIUS ENDA PANGGATI
Abstract:	This study aims to develop a decision support model to predict public interest trends in Indonesian pawning services while simultaneously quantifying uncertainty risks through a hybrid framework. Utilizing daily Google Trends time-series data as a real-time proxy for pawning demand over a five-year period (November 2020–October 2025), this research develops a hybrid model. It integrates Long Short-Term Memory (LSTM) to map complex non-linear patterns and long-term dependencies, with Monte Carlo simulation (via the MC Dropout technique) to generate probabilistic forecasting. Empirical results demonstrate that the deterministic LSTM component achieved "Highly Accurate" performance, measured by a Mean Absolute Percentage Error (MAPE) of 7.90% on the test set. The MAPE results prove the model's capability in separating fundamental trend signals from daily noise. Furthermore, the probabilistic Monte Carlo component successfully transformed single-point forecasts into measurable risk distributions. The model proved effective as an anomaly detector, identifying market surprises when actual values fell outside the 95% Confidence Interval. The novelty of this research lies in the hybrid integration of deep learning and stochastic Monte Carlo simulation applied to the underexplored pawning financial sector. This study contributes theoretically to time-series forecasting and pawning literature. Additionally, it offers managerial contributions, particularly for stakeholders in the pawning industry, by providing a tool for risk-based decision-making.
Keywords:	Time-Series Forecasting, Long Short-Term Memory (LSTM), Monte Carlo Simulation, Google Trends, Pawning Industry, Risk Management.
DOI:	https://doi.org/10.5281/zenodo.18666479
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	PHISHING DETECTION WITH SELF-ATTENTION AND NEURAL ODE LAYERS USING A TENSORFLOW-BASED DEEP LEARNING APPROACH
Author:	R ADINARAYANA , G VAMSI KRISHNA
Abstract:	Phishing remains the largest cyber security problem with which we have to put up with today and is one of the most common means under which users are deceived into providing their personal information (e.g., usernames, passwords, and credit card) because it can be cheaply mass-produced. The paper proposes an effective deep learning-enabled phishing detection model, which jointly applies self-attention modules and re-layered evaluated neural ordinary differential equation (ODE) blocks to achieve customizable traditional phishing detection systems. Multithread Self-Attention mechanism allows the model to attend which are the most relevant features within complex datasets, and capture elaborate relationships between categorical and numerical features. The use of Neural ODEs adds the possibility to model in continuous time, having the system getting information on dynamic and changing patterns of phishing attempts. The architecture also uses more elaborate preprocessing layers including normalization of numeric features and one-hot encoding pattern for categories, providing a powerful means to represent different sorts of data. Regularization techniques such as the weighted activation scaling (SELU) and Alpha Dropout add to the stability of the learning process by preventing over fitting. The model, which has been built in Tensor Flow, outperforms traditional models (e.g., KNN, logistic regression and SVM) with an accuracy of 96.1% and AUC = 0.99. These results demonstrate that the model generalizes well on unseen samples, which would be a scalable robust solution for phishing detection. This work contributes to advancements in cyber security by showing that self-attention and Neural ODEs can be usefully combined to meet the demands of an ever-evolving crime.
Keywords:	Phishing Detection, Self-attention, Neural ODE, Deep Learning, Tensor Flow, Cyber Security
DOI:	https://doi.org/10.5281/zenodo.18666499
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	PPBEFL: PRIVACY-PRESERVING BLOCKCHAIN ENABLED FEDERATED LEARNING FOR HEALTHCARE DATA SECURITY
Author:	TANISHA BHARDWAJ , SUMANGALI K
Abstract:	Blockchain (BC)-based platforms have emerged as a result of the technology's explosive expansion., each with distinct structures and consensus mechanisms. This has heightened the focus on blockchain interoperability, facilitating interactions between different platforms. In decentralized learning environments, maintaining security, transparency, and trust is particularly challenging, particularly in delicate sectors like financial services and healthcare, where data centralization is not a viable option. The Blockchain-Enabled Federated Learning (BEFL) approach, which combines blockchain technology with this work, federated learning (FL) is provided as a dependable and effective alternative. By using an aggregation technique to eliminate anomalous model parameters and integrating blockchain procedures and privacy strategies to synchronize client privacy protection, PPBEFL protects against poisoning attempts. The PPBEFL model trains the client’s datasets using local models of the Visual Geometry Group 19 (VGG19) for images and a Convolutional Neural Network (CNN) for the dataset. The updates are serialized to avoid concurrency issues and recorded immutably on the blockchain. Aggregated updates refine a global model iteratively in a transparent and verifiable way. Heart Disease (HD) and Breast Cancer (BC) Image Datasets (Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM)) are used for experiments, Brain Tumor Magnetic Resonance Imaging (MRI) Dataset, and Breast Cancer Wisconsin (Diagnostic) comparison with typical FL schemes like FedAvg, Federated Learning with Multi-Party Computation (FL-MPC), Federated Learning with Robust Aggregation in Edge Computing (FL-RAEC), and the privacy-preserving and efficient FL (PEFL) all show improved defense against different types of attacks based on metrics like throughput, precision, loss, and delay.
Keywords:	Blockchain, Consensus Mechanism, Differential Privacy (DP), Federated Learning (FL), privacy security, Neural Network, Data protection, Decentralized learning, Privacy-Preserving Blockchain Enabled Federated Learning (PPBEFL).
DOI:	https://doi.org/10.5281/zenodo.18666506
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

v

Title:	STRUCTURAL CAUSES OF THE DIGITAL DIVIDE AND ITS LONG-TERM CONSEQUENCES FOR ACADEMIC ACHIEVEMENT
Author:	MOHAMMAD BANI YOUNES, ISSA ALSMADI, OMER ABUSHQEER, NJOOD ALJARRAH, RAZAN OBEIDAT
Abstract:	The rapid digitalization of education has reshaped learning environments but also exacerbated longstanding social inequities, creating a multidimensional digital divide. This study examines the structural determinants of digital exclusion, differences in access, skills, and usage, and evaluates their influence on students’ academic performance. Using a mixed-methods design, we analyze micro-level data from the 2018 Program for International Student Assessment (PISA) alongside a thematic synthesis of recent case studies. Quantitative results show a strong positive association (p < 0.01) between a composite Digital Access Index and achievement in mathematics, reading, and science, independent of socioeconomic status. The qualitative findings reveal three recurring determinants: socioeconomic and infrastructural constraints, disparities in teacher digital preparedness, and variations in sociocultural and digital capital. Integrating both strands of evidence, the study demonstrates how digital exclusion contributes to persistent educational disadvantages and constrains long-term opportunities in higher education and the labor market. Our study that addresses the digital divide requires coordinated reforms targeting infrastructure, digital competency development, and equitable pedagogical design to prevent the further entrenchment of digital stratification.
Keywords:	C++, Code Reusability, Template Programming, Generic Programming, Function Templates, Class Templates, Standard Template Library (STL), Compile-Time Polymorphism, Template Metaprogramming.
DOI:	https://doi.org/10.5281/zenodo.18666522
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	UTILIZING GPT 3.5 FOR ARABIC INTENT CLASSIFICATION WITH PROMPTING
Author:	FATMA JAMAL HABIB HABIB , SALLY S.ISMAIL , ABEER M.MAHMOUD
Abstract:	Intent classification in natural language processing (NLP) is crucial for identifying user intents from their utterances. This task is particularly challenging for Arabic due to its complex morphology and high ambiguity. Recent advancements in NLP, including deep learning and transfer learning, have improved Arabic intent classification. However, a significant gap still exists compared to major languages like English. This paper proposes a new Arabic intent classification based on 1-shot and 3-shot techniques for investigating prompting strategies for Arabic in specific .The ’Temperature’ parameter is employed to modulate probability distributions and guide the model in classifying Arabic sentences into predefined categories. Results show that prompting, especially contrastive with dynamic reasoning, outperforms fine-tuning in both accuracy and resource eﬀiciency, high- lighting the effectiveness of prompting techniques in enhancing Arabic NLP applications. Contrastive with Dynamic reasoning reported accuracy of 0.98% for 1- shot and 100% for 3-shots while fine tuning recorded 0.87% for training of the used Arabic intent dataset.
Keywords:	Natural Language, Prompting, Intent classification, Few shots, Deep learning, Transfer learning, Softmax, Temperature Scaling, Morphology, Ambiguity
DOI:	https://doi.org/10.5281/zenodo.18666533
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	MODELING THE EFFECTS OF CHATBOT RESPONSIVESS, USEFULNESS, AND PERCEIVED RISK ON E-COMMERCE USER SATISFACTION AND BEHAVIORAL INTENTION
Author:	AUDREY ESTHER LITA , VIANY UTAMI TJHIN
Abstract:	As e-commerce grows quickly in Indonesia, more businesses are using AI chatbots to make their services more efficient and available. These chatbots are meant to provide quick, useful answers that can improve the customer experience and encourage people to keep using the service. Still, many users worry about how reliable chatbots are, whether their data is safe, and if the responses are accurate. These concerns can make people feel there is a high risk, which may lower their satisfaction and make them less likely to keep using the service. Although past studies have looked at these issues one at a time, there is research that examines them together. This study aims to analyze the influence of chatbot responsiveness, usability, and risk perception on user satisfaction and behavioral intentions, and to examine the mediating role of satisfaction in these relationships. This study used a quantitative approach with 403 active e-commerce users from Shopee, Tokopedia, Lazada, and Blibli. Data analysis was carried out using the Partial Least Squares Structural Equation Modeling (PLS-SEM) method with SmartPLS 4. To add depth to the findings, qualitative guerrilla interviews were also conducted to capture real user experiences. This result found that when chatbots are responsive and easy to use, customers are more satisfied and more likely to act positively. However, if customers feel there is risk, their satisfaction drops. Customer satisfaction partly explains how these factors affect what people do, but it does not account for everything. The study showed that when chatbots are responsive and helpful, users are more satisfied and more likely to use them again. However, if users feel there is a high risk, their satisfaction drops. These results help explain how people in emerging markets adopt AI chatbots and offer useful guidance for e-commerce companies looking to design more trustworthy, reliable chatbots.
Keywords:	E-Commerce, Chatbot, Responsiveness, Usefulness, Perceived Risk, User Satisfaction
DOI:	https://doi.org/10.5281/zenodo.18666543
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	THE BEHAVIOUR OF THE TWO LANES OF A QUASI-ONE-DIMENSIONAL SYSTEM CONNECTED BY AN ANISOTROPIC NODE: STUDY OF A MERGING NODE
Author:	AHLAM EL ATTARI, KAMAL JETTO2, ZINEB TAHIRI, ANASTASIA GENNAD’EVNA SHEVTSOVA, ABDELILAH BENYOUSSEF, ABDALLAH EL KENZ
Abstract:	In the light of the growing popularity of the so-called hybrid or multimethod modelling approaches, we attempted to solve a selected problem in the field of modelling transport processes using the AnyLogic general purpose-modelling tool. Using the example of the controlled selected intersection, the procedure of creating a model with a more detailed analysis of its selected parts was discussed. Later on, the same model is investigated in the light of statistical physics. Using a cellular automaton model, we studied the response of a quasi-one-dimensional system to an anisotropic merging node. The comparison of the two simulations showed non-usual behaviours through the latter. Accordingly, following the patterns of Vehicles’ extraction, we found that the exits of the two lanes are correlated, and a self-organization process governs the node. A cross-correlation test proved that the flows of both lanes are interdependent. Finally, we brought to evidence the importance of cellular automata models in understanding the process of traffic breakdowns in comparison to hybrid modelling programs.
Keywords:	Traffic Flow, Anisotropy, Anylogic; Merge, Correlation, Self-Organization.
DOI:	https://doi.org/10.5281/zenodo.18666550
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ENERGY-EFFICIENT AND QOS-OPTIMIZED IOT ROUTING USING PROPOSED QEER-MFO HYBRID ALGORITHM
Author:	S. BHARATHI, DR. D. MARUTHANAYAGAM
Abstract:	The Internet of Things' (IoT) rapid expansion has transformed connectivity in a number of industries, including smart cities, healthcare, transportation, and agriculture. However, because of their dynamic topology, device heterogeneity, and limited power resources, IoT networks continue to confront difficulties with energy management, routing efficiency, and Quality of Service (QoS). Conventional routing protocols sometimes don't adjust to the changing needs of IoT environments because they were mostly created for static or homogeneous networks. In order to overcome these constraints, this study suggests a brand-new hybrid routing architecture called QoS and Energy-Efficient Routing utilizing Moth-LSTM Optimization (QEER-MFO), which combines Long Short-Term Memory (LSTM) networks with the Moth-Flame Optimization (MFO) method. While the LSTM model forecasts network characteristics like traffic congestion and node energy depletion to enable proactive routing decisions, the MFO component makes sure that the best path is chosen through adaptive exploration and exploitation, reducing energy consumption and transmission latency. The Trust-Aware IIoT Routing Dataset is used to assess the suggested method, and the ns3-gym interface is used to simulate it in NS-3 combined with Python. According to experimental data, QEER-MFO outperforms current bio-inspired and conventional routing protocols, resulting in increased packet delivery ratios, decreased latency, and extended network lifetime. For scalable and intelligent IoT ecosystems, this hybrid architecture provides a self-learning, predictive, and energy-efficient routing approach.
Keywords:	Internet of Things (IoT), Quality of Service (QoS), Energy Efficiency, Bio-Inspired Optimization, Moth-Flame Optimization (MFO), Long Short-Term Memory (LSTM), Hybrid Routing, Machine Learning, NS-3 Simulation and Predictive Network Modeling.
DOI:	https://doi.org/10.5281/zenodo.18666562
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	RHYTHM-GUIDED CONVOLUTIONAL NEURAL NETWORK SEGMENTAL TRANSITION ENTROPY-MODULATED LESION DISCRIMINATION COTTON LEAF DISEASE IDENTIFICATION
Author:	K. KARTHIGA , Dr. B. RAJDEEPA
Abstract:	Cotton leaf disease presents significant diagnostic challenges due to inconsistent symptom expression, irregular lesion geometry, and background interference in open-field imagery. These issues reduce the reliability of conventional detection systems, which often struggle with symptom fragmentation, visual overlap, and class imbalance. To address these limitations, this research proposes a biologically inspired architecture titled CO-CNN (Caterpillar Optimization-Based Convolutional Neural Network). The model incorporates rhythm-guided adaptation, segmental coordination, and entropy-modulated parameter transitions to regulate convolutional behavior based on lesion-driven visual entropy. Unlike fixed-structure models, CO-CNN refines kernel scaling, dropout positioning, and gradient modulation in response to region-specific complexity, ensuring stability across asymmetric lesion patterns. The network performs segment-wise adjustment to retain lesion detail while suppressing irrelevant activations arising from field-induced noise. In evaluation, CO-CNN achieves 89.828% Overall Detection Efficiency and reduces Total Prediction Deviation to 10.172%, outperforming state-of-the-art models under identical test conditions. These results validate its ability to classify cotton leaf diseases with high sensitivity and precision, especially in real-world datasets where symptom boundaries are diffuse, visual distortions are common, and class distributions remain uneven.
Keywords:	CO-CNN, Cotton Leaf Disease, Entropy-Modulated Learning, Segmental Adaptation, Rhythm-Guided CNN, Lesion Discrimination
DOI:	https://doi.org/10.5281/zenodo.18666572
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	FEDERATED LEARNING WITH QUANTUM-INSPIRED BOLTZMANN WEIGHTING: ENABLING SECURE AND ROBUST HEART DISEASE PREDICTION
Author:	MANJIT SINGH, MONG-FONG HORNG, CHUN-CHIH LO, D.VETRITHANGAM, SIVA SHANKAR
Abstract:	Cardiovascular disease prediction using machine learning faces critical methodological challenges, including data leakage from improper preprocessing sequences, arbitrary feature subset selection without empirical validation, and privacy vulnerabilities in centralized model aggregation. This research proposes a federated quantum-enhanced learning framework that addresses these gaps through: (1) site-specific local preprocessing with global stratified train/test splitting, training-only scaling, and training-only SMOTE application to eliminate data leakage; (2) federated feature selection combining per-site RandomForest importance computation, server-side global aggregation, and multi-k evaluation (k ∈ {5, 7, 9, 11, 13}) with regularized optimization to determine empirically-validated optimal feature subsets (k* = 11); (3) quantum-inspired Boltzmann-weighted secure aggregation (weights ∝ exp(−β·Loss_i)) with convergence monitoring and CTGAN-based generative augmentation for robustness to heterogeneity; and (4) convergence speed metrics tracking accuracy history to identify earliest rounds achieving near-optimal performance, enabling computational efficiency gains. The proposed architecture employs a federated Random Forest classifier (100 trees) trained at each site on CTGAN-augmented data using an empirically optimized 11-feature subset, while maintaining patient data locally at each institution and enabling collaborative model training through encrypted parameter exchange using quantum-safe cryptography. Experimental evaluation on the UCI/Kaggle heart disease dataset demonstrates superior performance compared to a centralized Logistic Regression baseline(99.3% accuracy with AUC 0.9967 for the federated approach vs. 91.8% accuracy for centralized Logistic Regression training), enhanced privacy guarantees through lattice-based LWE-256 quantum-resistant encryption, improved robustness across heterogeneous sites (cross-site performance variance <0.3%), and 28% faster convergence (48 rounds vs. 67 rounds for standard FedAvg). This work advances the state-of-the-art in privacy-aware, distributed cardiovascular disease prediction and is suitable for real-world multi-institutional clinical deployment.
Keywords:	Federated Learning, Quantum Computing, Heart Disease Prediction, Privacy-Preserving Machine Learning, Feature Selection, Secure Aggregation, Distributed Healthcare
DOI:	https://doi.org/10.5281/zenodo.18666578
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	A COMPREHENSIVE ANALYSIS ON DEEP LEARNING TECHNIQUES FOR ELECTRO ENCEPHALOGRAM SIGNAL ANALYSIS AND APPLICATIONS
Author:	SYDA NAHIDA , SHAIK SAGAR IMAMBI
Abstract:	Electroencephalogram (EEG) analysis has emerged as a cornerstone for understanding brain dynamics and diagnosing neurological and cognitive disorders. While recent years have witnessed a surge in deep learning–based EEG studies reporting high classification accuracy, existing literature largely remains descriptive and model-centric, offering limited critical insight into why certain approaches succeed or fail under real-world EEG constraints such as non-stationarity, inter-subject variability, and low signal-to-noise ratios. This gap limits both scientific understanding and clinical translation.This survey addresses this limitation by presenting a critical and integrative analysis of deep learning techniques for EEG signal processing, moving beyond architectural taxonomies to examine their methodological assumptions, robustness, interpretability, and deployment feasibility. Unlike prior surveys, this work systematically evaluates how preprocessing strategies, feature learning paradigms, and emerging architectures—such as attention mechanisms, transformers, and graph neural networks—interact with the intrinsic properties of EEG signals.The study synthesizes recent advances to identify unresolved challenges in generalization, explainability, and data scarcity, while highlighting emerging trends such as self-supervised learning, multimodal fusion, and privacy-preserving federated frameworks. By framing deep learning–based EEG analysis through a critical lens, this survey provides actionable insights for researchers and clinicians, and contributes new conceptual understanding required for the development of reliable, interpretable, and clinically deployable EEG-based AI systems.
Keywords:	EEG Signals, Deep Learning, Brain Wave Analysis, Neural Activity Classification, Machine Learning Algorithms, Emotion Recognition, Neurological Disorder Detection
DOI:	https://doi.org/10.5281/zenodo.18666600
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	REAL-TIME ARRHYTHMIA DETECTION FROM WEARABLE ECG DEVICES USING LIGHTWEIGHT 1D CNN AND EDGE AI DEPLOYMENT
Author:	SHANMUGA SUNDARI M , T. LAKSHMI PRAVEENA , PUSHPA BHUPATIRAJU , ASHA VUYYURU , BUKKACHARLA KISHORE KUMAR , B. GOVARDHAN SIVA SAI
Abstract:	Cardiovascular diseases (CVDs) are still among the most prevalent causes of death globally, making it important to have quick, constant, and non-invasive cardiac monitoring. In this paper, a real-time system for arrhythmia detection using wearable ECG sensors combined with a low-power one-dimensional Convolutional Neural Network (1D-CNN) optimized for Edge AI deployment is introduced. The model is tailored to run efficiently on low-power wearables like smartwatches and ECG patches. A publicly available dataset, the MIT-BIH Arrhythmia Database, was used for training and validation. Preprocessing involved band-pass filtering (0.5–40 Hz), z-score normalization, and segmentation in 2-second windows. The 1D-CNN architecture with four convolutional blocks and a softmax classifier had an accuracy of 98.72 %, precision of 97.85 %, recall of 98.10 %, and F1-score of 0.981 on the test set. Model pruning and quantization-based edge optimization decreased memory consumption by 42 % and inference latency by 35 %, making it possible for real-time processing at 62 frames per second on the Raspberry Pi 4 platform. The system showed dependable skill in distinguishing different arrhythmias, including left bundle branch block (LBBB), atrial fibrillation (AF), and premature ventricular contraction (PVC). As a result, this pioneering method offers a technically low-power and readily scalable way to continuously monitor cardiac health, which can be used in both hospitals and patients homes.
Keywords:	Arrhythmia Detection, Wearable ECG, 1D Convolutional Neural Network, Edge AI, Real-Time Inference, MIT-BIH Dataset, Embedded Health Monitoring, Quantization, TinyML
DOI:	https://doi.org/10.5281/zenodo.18666631
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	THE ROLE OF INFORMATION TECHNOLOGY IN THE DEVELOPMENT OF CORPORATE COMMUNICATIONS AND PR
Author:	HASSAN ALI AL- ABABNEH, ZEYAD M. JAMHAWI, AHMED ALI OTOOM, AL REFAI MOHAMMED N, JAMEEL AHMAD KHADER, GANNA MYROSHNYCHENKO
Abstract:	The paper is aimed at the investigation of the information technology in corporate communications and public relations development in the digital transformation period. The importance of the study is determined by the need to develop system solutions that enable organizations to organize transparent, sustainable interaction and grow trust on the part of stakeholders. The aim of the paper is to create a model on how IT tools can be employed in corporate communication processes and PR management activity. The methodological framework of the study is modeling: conceptual model that incorporates major components (technological infrastructure, digital channels, analytical systems, feedback processes) and their relationship with each other was developed through analyzing existing theoretical concepts and large international companies’ applied practices for 2021-2024. The model provides the possibility of estimating the IT deployment influence on effectiveness of communications by characteristics such as reaction rate response, coverage, customization and content risk management reputation. The results of the study indicate that this developed model can be an instrumental tool to help in corporate communication planning and monitoring, as well in adapting strategies based on profiles of potential target publics. What is unique in the work, is the creation of a deep theoretical structure that contain technology and organization aspects with respect to which IT effects on corporate communication can be traced. There are practical implications of these findings for managing organizations and public relations practitioners working on digitally interacting with stakeholders that are looking to amplify the effectiveness of digital communication by utilizing communication activities.
Keywords:	Information Technology; Corporate Communications; Public Relations; Digital Transformation
DOI:	https://doi.org/10.5281/zenodo.18666653
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	A NOVEL IOT BASED PEST IDENTIFICATION AND CLASSIFICATION USING CNN AND LSTM
Author:	ARUNAPRIYA. R , Dr.S.P.VALLI
Abstract:	Citrus crops are often damaged by pest a infestation, which is a persistent problem for growers and often leads to significant declines in output and profitability. Traditional pest detection techniques rely on direct observation by field workers, which is labour-intensive, time consuming and sometimes unreliable. In order to detect and categories citrus pests in real time, this study proposes an internet of things (IoT) based automated bug detection system that makes use of convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) networks. Images from Iot-enabled field cameras are processed using a lightweight hybrid deep learning techniques to extract features and make fast conclusions on devices with limited processing power. Six important citrus pest species are represented by the 1,200 images in this experimental dataset which was collected from a citrus research station in Tamil Nadu, India. With an accuracy of 95.4%, the evaluation results demonstrate that our hybrid CNN-LSTM architecture performs better than both traditional CNN and conventional machine learning models. By facilitating automated field monitoring early insect detection and timely decision making the proposal method reduces crop losses and the need for pesticides.
Keywords:	Artificial Intelligence (AI), Internet of Things (IoT), Citrus, CNN, LSTM, Pests
DOI:	https://doi.org/10.5281/zenodo.18666661
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ENHANCING CLOUD RESOURCE UTILIZATION THROUGH INTELLIGENT TASK MIGRATION AND ADAPTIVE THRESHOLD LOAD BALANCING MECHANISM
Author:	CHARUL BHANAWAT , MANOJ KUMAR JAIN
Abstract:	Cloud computing has become the most popular way to provide services and tools for computers over the internet. The proposed method uses a dynamic threshold-based algorithm that constantly checks the server's load and moves jobs around to avoid hotspots. We use a hybrid load balancing method that takes the best parts of both static and dynamic algorithms and keeps transfer costs as low as possible. The Adaptive Threshold Load Balancing (ATLB) is used to transfer and balance jobs in cloud computing. ATLB can improve cloud resource management by streamlining migration decisions and altering criteria. The adaptive threshold adjusts load thresholds depending on real-time system data. The results of our experiments show that our method cuts average response time by 34.7% and makes better use of resources by 28.3% when compared to the most common round-robin and least-connection methods with 96.2% efficiency, the suggested system keeps the load balanced even when the amount of work changes. The comprehensive simulations used to test our approach's performance in different types of cloud settings confirm that it works.
Keywords:	Cloud Computing, Task Relocation, Load Balancing, Resource Usage, Time Optimization
DOI:	https://doi.org/10.5281/zenodo.18666670
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	REAL-TIME FOOD CLASSIFICATION USING VGG19 WITH GRAD-CAM VISUALIZATION FOR ENHANCED INTERPRETABILITY
Author:	D. LEELA DHARANI, Dr.YERUVA JAIPAL REDDY, Dr.VANKUDOTHU MALSORU, RAVIKIRANREDDY KANDADI, Dr..KUNAPULI SIVA SATYA MOHAN, Dr.GARLAPATI NARAYANA, Dr VENKATA RAMANA.N, JYOTHSNA PALAGANI, Dr. SIVA KUMAR PATHURI, Dr D. HARITHA
Abstract:	As nutritional disorders increase, because of unbalanced diets and numerous cuisines available on markets, physical and mental health of humans is more likely to be ruined. Appropriate food categorization is essential in tracking food consumption and sensitizing nutrition. The following paper proposes a food recognition model based on the VGG19 deep convolutional neural network with the incorporated Grad-CAM (Gradient-weighted Class Activation Mapping) visualization of the predictions made by the model to improve its interpretability. The model is tested against extensive dataset of 53 dishes in international cuisine on which both non-vegetarian and vegetarian food items are found. VGG19 is chosen because it performs well in feature detection and Grad-CAM gives the areas on each image that affect the decisions made by the model, making them more transparent, and trustworthy. This explains AI model enables users and researchers to visualize and check classification results. As measured by experimental analysis, the enhanced VGG19 classifier with Grad-CAM achieves an efficient accuracy of 96 percent, which is better compared to the traditional machine learning models including Naive Bayes and Decision Tree. This proposed approach is effective in real-time applications monitoring food intake and diet assessment.
Keywords:	Nutrition Disorder, Grad Cam, VGG-19, Naive Bayes , Decision Tree.
DOI:	https://doi.org/10.5281/zenodo.18738472
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	MICROCHAETUS RAPPI-INSPIRED SECURE AND REGENERATIVE ROUTING PROTOCOL (MR-SRRP) FOR ADAPTIVE AND ENERGY-EFFICIENT CLOUD COMMUNICATION NETWORKS
Author:	VATCHALA B , PREETHI G
Abstract:	Cloud communication networks require routing mechanisms that sustain security, adaptability, and energy efficiency under rapidly changing virtual workloads and dense topologies. Conventional routing approaches experience route instability, excessive energy usage, and delayed recovery in large-scale cloud environments. This study presents the Microchaetus Rappi-Inspired Secure and Regenerative Routing Protocol (MR-SRRP), a biologically driven framework that models regenerative behavior and energy-conserving movement patterns of Microchaetus rappi within an AODV-based routing structure. The goal of this work is to design a self-healing and secure routing protocol capable of restoring disrupted paths, balancing energy consumption, and maintaining low-latency communication. The underlying hypothesis assumes that embedding biological regeneration and adaptive intelligence into cloud routing decisions enhances route stability, packet reliability, energy utilization, and cryptographic diffusion strength. The protocol integrates encrypted tunneling, trust-driven authentication, regenerative key scheduling, and dynamic route rejuvenation across twelve coordinated phases. Simulation evaluation demonstrates improved delay, packet delivery, throughput, and energy conservation, with an avalanche diffusion strength of 86.14%, validating the protocol’s effectiveness for adaptive and secure cloud communication environments.
Keywords:	Routing, Cloud, Encryption, Security, Avalanche, Energy
DOI:	https://doi.org/10.5281/zenodo.18666685
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	INTELLIGENT OPTICAL CHARACTER RECOGNITION THROUGH CNN-LSTM FUSION WITH DICTIONARY VALIDATION AND ERROR CORRECTION
Author:	A.NARESH KUMAR , S.APARNA
Abstract:	OCR has revolutionized the process of text extraction and digitization, which is playing a key role in industries including document processing, healthcare, and finance. Although such models have been developed, conventional OCR systems are usually not capable of handling mixed, low quality, and noisy data. To overcome these limitations, the hybrid CNN-LSTM model combined with a dictionary-based confidence validation algorithm is proposed in this paper, where Convolutional Neural Networks (CNN) is used to extract spatial features effectively and Long Short-Term Memory (LSTM) networks to learn sequential data. Standard data preprocessing algorithms such as normalization, augmentation and Isolation Forest based outlier detection are applied to streamline the input data. A finite automata model represents the flow of data, and this gives a structured view of the model transitions. Also, a new confidence validation algorithm compares predictions to a medical dictionary, correcting low-confidence predictions, thus minimizing false predictions. An edit distance-based correction algorithm with first-character matching constraints further enhances accuracy by intelligently correcting OCR misrecognitions within a two-character tolerance, achieving only 1.4% false correction rate. The entire system of preprocessing has resulted in 6.3% increase in accuracy when compared to simple methods of normalization. This research methodology has significantly enhanced high text recognition accuracy and reliability to more efficient OCR systems with the capability to be tailored to meet arduous real-world environments in applications requiring high accuracy like domain-specific applications.
Keywords:	Optical Character Recognition, Hybrid CNN-LSTM Mode, Feature Extraction, Medical Text Recognition, Edit Distance Correction
DOI:	https://doi.org/10.5281/zenodo.18666690
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	HIERARCHICAL SOFTWARE-DEFINED NETWORKING ARCHITECTURE FOR NEXT-GENERATION MOBILITY MANAGEMENT
Author:	WONYONG YOON
Abstract:	Achieving Ultra-Reliable Low-Latency Communication (URLLC) in mobility scenarios for 5G or next-generation mobile networks is challenged by the dense deployment of small cells and the high frequency of handovers. Traditional 3GPP mobility protocols and existing centralized Software-Defined Networking (SDN) architectures suffer from signaling overhead and handover delay due to relying on the top-down orchestration of forwarding path changes. To overcome this fundamental limitation, we propose an architecture of hierarchical controllers that utilize a novel bottom-up orchestration strategy for mobility support. In this bottom-up orchestration, geographically located hierarchical controllers collocated with base stations can trigger handover execution, allowing target base stations to push user plane forwarding rules up the hierarchy. Through numerical analyses, we validate that our proposed hierarchical controller architecture and bottom-up orchestration reduces network bandwidth consumption for handover signaling by 55% and handover delay by 47% compared to the traditional tunneling-based mobility management method. The results support that the proposed architecture and orchestration mechanism can be a viable solution to provide efficient mobility support in 5G or next-generation dense cellular networks.
Keywords:	Software-Defined Networking, 5G Network, Mobility, Hierarchical Controller, Bottom-Up Orchestration.
DOI:	https://doi.org/10.5281/zenodo.18666705
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	OPTIMISED FINGERPRINT FEATURE REPRESENTATION FOR RELIABLE IDENTIFICATION
Author:	JAINY JACOB M. , D. SHANMUGAPRIYA
Abstract:	The fingerprint recognition has become a widely used procedure of biometric authentication but its functionality is typically affected by noise, mis-alignment and partial impression of the fingerprint making the existing feature extraction and classification scheme ineffective. In an attempt to address these issues, the research shall recommend an adaptive model of fingerprint recognition that shall incorporate the application of enhanced feature extraction and effective classification methods. Specifically, it introduces the proposed algorithm Selective Spatio-Temporal Residual Feature Framework (STRFF-Net) of discriminative spatio-temporal features extraction and Temporal Residual Flow Recognition Network (TRFRN) of the accurate fingerprint identification. STRFF-Net uses the residual flow modeling and attention to highlight salient ridge and minutiae patterns and discard the irrelevant or noisy regions to provide enriched feature representation. TRFRN uses the stream of attention-based classification model to investigate the spatial and temporal changes patterns of the fingerprints that enables reliable recognition of the prints of fingers in the event of smudged and partial fingerprints. Comparison of the experimental results with publicly available fingerprint data shows that experimental analysis with reference to feature improvement, strength of the activation as well as accuracy of classification is highly improved when compared to existing and machine-based learning techniques. The robustness and precision of the suggested framework makes it a suitable instrument in the realistic application of biometric authentication, forensic identification or application in areas of serious security concern.
Keywords:	Attention Mechanism, Biometric Authentication, Fingerprint Recognition, Residual Flow, STRFF-Net, TRFRN.
DOI:	https://doi.org/10.5281/zenodo.18666729
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	DECISION SUPPORT FRAMEWORK FOR EARLY PREDICTION OF DIABETIC PATIENT READMISSION USING HYBRID METAHEURISTICS
Author:	P. VENKATA KISHAN RAO , AARTI , A. SURESH RAO
Abstract:	Readmissions to hospitals are a significant health issue for people with diabetes and waste valuable hospital healthcare resources. The proposed model in this research, is a hybrid one combining feature selection using GRA and optimizing through PSO while tuning their hyperparameters by GWO. The most important predictors of risk of readmission from demographic, clinical and laboratory parameters were pre-processed and analyzed with the UCI Diabetic Readmission Database. GRA and PSO brought about good selection of informative features, and GWO tuned model parameters for base classifiers such as Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), Logistic Regression (LR) and k-Nearest Neighbours( KNN). The ensemble model reported here achieved the highest accuracy of 98.6+% compared with decision tree (DT) 97.3% and gradient boost (GB) of 97.6%. Other metrics – F1-score (98.5%), Kappa (0.987), AUC (0.999) further confirmed its robustness. Stratifying patients into high, moderate, low high risk enables earlier interventions and is a useful decision support system in diabetic readmission reduction initiatives and delivery of care.
Keywords:	Diabetic Readmission, Ensemble Learning, Grey Relational Analysis (GRA), Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO).
DOI:	https://doi.org/10.5281/zenodo.18666740
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	MACHINE LEARNING BASED GCN-LSTM MODEL FOR CROP YIELD PREDICTION USING SPATIAL-TEMPORAL FEATURE LEARNING
Author:	MAMTA KUMARI , SUMAN , DEVENDRA PRASAD
Abstract:	Prior research has identified limited data and minimal use of soil characteristics as significant shortcomings. To address these issues, this article conducted extensive data collection for bajra yield prediction and introduces a novel Graph convolution neural network and long short-term memory (GCN-LSTM) model, which consists of three main stages: data collection and processing; spatial feature learning; and prediction. The model breaks the limitations of the traditional methods by using data analytics based on IT and advanced deep learning, making more accurate predictions that can be used in smart agriculture, resource optimization, and food security. Unlike previously studied, deep learning models such as recurrent neural networks (RNN), Long Short-Term Memory (LSTM), and Convolutional neural networks (CNN), the proposed GCN-LSTM model does not assume independence among the districts in the crop yield prediction (CYP) data. Instead, it processes attributes related to crop yield prediction such as meteorological, soil, and climate data. The spatial feature learning is not neglected and is leveraged by the LSTM model for the temporal prediction of crop yield. The performance of the GCN-LSTM model is evaluated on RMSE, R², and correlational coefficients. The experimental results demonstrate that the proposed model significantly outperforms conventional models by effectively incorporating spatial information.
Keywords:	Graph Convolution Network, Bajra Crop, Soil Data, Spatial Feature Learning, And Soil Characteristics.
DOI:	https://doi.org/10.5281/zenodo.18666754
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	STATISTICAL FORECASTING BASED ON CONDITIONAL DENSITY ESTIMATION
Author:	RAVINDRA CHANGALA, K.KIRAN KUMAR, S. RENU DEEPTI, V. PREETHI, G. SUSHMA KIRAN KUMAR KAVETI, NATHA DEEPTHI
Abstract:	Many real problems such as stock market prediction, weather forecasting etc has inherent randomness associated with them. Adopting a probabilistic framework for prediction can accommodate this uncertain relationship between past and future. Typically the interest is in the conditional probability density of the random variable involved. One approach for prediction is with time series and auto regression models. In this work, liner prediction method and approach for calculation of prediction coefficient are given and probability of error for different estimators is calculated. The existing techniques all require in some respect estimating a parameter of some assumed solution. So, an alternative approach is proposed. The alternative approach is to estimate the conditional density of the random variable involved. The approach proposed in this thesis involves estimating the (discretized) conditional density using a Markovian formulation when two random variables are statistically dependent, knowing the value of one of them lets us get a better estimate of the value of the other one. The conditional density is estimated as the ratio of the two dimensional joint density to the one-dimensional density of random variable whenever the later is positive. Markov models are used in the problems of making a sequence of decisions and problem that have an inherent temporality that is consisting of a process that unfolds in time in time. In the continuous time Markov chain models the time intervals between two consecutive transitions may also be a continuous random variable. The Markovian approach is particularly simple and fast for almost all classes of classes of problems requiring the estimation of conditional densities.
Keywords:	Statistical Prediction, Unbiased Ness, Sufficiency, Smoothing, Univariate Time Series, Autoregressive, Markov Chains.
DOI:	https://doi.org/10.5281/zenodo.18666768
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	VESSEL LOOP STRUCTURE-AWARE MULTI-GRADE DIABETIC RETINOPATHY CLASSIFICATION USING ENHANCED TCLLKEN
Author:	JAGADEESH VUTLA , VISALAKSHI ANNEPU
Abstract:	In order to prevent individuals from blindness and blurred vision, timely and accurate Diabetic Retinopathy (DR) diagnosis is essential in the context of the medical field. Therefore, numerous studies have been implemented by a variety of researchers to classify DR using Artificial Intelligence (AI) algorithms and Retinal Fundus Images (RFIs). The fundamental biomarker for recognizing the DR is the vessel loop. However, the conventional works failed to concentrate on investigating the Vessel Loop Structure (VLS) during DR classification, thereby causing poor DR grading. Thus, to upgrade the diagnosis reliability, a significant VLS-aware multi-grade DR classification framework is pivotal. Therefore, This work is driven by the need to develop a more accurate, VLS-aware multi-grade DR classification method using Transfer Complementary Log-Log Karplus-EfficientNet (TCLLKEN). Primarily, the eye-retinal images are gathered and further pre-processed, followed by gray-scale conversion. After that, regions in the gray-scale images are grouped. In the meantime, from the pre-processed images, the Green Channel (GC) is extracted. Subsequently, vessel structure segmentation, vessel graph construction, and VLS extraction are carried out. Lastly, the extracted features, extracted VLS, and pre-processed images are inputted into the proposed TCLLKEN that classifies the multi-grades of the DR efficiently. Overall, through the inclusion of VLS analysis, DR grading accuracy is significantly enhanced across all stages. Hence, highly accurate multi-grade DR classification is achieved by the proposed VLS-aware TCLLKEN model with an accuracy of 98.99%, thus performing better than existing methods.
Keywords:	Diabetic Retinopathy (DR), Blood Vessels (BV), Vessel Loop Structure (VLS), Retinal Lesions (RL), Deep Learning (DL), Proliferative Diabetic Retinopathy (PDR), and Retinal Fundus Images (RFI).
DOI:	https://doi.org/10.5281/zenodo.18666781
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	LIGHTWEIGHT CNN-TRANSFORMER FUSION MODEL FOR AUTOMATED RIB FRACTURE LOCALIZATION IN RADIOGRAPHS
Author:	NAREGALKAR AKSHAYKUMAR RANGNATH , ANE ASHOK BABU , NAGAMALLESWARA RAO PURIMETLA , SHAIK SALMA BEGUM , DEEPAK V , S SINDHURA , N JAYA
Abstract:	This study proposes a Lightweight CNN–Transformer Fusion Model for the automated detection of rib fractures in chest radiographs. Rib fractures represent one of the most common trauma-related injuries; however, their identification through conventional radiography remains challenging due to overlapping anatomical structures and the subtle nature of fracture lines. The proposed hybrid architecture integrates Convolutional Neural Networks (CNNs) for capturing fine-grained local features with Transformer encoders to model long-range global dependencies, thereby enhancing interpretability and diagnostic accuracy. A dataset comprising 6,200 chest radiographs was utilized to evaluate the model’s performance against standard benchmarks. Experimental results demonstrated superior performance, achieving a precision of 94.6%, sensitivity of 92.3%, specificity of 95.7%, and an F1-score of 91.9%, outperforming CNN-only, Transformer-only, and CT-based baseline models. Furthermore, the lightweight architecture enabled real-time inference at 32 frames per second, ensuring suitability for deployment in resource-constrained healthcare environments. Overall, the findings highlight the potential of hybrid deep learning frameworks to significantly enhance clinical decision-making by providing accurate, efficient, and interpretable rib fracture detection in chest radiography.
Keywords:	Rib Fracture Detection, CNN–Transformer Fusion, Medical Image Analysis, Chest Radiographs, Deep Learning
DOI:	https://doi.org/10.5281/zenodo.18666797
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	MULTI MODAL AI FOR IMPROVING ACCESSIBILITY THROUGH SPEECH AND GESTURE RECOGNITION FOR THE DISABLED
Author:	RAVINDRA CHANGALA , Dr. K PAVAN KUMAR, Dr. S ARUNA, P MADHAVI, N SRINIVASA RAO, CH LAVANYA SUSANNA, Dr. C SUGANYA
Abstract:	In this paper, we report on designing and evaluating a new multimodal AI that multi-modally combines speech and gesture recognition to enhance accessibility for people with disabilities. We aim to create a flexible and resilient interface to allow individuals with various disabilities to interact with technology using one or two input modalities. The model comprises the Deep Learning DNN models for speech (LSTM-based) and gestures (CNN-based) and the multi-attention fusion model to combine both modality outputs. To that end, a proprietary speech dataset of people with speech disabilities and gesture videos of people with motor disabilities was used to train and test the system. The experimental results demonstrate that our multimodal model outperforms a speech-only and a gesture-only baseline system in terms of global F1-score (0.90) and reduces the task time duration by 28% compared to the baseline systems (Google Assistant and Microsoft Kinect). The acoustics-based system was also robust to noise, performing with 82% accuracy at high noise. The system accessibility, ease of use, and performance were very high. This paper shows how multimodal AI can contribute towards building inclusive, user-friendly technologies for people with disabilities. It is a good step for achieving greater real-life usability.
Keywords:	Multi-modal AI, Speech Recognition, Gesture Recognition, Accessibility, Deep Learning, Assistive Technology
DOI:	https://doi.org/10.5281/zenodo.18666805
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ENHANCING CONTENT RETRIEVAL WITH BIG DATA AND NATURAL LANGUAGE PROCESSING FOR SCALABLE AND SEMANTIC SEARCH SYSTEMS
Author:	S SUJANTHI , Dr. MULUMUDI SUNEETHA , NARASIMHA RAO THOTA , N SRIHARI RAO , CHITNEEDI KASI VISWANADHAM , Dr. B HEMANTHA KUMAR , P S V S SRIDHAR
Abstract:	Content search and retrieval systems are required to be more efficient due to the data's high volume and complexity. This paper presents a new way to combine Big Data techniques with high-end Natural Language Processing (NLP) models to improve the search procedure's accuracy, relevance, and scalability. We aim to build a system that effectively uses distributed Big Data infrastructure for data processing and cutting-edge NLP models for semantic query interpretation. We evaluate the system over three datasets: Common Crawl (web content), Medical Text Mining, and Amazon Product Reviews, and compare to traditional keyword-based search and TF‐IDF and Word2Vec‐based approaches. The experimental results show that our system achieves better precision, recall, F1-score, and Mean Average Precision (MAP) than previous works at a reasonable query response time. The combination of Big Data and NLP results was much more relevant and contextually aware. This work is a big step toward better content search in many application domains; it makes more accurate and efficient retrieval possible and proposes a personal search experience. The proposed integration of Big Data infrastructure with advanced NLP models enables scalable and semantically rich retrieval, addressing key limitations of existing keyword-centric and shallow semantic search systems.
Keywords:	Big Data, Natural Language Processing, Content Search, Semantic Search, Precision, Information Retrieval
DOI:	https://doi.org/10.5281/zenodo.18666811
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ENHANCING AUTOMATED CODE GENERATION WITH TRANSFORMER MODELS AND REINFORCEMENT LEARNING: A DEEP LEARNING APPROACH TO SOFTWARE DEVELOPMENT
Author:	T PRAVEEN KUMAR , ANNAPURNA GUMMADI , SYED NAFEES AHAMED , N SRIHARI RAO , CHITNEEDI KASI VISWANADHAM , DEVAKI K , P S V S SRIDHAR
Abstract:	Automated programming using deep learning can shorten the code development process and ensure the software built is of high quality. In this paper, we investigate the combination of transformer models and reinforcement learning (RL) for the automatic generation of code. The aim is to design a model that produces correct, consistent, and valuable code in different programming languages. Our test utilized 5 million pieces of code from several open-source repositories, and models were evaluated based on whether their output was grammatically correct, code quality, code execution accuracy, and code production speed. Our model outperforms conventional LSTM-based approaches and GPT-2, achieving excellent syntactic correctness and execution accuracy, the highest code quality marks (8.5/10), and completing tasks in less time. The findings demonstrate that combining deep learning and RL enables the creation of top-quality code efficiently. By applying AI to software development, this work finds that both speed and reliability noticeably improve, which is beneficial for all parties involved and the broader industry. Despite the success of deep learning in natural language processing, automated code generation continues to face challenges related to execution correctness, code quality, and scalability across programming languages. Experimental results demonstrated that the proposed transformer–reinforcement learning framework achieved higher syntactic correctness, execution accuracy, and reduced generation time compared to existing LSTM-based and transformer-only models, indicating its suitability for real-world software development tasks.
Keywords:	Automated Code Generation, Deep Learning, Reinforcement Learning, Transformer Models, Code Quality, Software Development
DOI:	https://doi.org/10.5281/zenodo.18666822
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ENHANCING AUTONOMOUS DRONE DELIVERY SYSTEMS: A HYBRID CNN-LSTM APPROACH FOR ROBUST OBJECT TRACKING IN DYNAMIC ENVIRONMENTS
Author:	ANNAPURNA GUMMADI , RAVINDRA CHANGALA , PULLAIAH PINNIKA, B SRIVANI, VIJAYAKUMARI RODDA, Dr. N NEELIMA, R KIRTHIGA
Abstract:	Strong object tracking is needed in autonomous drone delivery to manage changing surroundings appropriately. This research introduces an innovative method that uses CNNs to extract features and LSTM networks to predict the sequence in which objects appear in videos. The primary goal is to enhance the tracking system's reliability while objects move and are hidden and when the environment is unpredictable. We measured the precision, success rate, real-time FPS, and occlusion handling of the method with the help of the MOT Challenge and our custom drone flight dataset. Experiments found that the Proposed Model is more accurate than SiamFC, DeepSORT, and CNN + Kalman Filter, with a precision of 94% and a success rate of 92%, and can operate at 30 FPS in live tests. The model is trusted for uncrewed drone operations because it can overcome occlusions and accurately restore missing parts. This approach improves how autonomous drones navigate, making them useful for logistics, keeping watch, and handling emergencies. The primary contribution of this work lies in the development of a hybrid CNN–LSTM architecture integrated with sensor fusion for robust real-time object tracking. The proposed approach advances existing solutions by improving tracking accuracy, occlusion recovery, and real-time performance in dynamic drone delivery environments.
Keywords:	Autonomous Drones, Object Tracking, Deep Learning, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM), Sensor Fusion
DOI:	https://doi.org/10.5281/zenodo.18666829
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ANALYSIS ACCURACY OF CLASSIFICATION FOR MULTI-EVIDENCE DATASET CONTAINING DISCREATE AND CONTINUOUS VALUE USING CATEGORICALNB AND GAUSSIANNB
Author:	DR. VIJAY KUMAR VERMA
Abstract:	Machine Learning is one of the emerging fields in computer science, and Bayes’ theorem can be applied to predict class labels precisely and accurately. Naïve Bayes classification, specifically the Categorical Naïve Bayes (CategoricalNB or CNB) classifier, is a simple probabilistic model based on Bayes’ theorem and is highly effective in reducing computational cost, as conditional probabilities can be easily estimated from data. In this paper, we employ both Categorical Naïve Bayes (CNB) and Gaussian Naïve Bayes (GNB) classifiers. The CNB classifier is widely used for binary and multi-class classification problems in machine learning and performs well when the attribute values are categorical or discrete. However, it has certain limitations when handling continuous data. In real-world applications, datasets often contain attributes in mixed formats, including both discrete and continuous values. For continuous attributes, CNB requires discretization, whereas GNB can be directly applied to continuous data by assuming a Gaussian distribution. In this study, we use CNB and GNB to classify a dataset containing both discrete and continuous attributes and compare their classification accuracy. A real-life dataset obtained from the HDFC Bank home loan department was used for experimentation. The dataset consists of 6,000 records with 19 attributes. Experimental results show that the performance of the Gaussian Naïve Bayes classifier is superior to that of the Categorical Naïve Bayes classifier. However, a limitation of GNB is its assumption that continuous attributes follow a normal distribution.
Keywords:	Classification, Machine learning, Naive Bays, Continues, Discreate, Gaussian Naïve, Accuracy
DOI:	https://doi.org/10.5281/zenodo.18666839
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	ARTIFICIAL INTELLIGENCE-POWERED LANDSCAPE VISION TRANSFORMATION TECHNIQUE FOR EARLY FLOOD RISK DETECTION THROUGH SATELLITE MAPPING SYSTEM USING DEEP FEATURE ENGINEERING
Author:	Dr.B.SHARMILA , Dr.A.V.SANTHOSH BABU
Abstract:	Environmental development depends entirely on natural resources, such as covering landscapes based on human lives and technological improvement. Specifically, the sloping mountain regions of land and their nature depend on sliding partitions; due to large floods, human lives are affected mainly by sudden causes of disasters. This paper addresses the critical issue of early flood risk detection in sloping mountain regions, where human lives and technological infrastructure are heavily reliant on natural resources. Satellite images and AI techniques are crucial developments to make analysis effective. Still, traditional methods fail to analyze the feature mapping and presence in feature variability due to increasing time frames, resulting in a high recall, which means that fables are unable to detect the flood quantity from regions where rainfall occurred, regions, and landscape regions. To resolve this problem, we propose an AI-powered landscape Vision Transformation Technique (LVTT) For Early Flood Risk Detection through a Satellite Mapping System Using Deep Feature Engineering. Initially, the landscape region images are collected from the satellite monitoring system. An adaptive Gaussian filter carries out the preprocessing. The historic color difference-based Slice Window Canny Edge Detection (SWCES) is used to segment the two-layer flooding and water flow region with dependable features. Then, Region Scaled Ant Colony Optimization (RSACO) for selecting the variability region features mutually depends on detecting the flooding variation region. Finally, the Hyper Capsuled LSTM Gated Convolution Neural Network (LSTMG-CNN) predicts that the flood region will cover the risk region of landscapes for early detection of floods to safeguard human lives. The proposed system produces high performance by predicting in the early stage of flooding variation limits depending on the flood flow through the image, progressing to achieve a high precision rate and recall rate with improved detection accuracy than any conventional methods.
Keywords:	Risk Detection, Satellite Imagery, Deep Learning, Feature Engineering, Landscape Analysis, Artificial Intelligence, Convolutional Neural Networks, LSTM.
DOI:	https://doi.org/10.5281/zenodo.18666925
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	BLOCKCHAIN BASED TRUST AWARE FRAMEWORK FOR SECURE AND TRANSPARENT CONVENTIONAL ACADEMIC RESULT PUBLISHING SYSTEM
Author:	SHEELA D V , DR.ASHOK KUMAR T A
Abstract:	In the evolving landscape of educational assessment and result publishing, blockchain technology holds a transformative approach for managing the examination results. The traditional approach for publishing results possesses various challenges including data security vulnerabilities, delayed result distribution, lack of transparency, administrative inefficiencies and so on. Despit recent efforts using online examination platforms and blockchain based grade storage, existing solutions typically do not integrate decentralized tamper proof storage, automated access control and dynamic trust modeling within a single architecture, leaving result data exposed to manipulation, manual entry errors, weak auditability and limited interoperability. This research presents a novel framework for secure and efficient result publishing BTRP (Blockchain based Trust aware Result Publishing). Initially, the proposed model integrates blockchain technology with an efficient distributed database, utilizing Smart Contracts (SC) to ensure secure data storage and sharing. Secondly, we incorporate a trust-aware feedback mechanism into result publishing, allowing participants to receive feedback based on their behaviour. This feedback is dynamically updated on the blockchain through smart contracts, ensuring that only honest and authorized participants have opportunities for data sharing. Additionally, advanced cryptographic mechanisms are incorporated for feedback submission, which protects the privacy of participant information. Front end architecture of BTRP comprises Student, Examiner and Admin as the participants for result publishing; also authorized login dashboard is design for Examiner and admin participants for updation of results. Comparative analysis highlight BTRP’s superiority by addressing known limitations of existing e-learning and blockchain academic systems, such as manual data entry vulnerabilities, privacy concern and access control.
Keywords:	Result Data, Blockchain, Consensus, Smart Contract, Access Control
DOI:	https://doi.org/10.5281/zenodo.18666937
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	AL-BIRUNI EARTH RADIUS SPOTTED HYENA OPTIMIZER BASED TRANSFER LEARNING FOR BRAIN TUMOR CLASSIFICATION USING MRI
Author:	SRILAKSHMI ALURI , S. SAGAR IMAMBI
Abstract:	Brain tumor classification involves identifying and categorizing various types of tumors from Magnetic Resonance Imaging (MRI), a task critical for effective diagnosis and treatment planning. Existing methods often face challenges such as MRI noise, imprecise tumor localization, limited feature extraction, and high computational requirements, which can reduce classification accuracy and reliability. To address these challenges, an innovative technique named Al-Biruni Earth Radius Spotted Hyena Optimizer based Convolution Neural Network with Transfer Learning (BERSHO_CNN with TL) is proposed for classifying brain tumor. Initially, the pre-processing is carried out on MRI images using Non-Local Means (NLM) filtering. Tumor segmentation is then performed using the Multi-Attention Dense Network (MAD-Net). By utilizing the segmented images, feature extraction is conducted through a combination of 3D-Convolutional Autoencoder+Vision Transformer (3D-CAE+ViT). Finally, the brain tumors are classified using a proposed CNN with TL) model, which is trained using the BERSHO algorithm. BERSHO is developed by combining the Al-Biruni Earth Radius (BER) and the Spotted Hyena Optimizer (SHO). The proposed approach demonstrates enhanced classification accuracy, robust tumor localization, and improved feature representation compared to existing methods. The devised model has achieved a Negative Predictive Value (NPV) of 91.788%, accuracy of 92.847%, True Positive Rate (TPR) of 92.534%, Positive Predictive Value (PPV) of 92.277%, and True Negative Rate (TNR) of 91.747%. These findings indicate that the proposed framework provides a practical and efficient tool for automated brain tumor diagnosis, contributing new knowledge in integrating hybrid optimization with advanced feature extraction for improved MRI-based tumor classification.
Keywords:	Brain tumor classification, Magnetic Resonance Imaging, Al-Biruni Earth Radius, Spotted Hyena Optimizer, GoogLeNet.
DOI:	https://doi.org/10.5281/zenodo.18666947
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	TST-YOLO: TOKEN-SELECTIVE TWIN-ENHANCED YOLO FOR ROBUST UNDERWATER FISH DETECTION AND PHENOTYPE ESTIMATION ON AUVS
Author:	SIRIGINEEDI MANIKANTA , Dr. RAGHVENDRA KUMAR , Dr. R N V JAGAN MOHAN
Abstract:	Underwater fish detection remains difficult due to turbidity, color cast, motion blur, and the prevalence of small, fast, look-alike targets on embedded AUV hardware. We present TST-YOLO, a compact detector that combines four synergistic components: (i) a physics-aware digital-twin synthesizer that exposes the model to realistic water-optics shifts during training; (ii) a lightweight pre-enhancement fusion stage that mitigates color/contrast bias before inference; (iii) a token-selective transformer head that prunes and merges low-information tokens for AUV-grade efficiency; and (iv) a phenotype-guided auxiliary head that estimates lateral-line scale counts to regularize features toward biologically meaningful structure. Evaluated under identical training budgets on DeepFish and DePondFi’23, TST-YOLO improves mAP@50–95 by +6.1 and +6.0 points, respectively, over strong YOLOv8/YOLOv7 baselines, with +5.5 APS_SS gains on small targets. Confidence calibration also improves (ECE ↓34%, Brier ↓), while the token-selective head reduces transformer tokens by ≈30% at equal or better accuracy, cutting end-to-end Jetson latency by ~6%. Results are reported as 5× runs (mean±SD) with bootstrap confidence intervals, paired tests, and McNemar analyses. Beyond accuracy gains, the contribution is positioned at the systems level, demonstrating how domain-aware data synthesis, reliability-oriented evaluation, and token-efficient transformer design can be jointly integrated for trustworthy, resource-constrained intelligent perception. This emphasis on efficiency, robustness, and statistical rigor highlights the relevance of TST-YOLO as an information-technology solution for dependable autonomous sensing rather than a task-specific detector alone.
Keywords:	Underwater Object Detection; Autonomous Underwater Vehicles (Auvs); Digital-Twin Augmentation; Token-Selective Transformer; Underwater Image Enhancement; Small-Object Detection; Phenotype (Lateral-Line) Estimation.
DOI:	https://doi.org/10.5281/zenodo.18666965
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text

Title:	STYLISTICALLY-AWARE HINDI-ENGLISH POETIC TRANSLATION WITH MBART AND LLM-BASED POST-EDITING
Author:	PRAGYA TEWARI , ANURAG SINGH BAGHEL
Abstract:	Translating poetry across languages poses significant challenges, particularly for low-resource pairs like Hindi-English, where linguistic, cultural, and stylistic gaps are substantial. This paper introduces a domain-adapted machine translation framework designed to preserve the poetic essence like metaphor, emotion, tone, and rhythm of Hindi poetry in English translation. We fine-tune mBART50, a multilingual sequence-to-sequence model, on a curated parallel corpus of Hindi-English poetic pairs, augmented with stylistic tags. We also explore optional post-editing using large language models (LLMs) to enhance fluency and poetic expressiveness. Our approach outperforms standard translation systems, as demonstrated by both automatic metrics and human evaluations. To the best of our knowledge, this is the first systematic study of Hindi–English poetry translation that combines (i) the construction of a novel annotated parallel corpus, (ii) style-aware fine-tuning of mBART50, and (iii) LLM-based post-editing for poetic refinement. This work represents a step toward building translation systems that capture not just meaning, but the creative and emotional depth of poetry especially in low-resource language settings.
Keywords:	Neural Machine Translation, Hindi Poems, Large Language Models, Fine Tuning, Low Resource Language
DOI:	https://doi.org/10.5281/zenodo.18667066
Source:	Journal of Theoretical and Applied Information Technology 15^th February 2026 -- Vol. 104. No. 3-- 2026
Full Text