		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
July 2026 | Vol. 104 No.13

Title:	ENHANCING NOMA PERFORMANCE THROUGH TRANSFORMER-ASSISTED SIGNAL PROCESSING AND REINFORCEMENT LEARNING-BASED POWER ALLOCATION
Author:	MANCHURI TULASI, D.SREENIVASA RAO
Abstract:	This paper introduces NOMA-OptiNet, a novel Transformer-based deep learning framework integrated with Proximal Policy Optimization (PPO) reinforcement learning for enhanced signal detection, interference mitigation, and dynamic resource allocation in Non-Orthogonal Multiple Access (NOMA) systems. The proposed system leverages an Encoder-Decoder Transformer architecture, combined with AdamW optimization, to extract hierarchical features and improve convergence in complex multi-user environments. Through extensive simulations, NOMA-OptiNet achieves a Symbol Error Rate (SER) of 0.012 (User 1) and 0.015 (User 2) at 25 dB, outperforming SPICE (0.035), SIC (0.045), and CNN (0.025). The model also records an Average Error Rate (AER) of 0.009 and a Normalized Mean Square Error (NNMSE) of 0.005, demonstrating superior accuracy in user activity detection and channel estimation. Comparative analysis reveals that NOMA-OptiNet converges within 20 iterations, whereas traditional methods require 35-50 iterations. The CDF plot confirms that NOMA-OptiNet maintains higher probability density for lower error rates, ensuring robustness across SNR conditions ranging from 0 dB to 25 dB. Real-time adaptation tests underscore the effectiveness of the model to perform dynamic optimization of power distribution compared to the use of a fixed scheduling. These results establish NOMA-OptiNet as a scalable, computationally efficient solution that enhances next-generation wireless communication systems, paving the way for ultra-reliable low-latency NOMA networks.
Keywords:	Noma, Transformer, Symbol Error Rate, Deep Learning, Convergance.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	STAGED RECURRENT NEURAL NETWORK AND LONG SHORT TERM MEMORY BASED DEEP LEARNING ARCHITECTURE TOWARDS CYBERBULLYING DETECTION AND SEVERITY CLASSIFICATION IN ONLINE SOCIAL MEDIA PLATFORMS
Author:	Dr. S. UMA, Dr. C. THILAGAVATHY, S. GOKILAVANI, M. VINU, S. PRIYADHARASHINI, Dr.ARCHANA NANDIBEWOOR, M. SHANTHINI, G. RAJASEKAR
Abstract:	Online social media platforms gaining accelerated popularity in the last decade towards experiencing their views and opinion about personnel, product and social information. Cyberbullying is becoming severe concern to the online social media as it affects multiple categories of peoples and its societies. In literatures, many automated methods have been employed by researchers to detect the cyberbullying. Those architectures examine the social and cultural phenomena to identify the behavioral pattern of the post. Despite of the several advantages, those model fails considers contextual factors and computational complexities which leads to cold start issues, uncertainty, class imbalance problem, changes in velocity and volume. To mitigate those challenges, a new architecture has to be established to alleviate it by detecting and classifying the severity of the cyberbullying post. Stacked Recurrent Neural Network with Long Short Term Memory is designed and developed to process temporal and time varying data on computing the long term dependencies among the various users. Particular model was trained to extract the contextual, temporal, spatial and sentimental features of data. Those extracted features are represented as feature vector and it is processed further to construct the feature mapping using long short term memory layer of the model. Subsequently, the generated feature map is processed with classification function on the connected layer to detect and classify the severity of the post as mild, moderate and strong. Experimental analysis of the model is performed using twitter dataset in the python environment and it exhibits better processing capabilities and computational capabilities on reducing cold start and class imbalance issues.. Further performance analysis of the model is performed to compute prediction and classification accuracy which produces improved classification accuracy of 98.9 percent and predictive accuracy of 98.7 against different user characteristics. Finally it proved to be outperforming compared to conventional approaches
Keywords:	Cyberbullying, Recurrent Neural Network, Long Short Term Memory, Online Social Media, Twitter dataset
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A CROSS-DISCIPLINARY APPROACH TO OPPORTUNISTIC RESOURCE ALLOCATION IN MULTICELLULAR SYSTEMS: INSIGHTS FROM GIS AND BIOINFORMATICS
Author:	MOHAMED ATIFI, JAMILA DAHMANI, LAHCEN ZIDANE, AZIZ BOUJEDDAINE, SARA RIAHI, HAMID KHALIFI
Abstract:	GIS and Bioinformatics concepts are increasingly relevant in the analysis and modeling of complex systems. The rapid evolution of wireless and mobile systems has made the optimal allocation of radio resources an increasingly critical challenge. The growth of telecommunications networks relies on the efficient deployment of wireless technologies such as Wireless Fidelity (Wi-Fi) and mobile systems such as Long-Term Evolution (LTE). In this article, we propose an algorithm that improves overall radio resource allocation within a heterogeneous wireless and mobile network environment, using dynamic programming and the Bellman principle of optimality. User mobility is explicitly taken into account to enhance performance, and the proposed approach is validated through numerical testing. This optimization framework, grounded in dynamic programming, also echoes methodologies shared with bioinformatics (e.g., sequence alignment algorithms) and relies on a spatial zoning model that can be represented and analyzed through Geographic Information Systems (GIS).
Keywords:	Resource allocation; Optimization; GIS; Cellular systems; QoS; Bioinformatics
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	DEEP LEARNING-BASED MULTICLASS CLASSIFICATION OF MEDICINAL PLANT LEAF DISEASES USING EFFICIENTNET-B3, CONVNEXT-TINY, SWIN TRANSFORMER AND A HYBRID ENSEMBLE ON THE CIMD DATASET
Author:	K. NAGA DIVYA, AVANI ALLA , PAPPULA SARALA, S NAGA SINDHU, SURESH KUMAR GUDISE, BALAMURALIKRISHNA THATI
Abstract:	Plant diseases threaten the pharmacological quality of medicinal herbs cultivated across Central India and remain difficult to diagnose accurately without laboratory analysis. The Central India Medicinal Plant Dataset (CIMPD), released by Tomar et al. in Data in Brief (Elsevier, 2025), contains 9,130 high-resolution leaf images spanning 23 Indian medicinal plant species under both healthy and diseased conditions. No comprehensive multi-architecture deep learning benchmark has been published on this dataset beyond the ResNet18 baseline in the original paper. This work presents the first end-to-end transfer learning, ensemble, and GradCAM++ interpretability pipeline on CIMPD. Six architectures are benchmarked: a Simple CNN trained from scratch, ResNet50, EfficientNet-B3, ConvNeXt-Tiny, Swin Transformer-Tiny, and a soft-voting Hybrid Model that combines the three strongest backbones using validation-accuracy-proportional weights. Training employs MixUp augmentation (α=0.4), CoarseDropout regularisation, label smoothing (ε=0.1), AdamW optimisation, and cosine annealing with warm restarts. The proposed Hybrid Model achieves 90% test accuracy with weighted F1=0.885, Cohen's κ and MCC substantially above chance, and AUC=1.00 on all four reported class categories, outperforming the Simple CNN baseline by 15 percentage points and the prior ResNet18 result by approximately 7 points. GradCAM++ saliency maps confirm that model attention concentrates on biologically meaningful leaf regions. All experiments are reproducible through a single self-contained Jupyter notebook.
Keywords:	Medicinal plant disease classification, CIMPD, EfficientNet-B3, ConvNeXt, Swin Transformer, transfer learning, ensemble learning, GradCAM++, MixUp augmentation, deep learning, precision agriculture
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN INTELLIGENT NSNT-QRNN FRAMEWORK FOR AUTOMATED STAGE-WISE PANCREATIC CANCER DETECTION AND CLASSIFICATION FROM CT IMAGES
Author:	DHISHYA. E , ANANDABABU. P
Abstract:	One of the deadliest and most aggressive types of cancers, pancreatic cancer has historically been difficult to diagnose and treat because it is typically detected when it reaches a late stage, and there are limited options for treatment. As such, there is an urgent need to develop novel methodologies to provide earlier detection and enable radiologists to make informed decisions about patient care. To meet this demand, a fully automatic stage-wise pancreatic cancer disease detection and classification model has been proposed. This approach uses a Quasi-Recurrent Neural Network (QRNN), specifically designed to use NasNet-based architectures with a Quasi-Recurrent Neural Network. The NasNet-based architectures used within our proposed methodology are capable of extracting hierarchical deep features from CT scan images. These features are learned through the process of training on large amounts of labelled data. These deep features are then fed into a Quasi-Recurrent Neural Network (QRNN). It uses both convolutional and recurrent layers. These layers enable the QRNN to learn how to associate information across time. The association capabilities provided by the quasi-recurrent layers allow the QRNN to identify relationships between input information. The proposed framework consists of three main components. Initially, a pre-processing technique - the Wiener filter was used to reduce the image’s noise. Secondly, Contrast-Limited Adaptive Histogram Equalization (CLAHE) was used to enhance the contrast of CT images. Following that, the TransUNet model was used to segment pancreatic tumors. The Quasi-Recurrent Neural Network is trained to classify pancreatic tumors based on their stage. The effectiveness of the proposed methodology is assessed through a comparative analysis with other state-of-the-art methodologies. Results show that the proposed framework provides a classification accuracy of 98.90% and performs better than previously reported methodologies.
Keywords:	Deep Learning, Pancreatic Disease, NasNet Segmentation, Computed Tomography, Image Processing
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	DETERMINANTS OF ARTIFICIAL INTELLIGENCE ADOPTION IN MALAYSIA’S PUBLICLY LISTED MANUFACTURING SECTOR
Author:	SHIAU WEI CHAN, KONG CHOI CHEW, HIROHIKO MORI, DIYANA SYAFIQAH ABD RAZAK
Abstract:	Artificial intelligence (AI) has the potential to transform organizational performance; however, AI adoption among Malaysian publicly listed manufacturing companies (PLMCs) remains relatively low and uneven. Key challenges include limited technological readiness, insufficient organizational support, and unclear external pressures influencing adoption decisions. Therefore, this study aims to (1) examine the current level of AI adoption, (2) determine the relationships between Technology–Organization–Environment (TOE) determinants and AI adoption, and (3) assess the impact of these determinants on AI adoption among PLMCs in Malaysia. A quantitative research design using a descriptive correlational approach was employed. The respondents comprised 126 top management members, managers, and supervisors representing PMLCs. Simple random sampling was used to select the sample. Data were collected through a structured self-administered online questionnaire using a 7-point Likert scale. Statistical analyses, including descriptive statistics, correlation analysis, and multiple regression analysis, were conducted using SPSS. The results indicate that AI adoption among Malaysian PLMCs is still at an early stage, with most firms planning or partially implementing AI. Correlation analysis revealed that all TOE determinants significantly relate to AI adoption, although the relationships were weak to moderate. Regression analysis showed that competitive pressure is the only determinant that significantly influences AI adoption, while technological and organizational factors play supportive roles. The findings provide important theoretical and practical implications by highlighting the dominant role of environmental forces in driving AI adoption. The study contributes to the TOE literature and offers insights for managers and policymakers to design strategies and policies that promote sustainable AI adoption in the manufacturing sector.
Keywords:	Artificial Intelligence (AI) Adoption, Technology–Organization–Environment (TOE) Determinants, Manufacturing Companies, Publicly Listed Companies
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	ALGORITHMIC PERSUASION, TRUST EROSION, AND THE TRUST–BEHAVIOR GAP IN DIGITAL COMMERCE
Author:	SRINIVASAN JAYASANKAR, VIJAYALAKSHMI RAMACHANDRAN, AMBULI VELAYUDHAM, KANNAMUDAIYAR SANKARANARAYANAN
Abstract:	Algorithmic personalization shapes how consumers encounter persuasive messages in digital commerce. Platform success is often inferred from behavioral engagement metrics, treating continued engagement as indicating healthy platform-consumer relationships. Drawing on Persuasion Knowledge Theory and attribution theory, this study examines how consumers interpret algorithmic persuasion cues and their impact on trust and disengagement intentions. Survey data from 737 digital commerce users in India, analyzed through structural equation modeling, shows that perceived algorithmic manipulation predicts inferred manipulative intent, which erodes relational trust. Advertising skepticism weakens this relationship by attenuating the effect of inferred intent on trust erosion. Trust erosion does not predict platform disengagement intentions, revealing a trust-behavior gap in algorithmic environments. Continued engagement may coexist with declining relational evaluations, suggesting engagement metrics can mislead as indicators of communicative effectiveness. This study advances research on trust and algorithmic influence by reframing algorithmic persuasion as an interpretive process rather than purely behavioral.
Keywords:	Algorithmic Personalization, Digital Commerce Platforms, Consumer Trust, Persuasion Knowledge, Advertising Skepticism.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	ENHANCING OBJECT DETECTION ACCURACY IN LOW LIGHT ENVIRONMENTS THROUGH GDIP BASED ADAPTIVE IMAGE PROCESSING AND CONTRAST OPTIMIZATION TECHNIQUES
Author:	ZEBA MASROOR, DR.V. K. DAYA SAGAR
Abstract:	Accurate object detection under low-light conditions remains a persistent challenge in computer vision due to poor visibility, high noise levels, and diminished edge clarity. This research addresses these limitations by introducing a structured image preprocessing framework that combines contrast enhancement, noise suppression, and Gradient Domain Image Processing (GDIP)-based adaptive edge amplification to improve detection performance. The enhancement pipeline is implemented using Python and OpenCV, while YOLOv5 serves as the detection backbone. Raw low-light images are sequentially processed through histogram equalization, bilateral filtering, and GDIP transformation before being passed to the detection network. Performance evaluation was conducted on the ExDark and LLVIP datasets. The proposed method achieved a precision of 81.3%, recall of 74.2%, and mAP@0.5 of 78.5%, representing significant gains over the baseline YOLOv5 model, which recorded 64.2% precision, 58.6% recall, and 61.0% mAP@0.5. Despite a modest increase in computational cost, with total inference time rising to 28.6 milliseconds per image, the system maintained 34.9 FPS, confirming its real-time applicability. This integrated enhancement approach not only boosts detection accuracy but also supports reliable deployment in low-light scenarios such as nighttime surveillance, autonomous navigation, and smart urban environments.
Keywords:	Low-Light Object Detection, GDIP Enhancement, Contrast Adjustment, Noise Reduction, YOLOv5, Image Preprocessing
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN ADAPTIVE FRACTIONAL DEEP LEARNING MODEL FOR ACCURATE MEDICAL IMAGE SEGMENTATION
Author:	L. PRASANNA, SITA SOWJANYA PRAKHYA, V. SIVA NAGA MALLESWARI, CH. BHAVANI, K. PAVANI, D. ANUSHA, DASARI YUGANDHAR, NIMMAGADDA MURALIKRISHNA, DARA RAJU
Abstract:	Medical image segmentation plays a crucial role in computer-aided diagnosis, disease monitoring, and clinical decisions. Existing deep learning segmentation methods perform poorly in low-contrast regions, at edges, and across different anatomical regions of the human body. This investigation presents an Adaptive Fractional Deep Learning Model (AFDLM) for high-accuracy medical image segmentation. The model integrates Adaptive Fractional Convolution (AFC) Operations, the Fractional Adaptive Attention Module (FAAM), and the Fractional Gradient Optimization Strategy (FGOS). The proposed framework provides a more informative context for feature extraction. It also maintains high accuracy in delineating fine anatomical details and helps preserve segmentation accuracy, even in the presence of imaging noise. A series of experiments has been carried out on normalized metrics, including Dice Similarity Coefficient (DSC) for brain tumor MRI, Lung CT, and Retinal Vessel data sets, as well as Intersection over Union (IoU), Precision, Recall, and Hausdorff Distance (HD). The proposed AFDLM has achieved high performance, outperforming previous models such as U-Net, Attention U-Net, DeepLabV3+, and TransUNet, with an average DSC of 94.0% and an IoU of 89.3%. The visual analysis also showed that the vessels were preserved, lesion edges were well delineated, and extraction of low-contrast lesions was improved. Adaptive fractional operations enable robust segmentation across multiple modalities is achieved using adaptive fractional operations, which yield inter-segmentation noise reduction and improved structural consistency. The proposed framework suggests that fractional deep learning is an intriguing future path for a next-generation medical image segmentation system with enhanced capabilities to support diagnosis and high clinical reliability.
Keywords:	Medical Image Segmentation; Fractional Deep Learning; Adaptive Attention Mechanism; Brain Tumor Segmentation; Lung CT Analysis; Retinal Vessel Extraction
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AI-DRIVEN ENERGY EFFICIENT EDGE COMPUTING FRAMEWORK FOR PREDICTIVE SMART STREET LIGHTING SYSTEMS
Author:	B. SARITHA, M. SRIRAMA LAKSHMI REDDY, S. KOMAL KOUR,V. V. RAMA KRISHNA, S. BHANU PRAKASH,TADAVARTHI NAGA NAVYA, N. PUSHPA LATHA, PUNYALA RAMADEVI, DARA RAJU
Abstract:	The rapid growth of the smart city infrastructure has driven the demand for smart and energy-efficient street lighting systems which can operate autonomously and make decisions in real-time. The current street lighting systems have significant communication overhead, are slow to respond, and consume excessive energy because they use a static algorithm for controlling the lamps, with the resulting control decisions sent up to the cloud. This research presented an Energy-Efficient Edge Computing Framework for AI Based Smart Street Lighting Systems that integrate IoT sensors, distributed edge computing and hybrid CNN-LSTM deep learning model for predictive adaptive illumination control. The proposed framework involved using edge devices for local processing of traffic surveillance images and environmental sensor data, so as to predict pedestrian and vehicle activity in real time and reduce the reliance on the cloud. The CNN model was used for spatial traffic feature extraction, and the LSTM network was used for the temporal activity feature of intelligent brightness optimization. Experimental tests were performed on about 120,000 traffic images and 1.8 million logs of time-synchronized IoT sensor data acquired under different environmental conditions. The proposed framework has been able to obtain 97.2% of prediction accuracy, 41.8% energy savings, 52.4% reduction of communication latency, 46.1% improvement of response time, and 61.6% reduction of bandwidth utilization compared with existing smart lighting systems. The findings showed that embedding edge intelligence into deep learning models based on hybrid architectures resulted in substantial gains for real-time responsiveness, scalability, and efficiency of operations. The proposed framework offers a low latency, scalable, and sustainable solution for future smart city street light infrastructures powered by AI. The experimental results validate that the proposed framework offers an effective, energy-efficient, and low-latency solution for intelligent smart street lighting, underscoring the promise of AI-driven edge computing for the future of smart city infrastructures.
Keywords:	Edge Computing, Smart Street Lighting, CNN-LSTM, Energy Efficiency, IoT Sensors, Predictive Illumination Control.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	MULTI-CLASS SEVERITY DETECTION OF CYBERBULLYING IN SAUDI-DIALECT TWEETS USING DEEP LEARNING
Author:	BADER AZI ALANAZI, CHIN-TENG LIN2, YU-CHENG FRED CHANG
Abstract:	Cyberbullying in Arabic social media poses a growing threat to users' psychological well-being, yet automated detection systems largely overlook both severity levels and dialectal variation, particularly the Saudi dialect. This gap limits the practical utility of existing systems for real-world content moderation. This study develops an automated deep-learning approach to classify the severity of cyberbullying in Saudi-dialect tweets into four levels: high, medium, low and non-cyberbullying. Twenty-eight experimental scenarios were constructed by systematically combining Arabic NLP toolkits, stop-word removal strategies and class-balancing techniques. Four deep learning models (CNN, LSTM, BiLSTM and GRU) were trained and evaluated. Data balancing and preprocessing improved classification performance; CAMeL with random oversampling produced the most consistent gains across models. Compared with previous machine learning baselines on the same dataset (maximum accuracy 92.23%), CNN achieved 95.92% and LSTM reached 95.59%. These findings have direct implications for automated content moderation platforms serving Arabic-speaking communities, enabling graduated responses calibrated to cyberbullying severity. To the best of our knowledge, this is the first study that uses deep learning to detect cyberbullying severity in Saudi-dialect tweets, providing a scalable, linguistically informed solution.
Keywords:	Text Classification, Deep Learning, Cyberbullying Detection, Arabic social media, Saudi dialect.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AI-DRIVEN CYBERSECURITY FOR DIGITAL TWIN-ENABLED IOT SYSTEMS: CHALLENGES, SOLUTIONS, AND PERFORMANCE EVALUATION IN SMART CITIES
Author:	V. PURUSHOTHAMA RAJU, DR. MD SIRAJUL HUQUE, DR. SWATHI AGARWAL, JAGADEESH SUNDHARAMOORTHY, SUDHEER CHOUDARI, Dr. R. SUBA SREE
Abstract:	The large-scale Internet of Things (IoT) deployments over smart grids, intelligent transportation, healthcare and environmental monitoring are increasingly relied upon by smart cities, which generate enormous real-time telemetry but remain highly vulnerable to advanced cyber threats such as data manipulation, ransomware and zero-day attacks. The conventional signature-based intrusion detection systems are not effective against dynamic threats, not real time bidirectional synchronized, root cause analysis and safe testing facilities without disrupting the live operations. This work fills in these key gaps by proposing Urban Twin Secure, an active hybrid framework that tightly integrates digital twins to provide a two-way synchronization system and a sandboxed simulation system, permissioned blockchain to provide a system of immutable logging and access control, a explainable root-cause diagnosis and impact prediction system, and a feature prioritization system that relies on explainable artificial intelligence, using SHAP as a framework. The framework achieves macro F1-score of 0.942, which is 15-18% higher than the state-of-the-art baselines, and reduces the mean time to diagnosis by 65-70, the false positive rate by 20-25, and maintains sub-second inference on edge devices. These findings show a better detection accuracy, diagnostic efficiency, and strength in multi-domain settings. Urban Twin Secure, the first-of-its-kind comprehensive smart-city-scale integration of digital twin sandboxing, blockchain integrity, and dual contrastive-causal explainable intelligence, provides a scalable, tamper-resistant solution that can improve cyber resilience, help maintain regulatory compliance, and strengthen operational continuity in critical urban infrastructure.
Keywords:	Causal Learning, Digital Twin, IoT Cybersecurity, Smart Cities, Blockchain
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN INTEGRATIVE DEEP LEARNING FRAMEWORK FOR MULTI-MODAL RARE DISEASE DIAGNOSIS
Author:	Dr SUTHA K, RAJEASHWARI SRINIVASA RAGAVAN, Dr N SUBHASH CHANDRA, Dr S.PAVAN KUMAR REDDY, D.BHAVANA, Dr. G B HIMA BINDU, PACHIYAPPAN C, Dr B.Jalender, , DR.T.VENGATESH, Dr.BH.KRISHNA MOHAN, B.SRILAKSHMI
Abstract:	Rare diseases affect 300-400 million people globally, yet individual conditions are so uncommon that few clinicians gain sufficient experience, and data scarcity prevents conventional AI from learning reliable diagnostic patterns. This paper creates three new knowledge contributions. First, while prior multimodal efforts have combined at most two data types (images+HPO terms or genomes+simple phenotypes), we introduce the first end-to-end deep learning architecture that integrates three raw, high-dimensional modalities: unstructured clinical text, unconstrained facial photographs, and full genomic variant sequences. Second, we propose a gated cross-attention fusion mechanism that learns to dynamically weight modalities for Williams-Beuren syndrome, the model attends 62% to facial features; for Smith-Magenis syndrome, it shifts to 58% attention on behavioral text descriptions providing interpretable, clinician-friendly explanations. Third, we demonstrate that embedding this architecture within a Few-Shot Learning (FSL) paradigm is not incremental but fundamental: removing FSL drops accuracy from 94.7% to 81.3% (-13.4%), proving that standard supervised learning fails catastrophically in data-scarce medical domains. On a curated dataset of five rare genetic syndromes (n=5,250 multimodal samples), our Unified Deep Learning Framework (UDLF) achieves 94.7% accuracy and 0.92 F1-score in a 5-way 5-shot setting, significantly outperforming unimodal baselines (<80%) and bimodal fusion (<87%). This establishes that principled multimodal integration is not merely helpful but essential for building robust, generalizable AI systems for rare disease diagnosis
Keywords:	Rare Diseases, Deep Learning, Multi-Modal Learning, Feature Fusion, Medical AI, Few-Shot Learning, Genomics, Phenotypic Analysis.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	EVOLUTIONARY SOFTWARE TEST CASE OPTIMIZATION USING REAL-CODED GENETIC ALGORITHMS
Author:	REVATHI TALARI, CH. SUDHEER, PEDDIREDDY SOWMYA REDDY, N. S. L. KUMAR KURUMETI, GUNDALA VENKATA RAMA LAKSHMI, SRINIVASA RAO NIDAMANURU, M VENUNATH, 8SESIKALA BAPATLA
Abstract:	Software testing is a critical phase in software development, making sure the product is reliable correct, and has good overall quality. Still, with big software systems, you can end up with a substantial number of test cases, which then ramps up execution time, computational cost, and yes the maintenance effort too. The whole idea of test case optimization is to shrink the test suite size, but still keep strong fault detection ability and solid code coverage. In this paper we present an Evolutionary Software Test Case Optimization framework, built around a Real-Coded Genetic Algorithm, RCGA, to pick an optimal subset of test cases from a large repository, in an efficient way. Compared with the usual binary genetic approaches, the proposed RCGA uses real-valued chromosomes to represent candidate solutions, so the search space gets explored more smoothly, and the method converges faster toward better solutions. The optimization process starts by creating an initial population of test case subsets, then it runs fitness evaluation using multiple criteria, including code coverage, fault detection effectiveness, execution cost, and also redundancy reduction. A composite fitness function is meant to maximize the effectiveness of tests and to minimize the number of test cases selected. High-quality offspring and the diversity of the population throughout the evolution are created using the real coded crossover and mutation operators. The experimental results were obtained using the benchmark software testing dataset of test cases with 500 to 2,000 problems. Results show that the proposed RCGA can reduce the size of test suite significantly: only 35-45% of the original test set is kept, and fault detection capability and coverage is kept over 95%. The average number of test cases was decreased from 1500 to about 620 optimized test cases. Moreover, the fitness value gradually increased in the generations from 0.62 to 0.96, which is an excellent convergence toward a high fitness solution. The results of Comparative analysis with the conventional methods of Genetic Algorithms and random selection methods reveal that RCGA outperforms the conventional methods in terms of fitness values, convergence and test suite reduction. The above results validate the proposed evolutionary optimization method for reduction of number of test cases whilst achieving maximum fitness and quality of the overall software testing environment.
Keywords:	RCGA, Optimization, Testing, Fitness, Quality.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A NOVEL ATTENTION-BASEDSKIN CANCER DETECTION USING DERMATOLOGICAL IMAGES
Author:	VIDHYASHREE B , Dr PRASANNA LAKSHMI G S , CHETLA CHANDRA MOHAN , G KIRAN KUMAR , Dr CH BHAVANI , ANNAMARAJU THANUJA , PAVINDRA REDDY PONDUGALA , MELAM NAGA RAJU
Abstract:	Skin cancer is characterized by the abnormal growth of skin cells, caused by various factors. Recently, the number of people affected by skin cancer has been steadily increasing. Early diagnosis plays a vital role in improving patient outcomes, while late detection can lead to complications and, in some cases, mortality. Traditionally, skin cancer detection relies on visual inspection by dermatologists, a process that can be time-consuming, requires significant expertise, and is prone to errors. These challenges have attracted more researchers to develop autonomous systems for skin cancer detection. However, several issues remain unresolved, such as achieving high accuracy, which is still difficult due to minimal variation between benign and malignant cells, as well as the dermatological images being obscured by hair, moles, or skin texture. While some models achieve higher accuracy, they are highly complex and unsuitable for deployment on edge devices. To address this, the research proposed a novel Attention-based MobileNet-Transformer for skin cancer detection. The processed images are given to MobileNet for effective feature extraction with minimal parameters, and the extracted features are refined using a Convolutional Block Attention Module (CBAM) attention mechanism. The refined features are then given to a transformer module equipped with a fully connected layer (FCL) for capturing temporal dependencies and performing classification. The proposed model effectively extracts all the features from the skin images to achieve better accuracy. When compared with existing models like MobileNet, Xception, and DenseNet, the proposed approach achieved the highest accuracy of 94%, whereas the others remained below 90%. The ablation research and state-of-the-art comparison highlight the model’s effectiveness. The experimental analysis shows that the proposed network is suitable for real-time skin cancer detection because of its reliable output and lower model complexity.
Keywords:	Skin Cancer, Dermatological Images, Deep Learning, MobileNet, Transformer, Attention Mechanism, Confusion Matrix
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	STACKING-BASED ENSEMBLE CLASSIFICATION OF HAND MOVEMENTS USING WAVELET SCATTERING FEATURES EXTRACTED FROM RAW AND FILTERED SURFACE EMG SIGNALS
Author:	MOHAMED ELMEHDI AIT BOURKHA, ABDERRAHMANE BOUDRIBILA, ANAS HATIM, ASSIA SAYDTAHIRI
Abstract:	Surface electromyography (sEMG) signal classification plays a critical role in the development of intelligent prosthetic control systems and rehabilitation technologies for individuals with upper-limb motor impairments. Although numerous machine learning approaches have been proposed for sEMG-based hand gesture recognition, achieving high classification accuracy across multiple grasp types remains challenging, particularly when six or more movement classes are involved. The primary goal of this study is to develop and rigorously validate a robust yet computationally lightweight framework that attains high six-class classification accuracy while simplifying the conventional preprocessing pipeline required for real-time myoelectric control. To this end, two hypotheses are tested: first, that the wavelet scattering transform produces a feature representation sufficiently noise-robust to render conventional band-pass and notch filtering unnecessary; and second, that combining complementary K-Nearest Neighbors classifiers through a two-level stacking ensemble yields higher accuracy than any individual classifier or existing single-model approach reported on the same benchmark. The proposed framework combines wavelet scattering transform (WST)-based feature extraction with a stacking ensemble learning strategy for the recognition of six fundamental hand movements (Cylindrical, Tip, Hook, Palmar, Spherical, and Lateral) from a publicly available benchmark dataset (UCI sEMG, five subjects, two channels, 500 Hz). The wavelet scattering network decomposes each raw sEMG signal into a compact 6 × 202 time–frequency feature tensor, achieving a 59.60% dimensionality reduction while preserving discriminative information across scattering orders. A systematic comparison between filtered and raw sEMG signals demonstrates that the scattering transform renders conventional band-pass and notch filtering unnecessary, as both signal types yield comparable classification performance. At the classification stage, a two-level stacking architecture is employed: K-Nearest Neighbors (K = 3) and ensemble subspace KNN (20 learners) serve as base-level classifiers, and their concatenated predictions are fed into a meta-learner (ensemble subspace KNN, 30 learners). Evaluated through five-fold cross-validation, the proposed framework achieves an overall classification accuracy of 99.67%, with average sensitivity, specificity, precision, and F1-score of 99.67%, 99.93%, 99.67%, and 99.67%, respectively, outperforming all existing methods reported on the same dataset. These results demonstrate the effectiveness of combining scattering-based representations with ensemble stacking for robust and computationally efficient sEMG classification, with potential applications in myoelectric prosthetic control and clinical rehabilitation.
Keywords:	Surface Electromyography (sEMG), Hand Movement Classification, Wavelet Scattering Transform, Ensemble Learning, Stacking, K-nearest Neighbors, Prosthetic Control
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	FUSIONSCANAI: AN ATTENTION-BASED MULTI-MODAL FRAMEWORK FOR CANCER DETECTION FROM MRI AND CT IMAGES
Author:	DR. S. SWAPNA RANI, JOHN BABU GUTTIKONDA, KAPARTHI UDAY, MV NAGESH, DAMERA VENKATESH
Abstract:	Diagnosis of metastatic cancer remains one of the challenging problems in clinical oncology, as disease spread is highly complex and heterogeneous by nature. Widely used deep learning models extract features from single-modality inputs, including imaging or clinical data, but are thus hindered in their diagnostic robustness and generalizability. Meanwhile, existing fusion strategies do not sufficiently learn the complex cross-modal interactions that are essential for reliable metastasis prediction. To overcome these limitations, we propose MetaBreastAI, a new explainable multi-modality metastasis detection framework, combining medical imaging and clinical information. The central component of the framework is a hybrid dual-stream architecture (MetaBreastNet) that fuses a ConvNeXt-based CNN and a Swin transformer-based encoder to extract informative local and global features from imaging and clinical data, respectively. Firstly, we introduce a multi-head cross-attention-based interaction fusion module to exploit the inter-modal interactions. Then, the gated attention is used to refine the salient features further. To gain better accuracy and interpretability for metastatic prediction, the fused embeddings are routed through a fully connected classifier. Extensive experimental evaluations on benchmark datasets show that the proposed model outperforms the existing baselines, achieving state-of-the-art AUC, accuracy, and F1-scores. We provide ablation studies confirming that each of our architectural modules contributes to the performance gains, and results from an explainability method allowing us to visualise the model’s decision process. Abstract Introduction the MetaBreastAI framework not only enhances metastatic cancer diagnostic accuracy but also provides a clinically aligned decision support tool. The modular architecture and explainability capabilities make it applicable to more cancer types and multi-source medical datasets.
Keywords:	Metastasis Detection, Multi-Modal Learning, Explainable Deep Learning, Swin Transformer, Medical Image Analysis
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	OPTIMIZING SECURITY IN IOT ECOSYSTEM WITH BLOCKCHAIN INTEGRATED ARTIFICIAL INTELLIGENT BASED INTRUSION DETECTION
Author:	SUBBAIAHGARI R AJITHA , Dr G V RAMESH BABU
Abstract:	The Internet of Things (IoT) is an emerging technology that has enhanced connectivity across multiple industries such as Industry 4.0 and smart cities. Intrusion Detection Systems (IDS) play a crucial role in identifying and mitigating cyber threats in such environments. The centralized architecture of traditional security systems poses challenges for scalability and performance, which hinders their ability to handle the complexity and volume of IoT devices. The problems cause latency, power usage and inefficient intrusion detection. This research proposes an intrusion detection system using artificial intelligence and blockchain concepts to boost security in the Internet of Things (IoT) environment. The primary goal is to use blockchain to create a secure framework without the need for a central authority to oversee the data. Feature extraction is done by a dual-ResNet model and feature selection is done by modified Coati Optimization (MCO) algorithm. The resulting feed-forward artificial neural network (FF-ANN) is then re-used for intrusion detection, greatly enhancing the security of IoT networks. The main IT contribution of this research is the creation of the intelligent and decentralized cybersecurity framework, which uses deep learning to detect intrusions and blockchain to guarantee data integrity. The proposed approach also has advantages compared to the traditional centralized IDS solutions, such as trust, less centralized IDS monitoring dependency, and reliable attack identification in the resource-constrained IoT environment. The combination of Dual-ResNet feature extraction, optimized feature selection, and blockchain verification adds to the scalable and secure IoT infrastructure.The proposed model is evaluated using the ToN-IoT and CICIDS-2017 datasets, achieve exceptional predictive accuracy of 98.8 % and 98.937%, respectively, outperforming existing models.
Keywords:	Blockchain, Artificial Intelligent, Intrusion Detection System, Internet Of Things, Secure Data Management
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	HYBRID CNN LSTM MODEL FOR EARLY HEART DISEASE PREDICTION USING ECG AND CLINICAL DATA
Author:	DIVYA LINGINENI, NAGASIVA JYOTHI KOMPALLI, CHINTA SRI DIVYA, G. MADHAVI, S. VANISRI, VITHYA GANESAN, DASARI YUGANDHAR, ANIL KUMAR KATRAGADDA
Abstract:	The early identification of heart disease is not only crucial for cardiovascular mortality reduction, but for clinical decision-making as well. But, the common machine learning methods do not account for complex spatiotemporal ECG patterns and do not take into account valuable clinical patient information. The proposed research aimed to present a novel Hybrid Convolutional Neural Network - Long Short-Term Memory (CNN-LSTM) for early prediction of heart disease based on ECG signals and clinical data. The methodology proposed in this paper combined ECG preprocessing, CNN-based spatial feature extraction, LSTM-based temporal learning, adaptive multimodal clinical data fusion, and attention-based optimization in a single framework to realize deep learning. The experimental evaluation was carried out using benchmark datasets, such as the MIT-BIH Arrhythmia Dataset, the PTB Diagnostic ECG Dataset, and the Cleveland Heart Disease Dataset. The proposed model was found to perform better with the reported accuracy of 98.7%; precision, recall, F1, and AUC were 98.2%, 98.5%, 98.3%, and 99.1%, respectively, when compared with the conventional ML and standalone deep learning models. Adding deep features from ECG to clinical parameters led to a substantial boost in prediction accuracy, a decrease in false diagnosis, and an added benefit in the detection of cardiovascular disease at an early stage. The experimental outcomes illustrated that the introduced Hybrid CNN-LSTM framework is an effective, sturdy, and intelligent method for real-time heart disease prediction and superior healthcare monitoring applications.
Keywords:	Heart Disease Prediction, ECG Signal Analysis, Hybrid CNN-LSTM, Deep Learning, Clinical Data Fusion, Cardiovascular Disease Detection
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	POLYCYSTIC OVARIAN SYNDROME PREDICTION USING AN EFFICIENT NEURAL NETWORK-BASED LEARNING MODEL
Author:	B. BAZEER AHAMED, KALAICHELVI NALLUSAMY, NANDHAGOPAL SUBRAMANI, V. JEMMY JOYCE, J. ALFRED DANIEL, K. ARUNKUMAR, P. SHERUBHA
Abstract:	Polycystic ovarian syndrome (PCOS) is a hormonal disorder diagnosed during a woman's reproductive years. Several techniques exist to detect PCOS, but few methods address the combination of mental health issues with PCOS. This proposed system features an automated early detection model that assesses the likelihood of both PCOS and mental health problems accurately. Many individuals respond in natural language during real-life applications to share their health status in response to clinician inquiries. The underlying linguistic structure, which links symptomatic manifestations to the diagnostic process of Polycystic Ovary Syndrome, is exploited through a learning-based modelling paradigm. This approach is called Bounding-box based Convolutional Neural Networks (BB-CNN). The model effectively learns and extracts the most influential features to improve prediction accuracy. The proposed BB-CNN achieves 97.34% accuracy and 96.01% precision, recall, and F-measure. Various statistical measures are evaluated and compared with other methods.
Keywords:	PCOS, Hormonal Disorder, Deep Learning, Convolutional Neural Networks, Bounding Box, Statistical Measures
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A NOVEL SENTIMENT ANALYSIS APPROACH: INTEGRATING SEMANTIC UNDERSTANDING, DEEP FEATURES, AND MACHINE LEARNING FOR ACCURATE CLASSIFICATION
Author:	V.SUGANYA , K. PREMA
Abstract:	This paper introduces an advanced analysis framework meticulously crafted with an optimized approach for the evaluation of social data and the classification of sentiments. In the contemporary digital landscape dominated by social media and online interactions, the sheer volume and complexity of user-generated data present a formidable challenge. Traditional sentiment analysis methodologies often struggle to effectively interpret the nuances and evolving nature of language expressed in online spaces. To overcome these limitations, our proposed framework leverages cutting-edge technologies such as natural language processing (NLP), deep learning (DL), and machine learning (ML) algorithms. The primary contribution of this research is the development of an optimized unified framework which integrates advanced pre-processing, DL-based sentiment classification, and feature extraction. The developed approach contributes new knowledge by addressing constraints observed in large-scale unstructured social data interms of classification efficacy and contextual understanding. The framework's multi-faceted approach encompasses thorough data preprocessing, sophisticated NLP techniques for semantic understanding, utilization of deep learning models for feature extraction, and the implementation of an optimized sentiment classification algorithm. This integrated methodology not only enhances the accuracy of sentiment analysis but also ensures scalability, adaptability, and real-time capabilities. This paradigm becomes an essential resource as social media platforms develop for academics, corporations, and decision-makers who want to glean insights from the immense ocean of online interactions and comprehend the complex web of emotions influencing the digital conversation.
Keywords:	CNN, NLP, Deep learning, ConvNet-SVMBoVW Model, Adam Optimization Algorithm, Sentiment Classification
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN EXPLAINABLE HYBRID CNN TRANSFORMER AND ATTENTION REFINEMENT FRAMEWORK FOR INTELLIGENT UAV BASED STRUCTURAL DAMAGE DETECTION
Author:	RAMANJANEYULU SEGGEM, G. SUSHMA, LAKSHMI PANUGANTI, CH. NIRANJAN KUMAR, GUNAY ALIYEVA SAFI, NIRMAL MAHESH, P.V.S. MARUTHI KRISHNA, GARIGIPATI RAMA KRISHNA
Abstract:	Structural aging, environmental stress, corrosion, mechanical fatigue, and similar factors are causing an increasing number of critical civil infrastructure elements—such as bridges, tunnels, dams, and high-rise structures—to deteriorate. Current manual inspection approaches are prone to human error. They are also time-consuming, expensive, hazardous, and restrict effective SHM (Structural Health Monitoring). In this research, a novel Explainable Artificial Intelligence (XAI) guided UAV (Unmanned Aerial Vehicle)-based structural damage detection framework is proposed. This Hybrid Explainable Structural Vision Network (HESVN) aims to provide accurate, interpretable, and reliable infrastructure inspection methods. The approach combines a UAV (Unmanned Aerial Vehicle) for image acquisition and EfficientNet-B4, a type of convolutional neural network, for multi-scale feature extraction. It also includes transformer attention (a deep learning mechanism for focusing on relevant parts of the data) for encoding structures, Explainability-Guided Attention Refinement (EGAR, which enhances model interpretability), and Grad-CAM visualization (a technique for visualizing model decision areas) for transparent structural damage diagnosis. Several benchmark datasets were used under different environmental conditions. These include SDNET2018 (a concrete crack image dataset), CFD (Concrete Fracture Dataset), datasets collected with UAVs for bridge inspection, and images captured by a self-collected UAV. The HESVN framework outperformed other models, including CNNs (Convolutional Neural Networks), YOLOv5 (a real-time object detection system), EfficientNet, and the Vision Transformer. It achieved higher accuracy (96.8%), precision (95.9%), recall (96.3%), F1-score (96.1%), and structural explainability index (SEI) (91.4%). Experimental results were robust even under low-light, motion-blur, and shadow-challenging conditions. The framework also created reliable, easily interpretable visual explanations of damage localization. In safety-critical infrastructure monitoring, incorporating comprehensibility greatly enhances engineers' trust, understandability, and decision-making dependability. The proposed framework provides a robust platform for next-generation autonomous structural health monitoring systems, including smart city infrastructure management and explainable AI-assisted inspection technologies.
Keywords:	Explainable Artificial Intelligence (XAI), UAV based Structural Inspection, Structural Damage Detection, Deep Learning, Transformer Attention, Structural Health Monitoring (SHM)
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	ADAPTIVE MACHINE LEARNING FRAMEWORK FOR REAL TIME HEALTH RISK ASSESSMENT USING WEARABLE IOT SENSOR DATA
Author:	S SINDHURA, K VENUGOPAL, U. POORNA LAKSHMI, SANDEEP KOTTE, SWATHI VODDI, YALANATI AYYAPPA, ROJA D
Abstract:	The sudden rise in the development of wearable Internet of Things (IoT) technology has opened the door to new opportunities in terms of real-time health risk assessment due to the opportunity to continuously monitor physiological indicators. To evaluate risks based on the data obtained with help of wearable sensors such as heart rate, temperature, blood oxygen saturation (SpO 2) and activity levels, this study suggests a flexible machine learning structure. The system uses several deep learning and machine learning algorithms when it comes to data preparation and feature extraction. These are LSTM, CNN, RF and SVM. The streaming sensor data is continuously updating the model parameters by the adaptive learning process in the framework, enhancing the quality of prediction and system reliability. As per the experimental data, the LSTM model proves to be the best means of evaluating the time-series health data, and it is better than the alternative approaches in accuracy, precision, recall, and Area Under the Curve (AUC). Lastly, the proposed solution proposes an ingenious and adaptable resolution to the real-time health monitoring and risk prediction problems by capitalizing on the possibilities of wearable Internet of Things sensors.
Keywords:	Wearable IoT, Machine Learning, Health Risk Prediction, LSTM, Real-Time Monitoring.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN EFFICIENT DEEP LEARNING-BASED FRAMEWORK FOR EARLY MALNUTRITION DETECTION USING RESNET-101 AND RF–ADABOOST CLASSIFIER
Author:	Dr. LAKSHMI NAGA JAYAPRADA GAVARRAJU, VENKATESWARLU SUNKARI, Dr. SREENIVAS MADUPU, KANDULA ROJARANI, MVVS SUBRAHMANYAM, Dr. T. PREM CHANDER, Dr.G.KALPANA, AJAY KUMAR VEGI, Dr. SIVA KUMAR PATHURI
Abstract:	Malnutrition remains one of the most important public health problems in the world, especially in developing countries, with severe health consequences in children in the absence of early diagnosis. Traditional screening techniques can be labor-intensive, resource-dependent, and require trained health care workers to operate, making them less effective in achieving large-scale and timely screening. To overcome this issue, this study presents a novel deep ensemble efficient system for automated malnutrition detection system based on image data. The proposed method uses ResNet-101 to extract visual features to distinguish the malnutrition and uses two classifiers, namely: Random Forest (RF) and AdaBoost classifiers to enhance the classification accuracy and robustness. The performance of the framework is compared with the already available deep learning models such as RegNetY and VGG19. From the experimental results, it is observed that the proposed ResNet-101 with RF–AdaBoost model gives an accuracy of 95% which is higher than RegNetY (89%) and VGG19 (87%). The results show that deep extraction and ensemble learning is a reliable and efficient approach to detecting malnutrition. So, the suggested framework can be useful for large scale screening and regular monitoring which would help in early detection as well as management of healthcare for the children who are prone to malnutrition.
Keywords:	Malnutrition, Deep Learning, ResNet-101, RegNetY, Random Forest, ADAboost VGG19
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A HYBRID TEXT-GNN FRAMEWORK FOR SENTIMENT ANALYSIS AND CYBERBULLYING DETECTION IN SOCIAL MEDIA TEXT
Author:	D. JAGANNATHAN, N. V. BALAJI
Abstract:	The fast expansion of social media platforms has resulted in a higher amount of user-created text content, which requires automatic sentiment assessment and cyberbullying identification systems to ensure secure and positive online spaces. Traditional text classification methods fail to handle informal online language because they cannot understand its complex contextual relationships and semantic dependencies. The assessment of text contains three elements that create difficulties for accurate detection: noisy text elements, different linguistic patterns, and hidden abusive language. The existing system requires a framework that combines efficiency with intelligent capabilities to extract essential text-based data and select important features for accurate classification of sentiment and cyberbullying content across multiple datasets. This paper proposes a novel hybrid sentiment classification and cyberbullying detection model. The model integrates Multi-Resolution N-gram Embedding Fusion (MR-NEF)-based feature extraction, Dynamic Binary Swordfish Movement Optimization Algorithm (DBSMOA)-based feature selection, and Graph Neural Network with Text Graph Construction (Text-GNN)-based classification. For this research, IMDb, Yelp Polarity, and the Cyberbullying Classification datasets are applied as input datasets. Initially, the input data were preprocessed using NLP techniques like tokenization, stop word removal, and lemmatization. The developed MR-NEF + DBSMOA + Text-GNN model is assessed using three datasets individually. The model achieved its best performance with results like 98.64% accuracy, 98.57% precision, 98.52% recall, and 98.50% f1-score. Compared and validated with the current models and baseline variants, the developed model outperformed all the models with improved performance.
Keywords:	Cyberbullying Detection, Sentiment Classification, MR-NEF, DBSMOA, Text-GNN.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	ADAPTIVE POST-PROCESSING FUSION FOR MULTI-MODEL YOLO-BASED INSECT DETECTION
Author:	AKARID ABDERRAHIM, AIT EL ASRI SMAIL, EL ADIB SAMIR, RAISSOUNI NAOUFAL
Abstract:	In recent years, object detection has made tremendous progress, yet the best current detectors still fall short of achieving ultimate performance. Insect detection is vital for precision agriculture but remains challenging due to species similarity and complex environmental conditions. This study introduces a generic adaptive post-processing fusion framework that dynamically combines predictions from multiple heterogeneous object detectors using Non-Maximum Suppression (NMS) and Weighted Box Fusion (WBF). The proposed approach was evaluated on the IP102 dataset, focusing on 10 pest species. Individual You Only Look Once (YOLO) models (YOLOv5nu, YOLOv8n and YOLOv10n) attained accuracies of 95.73%, 95.37%, and 95.37% with Mean Average Precision (mAP) at Intersection over Union (IoU) 0.5 scores of 92.91%, 91.32%, and 92.57%, respectively. The adaptive fusion approach outperformed all single models, reaching a mAP at IoU 0.5 of 97.22% and an accuracy of 98.58%, representing a gain of approximately 5% and 2.85%, respectively, and surpassing fixed-weight NMS and WBF baselines by approximately 6% in mAP. These findings demonstrate that combining YOLO models through adaptive fusion significantly enhances insect detection performance, providing a reliable building block that can be integrated into future pest monitoring pipelines for smart agriculture systems.
Keywords:	Object Detection, YOLO, Weighted Box Fusion, Adaptive Ensemble Learning, Insect Pest Detection
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	DEEP LEARNING FRAMEWORK FOR CORROSION DETECTION IN UNDERGROUND PIPELINES USING GPR IMAGES
Author:	ANGIDI VEERABABU, K. SAI MADHURI, S.USHA BABY, N. S. L. KUMAR KURUMETI, P. KIRAN KUMAR, MANIKANDAN MOOVENDRAN, M. VENUNATH, KOMALI GOVINDU
Abstract:	Corrosion of underground pipelines is a major issue in oil, gas, and water transportation systems, as it affects physical stability, environmental safety, and maintenance costs. This study proposes a novel Attention-Guided Multi-Scale Deep Learning (AGMS-DL) framework for automatic detection of corrosion in underground pipelines using Ground Penetrating Radar (GPR) images. The proposed framework integrates adaptive preprocessing, multi-scale convolutional feature extraction, spatial attention learning, and residual feature combination, achieving accurate identification of corrosion patterns amid noisy conditions in underground environments. For experimentation, a GPR radargram dataset comprising 12,500 images of corroded and non-corroded pipelines was used under various soil conditions. The performance of the framework was assessed in terms of accuracy, precision, recall, F1 Score and Area Under Curve (AUC). Experimental results showed that the proposed AGMS-DL model outperforms other models, such as SVM, Random Forest, CNN, ResNet50, and EfficientNet, in terms of accuracy, precision, recall, F1-score, and AUC, with 98.4%, 97.9%, 98.1%, 98.0%, and 98.7%, respectively. The thorough analysis showed that the spatial attention mechanism was highly effective at localizing corrosion while minimising false detections caused by underground clutter and signal distortion. The proposed solution is not just reliable but also economical in terms of computationally efficient and intelligent enough to be implemented in real-time underground pipeline health-monitoring and predictive maintenance applications.
Keywords:	Underground pipeline corrosion, Ground Penetrating Radar (GPR), Deep learning, Spatial attention Mechanism, Multi-scale CNN, Corrosion detection
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	FUSING SWIN TRANSFORMER AND CNN FEATURES FOR ROBUST BRAIN TUMOR SEGMENTATION IN MRI
Author:	MOULIESWARAN ELAVARASU, PANIBHATE NEELAKANTESWARA, S. SRINIVASAN , Dr R USHA RANI, Dr S.PAVAN KUMAR REDDY, PANI BHATE VISWANATH, Dr. G B HIMA BINDU, Dr.K.HAZEENA, 9DR.T.VENGATESH, Dr.BH.KRISHNA MOHAN, A.ARUN
Abstract:	Automated and precise segmentation of brain tumors from Magnetic Resonance Imaging (MRI) is critical for diagnosis, treatment planning, and monitoring. While Convolutional Neural Networks (CNNs) have become the de facto standard, their limited receptive field often hinders performance in complex tumor sub-regions with diffuse boundaries. Recently, Transformer architectures, notably Swin Transformers, have demonstrated remarkable success in capturing long-range dependencies in visual data. This paper proposes a novel hybrid architecture, Swin-Net, that synergistically fuses a CNN encoder with a Swin Transformer branch for robust brain tumor segmentation. We quantitatively analyze the contribution of each component by comparing our model against state-of-the-art CNN-based (U-Net, nnU-Net) and Transformer-based (Swin-Unet) baselines on the BraTS 2020 dataset. Results demonstrate that Swin-Net achieves superior performance, with mean Dice scores of 92.1% for the whole tumor (WT), 88.5% for the tumor core (TC), and 84.2% for the enhancing tumor (ET). Extensive ablation studies confirm that the feature fusion mechanism is the primary contributor to this performance gain, providing a 4.7% boost in Dice for the gchallenging ET region over a pure CNN model. Our work establishes that the fusion of hierarchical CNN features with global Swin Transformer contexts sets a new benchmark for robust and accurate brain tumor segmentation.
Keywords:	Brain Tumor Segmentation, Swin Transformer, Convolutional Neural Networks, Feature Fusion, Deep Learning, Medical Image Analysis, BraTS.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AUTOMATION OF KNEE OSTEOARTHRITIS SEVERITY ANALYSIS OVER RADIOGRAPHIC IMAGING USING WEIGHTED ENSEMBLE LEARNING
Author:	B. CHRISTINA SWEETLINE, T. TAMILSELVI, G. JEEVAN NIRMAL RAJ, K. REVATHI, S. J. VIVEKANANDAN
Abstract:	Knee osteoarthritis (KOA) is a degenerative joint condition that gradually diminishes mobility and quality of life, particularly in older adults. Timely treatment relies heavily on accurate assessment of KOA severity, yet conventional diagnosis through the Kellgren–Lawrence (KL) grading system (grades 0–4) necessitates manual analysis of radiographs, which is labor-intensive and prone to inter-observer variability. To address these challenges, this research introduces an automated KOA grading framework based on deep learning, utilizing convolutional neural networks (CNNs) integrated with a cumulative ordinal regression network (CORN) architecture to comprehend the inherent order of KL grades. Three different CNN architectures ResNet50, DenseNet169, and DarkNet19 were trained individually on knee radiographs from the Osteoarthritis Initiative (OAI) dataset, followed by weighted ensemble evaluations to harness their unique strengths. Among the standalone models, ResNet50 attained the highest testing accuracy (68.4%) and demonstrated strong ordinal agreement (Quadratic Weighted Kappa = 0.83), although all single networks exhibited sensitivity to class imbalance and grades that were on the borderline. The ensemble configurations significantly enhanced robustness, with the ResNet50 + DarkNet19 ensemble reaching the highest accuracy of 70.47% (Quadratic Weighted Kappa (QWK) = 0.835), while the proposed three-model ensemble (ResNet50 + DenseNet169 + DarkNet19) yielded the most balanced and generalizable outcomes, achieving 69.99% test accuracy, QWK = 0.829, UAR = 0.668, Adjacent Accuracy = 0.932, and Macro AUC = 0.908. The ensemble also exhibited consistent performance on the auto-test set (accuracy = 69.3%) and obtained the best Macro PR-AUC = 0.940, indicating excellent precision-recall stability across all KL grades. In conclusion, the weighted CORN ensemble put forth effectively overcomes the drawbacks of manual grading and single-model CNNs, delivering a dependable, objective, and scalable solution for automated radiographic assessment of KOA severity in clinical settings.
Keywords:	Knee Osteoarthritis (KOA), Kellgren–Lawrence (KL) grading, Deep Learning, Ordinal Regression (CORN), Convolutional Neural Networks (CNNs), Ensemble Learning
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A DENSE RECURRENT NEURAL NETWORK WITH ATTENTION MODEL AND PRECISE FEATURE RETRIEVAL FOR CUSTOMER BEHAVIOUR PREDICTION TO ENHANCE MARKETING DECISIONS
Author:	B.SAKTHI, Dr.D.SUNDAR
Abstract:	Online platforms are essential for consumer interaction in the digital environment. In the current situation, it is critical to comprehend consumer behaviour in order to improve user experience, maximize marketing efforts, and improve company expansion. Nowadays, the majority of consumer behaviour models are constructed by applying machine learning and data mining techniques to real customer data. Forecasting customer behaviour is a difficult and unpredictable task. Therefore, sophisticated techniques and approaches are required to detect the client’s behaviour. Thus, a new customer behaviour prediction model is implemented in this work to enhance consumers’ experience in digital environments like online shopping. First, the customer behaviour data is collected from benchmark sites. Then, the collected data is utilized by the Restricted Boltzmann Machines (RBM) for retrieving useful features. These features are processed by the proposed Dense Recurrent Neural Network with Attention (DRNN-A) for predicting customer behaviours. This model is helpful for business organizations to increase their profits by showcasing the products that are frequently purchased by customers. Experimental evaluations are performed to check the prediction efficacy of the developed model.
Keywords:	Customer Behaviour Prediction; Marketing Decisions; Restricted Boltzmann Machines; Dense Recurrent Neural Network with Attention
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN INTELLIGENT HYBRID METAHEURISTIC FRAMEWORK FOR ADAPTIVE LOAD BALANCING USING GREY WOLF OPTIMIZATION AND CENTROID OPPOSITION-BASED LEARNING
Author:	P. RAJESH , K. MAHARAJAN , M. JAYALAKSHMI
Abstract:	The problem of task scheduling into the VMs of different capacities is one of the most important problems in the cloud computing domain. Bad scheduling leads to long makespan, unutilized servers, and violation of Service Level Agreements (SLAs). In cloud computing scheduling problem, Grey Wolf Optimization (GWO) is widely used but it suffers from premature convergence in high dimensional problems. In dynamic cloud environments, GWO gets trapped in local optima, which is one of its weaknesses that needs to be addressed. A novel algorithm called Centroid Opposition-Based Learning (COBL) is introduced, where opposite solutions are obtained by using the centroid of the current population in search space as an opposite agent rather than by taking the limits of the search space as an opposite agent. Using COBL keeps the swarm more diverse and avoids premature convergence and improves the exploitation of the solutions without compromising on the quality of results. To overcome the premature convergence and local optima trap problem of GWO, we propose Centroid Opposition-Based Learning Grey Wolf Optimization algorithm (COBL-GWO). We implement the algorithm in CloudSim and compare COBL-GWO with Improved GWO (IGWO), PSO, CPSO, GA and standard GWO algorithm for 10–50 tasks on 5 and 10 VMs. On 5 VMs it lowered the average makespan to 458.4 s against 573.6 s for GA, a 20.1% reduction, and raised average resource utilization to 0.854 against 0.422; on 10 VMs the makespan gap widened to 25.3%. But, as we report, there is a price: COBL-GWO takes about 38% more execution time and requires more memory than GA due to the opposition step that is performed during each iteration. Thus, COBL-GWO may be more suitable for the use cases where resource balancing and good resource utilization take precedence over scheduling overhead.
Keywords:	Cloud Computing, Load Balancing, Task Scheduling, Grey Wolf Optimization, Centroid Opposition-Based Learning, Metaheuristic Optimization, Resource Utilization; Energy-Efficient Computing, Sustainable Digital Infrastructure.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	UNSUPERVISED IMAGE CLUSTERING FOR IDENTIFYING INDUSTRIAL CUSTOMER POTENTIAL IN SIDOARJO USING MULTISPECTRAL SATELLITE FEATURES
Author:	I PUTU WAWAN SANJAYA PUTRA, IMAM YUADI, HENDRO MARGONO
Abstract:	This study investigates the use of unsupervised clustering techniques to identify potential industrial customer zones in Sidoarjo, East Java, using multispectral satellite imagery. The study employs Landsat 8 imagery, applying key spectral indices including Normalized Difference Vegetation Index (NDVI), Normalized Difference Built-up Index (NDBI), and Normalized Difference Water Index (NDWI) to classify land areas based on vegetation, built-up areas, and water bodies. K-means clustering was used to segment the region into distinct clusters, highlighting potential industrial areas. The industrial areas were further analyzed by calculating their spatial coverage using pixel count and spatial resolution. The estimated area of industrial zones in Sidoarjo was calculated to be 29,590 hectares, providing valuable insights for strategic energy distribution and industrial planning.
Keywords:	Satellite Imagery, Unsupervised Clustering, Industrial Customer Potential, Sidoarjo, ENVI, NDVI, NDBI, K-Means Clustering, Area Calculation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	A REAL-TIME IOMT-BASED WEARABLE SYSTEM FOR SIMULTANEOUS STRESS AND FATIGUE DETECTION IN UNIVERSITY STUDENTS USING PHYSIOLOGICAL SENSORS AND MACHINE LEARNING
Author:	ABDULLAH ALSHBATAT, RANDA QEISIEH, KHALID ALEMERIEN
Abstract:	The increasing lack of knowledge of the impact of stress and fatigue and how they can affect students' daily life has led to an increase in mental illnesses like anxiety, depression, and social anxiety and less productivity. Furthermore, they exacerbate physical illnesses especially hypertension, cardiovascular disorders, strokes, and obesity. In response to these growing concerns, this research proposes an Internet of Medical Things (IoMT)-based system to detect stress and fatigue in a real-time environment. This system collects physiological data via vital sensors. These data are used by machine learning (ML) algorithms to detect stress and fatigue and predict the students’ mental health at an early stage. Furthermore, the proposed system alerts students and suggests recommendations to reduce or prevent stress and fatigue levels. These alerts and recommendations assist students in maintaining their mental health before conditions deteriorate. This study proposes a real-time IoMT-based wearable system combining multiple physiological sensors and machine learning algorithms for the simultaneous detection of stress and fatigue in university students. Our work differs from prior work that considers single condition or limited physiological signals, in the integration of multiple biosignals, the system provides real-time personalized alerts and recommendations. Moreover, we validate the proposed prototype through experiments involving real students in realistic scenarios. The obtained results indicated strong correlations between HR and SpO2. RF and KNN outperformed other models in terms of accuracy scores.
Keywords:	Wearable Device, Stress Detection, Fatigue Prediction, Machine Learning, Biosignals, Sensors, IoMT.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	RIME-TB: A ROBUST AND INTERPRETABLE META-ENSEMBLE FRAMEWORK FOR ACCURATE AND TRANSPARENT TUBERCULOSIS DIAGNOSIS AND FUTURE-READY CLINICAL DECISION SUPPORT
Author:	DR. P. SAMBASIVA RAO, DR. AVANI ALLA, DR. MAKKAPATI SATYA SUKUMAR, DR.M.SRI DEVI SAMEERA, DR. BABU RAJENDRA PRASAD SINGOTHU, NARASIMHA RAO TIRUMALASETTI
Abstract:	The key issue is to ensure the optimal compliance with tuberculosis (TB) treatment regimens because non-compliance may result in the appearance of drug-resistant TB infections and an increase in the spread of the disease. This study deals with the shortcomings of the current methods used in adherence monitoring, which are more reactive and lack personalisation. The proposed study aims to use the predictive analytics framework to predict non-adherence risk in the case of TB patients using machine learning models. In this work, Robust Interpretable Meta-Ensemble for TB (RIME-TB) was designed and it will be used to combine various models which they will incorporate into multi-dimensional data that will include trends in behavior, socio-economic factors and indicators of healthcare accessibility. Data preprocessing activities such as normalization, missing value replacement, and Synthetic Minority Oversampling Technique (SMOTE) will be used to address the imbalance in the classes. Recursive Feature Elimination (RFE) will be used to select the most predictive variables. Such models as Decision Tree, Random Forest, XGBoost, and Support Vector Machines will be compared and evaluated on the basis of such metrics as accuracy, precision, recall, and the F1-score. Improvements in results are achieved and arrive at greatest efficiency of 96.8 percent in forecasting capacity compared to predictive capacity than baseline adherence-monitoring techniques. The RIME-TB model also performs better than regular classifiers, being more robust and providing a high degree of interpretability. This study actively detecting and eliminating adherence risk and enhances a better treatment outcome, prevent drug resistance, and a global control of TB.
Keywords:	CNN, Feature Selection, Grad-CAM, Meta Ensemble Learning Models, RFE, RIME-TB
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text

Title:	AN ENHANCED BLOWFISH ALGORITHM WITH DYNAMIC S-BOX EXCHANGE AND MERGE PATTERNS
Author:	Dr.S. SWEETLIN SUSILABAI, Dr. J. JEBAMALAR TAMILSELVI, Dr. K. SUTHA, Dr SURYA SUSAN THOMAS, Dr. KAYATHRI K
Abstract:	The rapid expansion of digital data transmission and the rise in cyber threats have made information security a crucial component of the modern communication systems. Sensitive data integrity, confidentiality, and authenticity are all greatly enhanced by encryption techniques. The Blowfish algorithm is a popular symmetric key cryptography algorithm due to its ease of use, adaptability, and computational effectiveness. However, the majority of current Blowfish-based methods rely on static S-box structures, which lessen algorithmic unpredictability and increase the algorithm's susceptibility to sophisticated cryptanalytic assaults like differential and linear cryptanalysis. Few studies have concentrated on dynamically altering S-box exchange techniques throughout both key allocation and encryption phases while maintaining computing efficiency, despite the fact that numerous improvements have been suggested in the literature.The primary objective of this research is to use the proposed - Exchange and Merge Patterns (EMPAT) method to create an improved Blowfish encryption model. To increase confusion, diffusion, randomness, and resistance to cryptanalytic attacks, the study incorporates dynamic S-box exchange operations and inter-bit shifting methods during encryption rounds and key scheduling.This paper proposes an improved Blowfish method based on the Exchange and Merge Patterns (EMPAT) approach to close this research gap. The proposed approach increases non-linearity, diffusion, and key dependency during encryption by introducing dynamic S-box exchange patterns and inter-bit shifting operations utilizing four different exchange combinations. The recommended EMPAT approach continuously rearranges S-box operations, which improves unpredictability and resilience against cryptanalytic assaults without significantly increasing computing overhead, in contrast to traditional Blowfish implementations.By offering a dynamic S-box exchange framework that improves cryptographic robustness while preserving the speed and efficiency features of the original Blowfish encryption, the study adds new information. The results verify that the EMPAT method provides a smaller and more secure encryption mechanism appropriate for modern secure communication applications.
Keywords:	Cryptography, Cipher Text, Blowfish Encryption algorithm and Data Security.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 15^th July 2026 -- Vol. 104. No. 13-- 2026
Full Text