		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
October 2026 | Vol. 104 No.20

Title:	TEMPORAL TAG–GUIDED MULTI-TASK GENERATIVE TRANSFORMER FOR HORIZON-AWARE TWEET-DRIVEN STOCK DIRECTION FORECASTING
Author:	S. PUNITHA , G.K. ARUN
Abstract:	Understanding the influence of social sentiment on stock market behavior has become essential in financial forecasting. Tweets, rich in emotion and immediacy, often reflect public reactions to market events before traditional indicators respond. This work introduces a temporal tag–guided, multi-task text generation framework that interprets tweet sentiment for stock movement prediction. By embedding financial time context directly into each input, the model learns to align language with specific market phases. The architecture simultaneously decodes sentiment classification, short- and long-term price direction, and event-driven rationale in a structured output format. A text-to-text transformer approach ensures adaptability across tasks without requiring separate models. The integration of temporal understanding with unified decoding enables more accurate, interpretable, and horizon-aware forecasts from unstructured tweet streams. This design bridges the gap between informal emotional language and structured financial logic, supporting intelligent systems that respond to the rhythm of market sentiment in real time.
Keywords:	Financial Sentiment Analysis - Tweet-Based Stock Prediction - Temporal Tagging - Multi-Task Text Generation - Text-to-Text Transformer - Horizon-Aware Market Forecasting
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A FINANCE-INSPIRED EDGE AI FRAMEWORK FOR COST-EFFICIENT AND HIGH-THROUGHPUT URBAN TRAFFIC MANAGEMENT
Author:	NAGA CHARAN NANDIGAMA , YOGITA GUPTA, RANGANATHAM GANGINENI , Y. LAKSHMI , SREE LAKSHMI MOORTHYGARI , MAHADEVAN M
Abstract:	Urban traffic management in Indian megacities faces dual challenges: vehicular density exceeding 17 million daily vehicles across six major metros and severe computational constraints at resource-limited edge nodes. Current deep reinforcement learning (DRL) methods can provide a relatively ideal performance in signal optimization, but are computationally expensive and do not consider the cost of traffic delay as a tangible economic risk. Researcher propose the edge AI framework FinEdge-Traffic, which uses quantitative finance tools to reshape the problem of signal phase allocation into a multi-asset portfolio problem, with the length of green-phase durations playing the role of the asset weights, lengths of the queues as risk measure proxies, and the maximization of the throughput as the return. The framework consists of a YOLOv9-nano detector (12 ms inference), a 4-layer temporal Transformer encoder capturing 30-minute traffic memory, an Optimal Execution Engine, which is inspired by the finance world and optimizes over mean-CVaR objectives, and a centralized-training decentralized-execution (CTDE) MARL coordination layer. Compared to the best DRL baseline, FinEdge-Traffic achieves 12 ms latency, 47.2% higher throughput, and 46.7% lower average delay (38.6 s vs. 72.4 s) at 2.1× the cost of the fixed-time controller (9.7M annotated frames over 18 months at 550 signalized intersections across Delhi, Mumbai, Bangalore, Chennai, Hyderabad and Kolkata). FinEdge-Traffic is superior to all three of its strong baselines on the five metrics (p < 0.001). The finance engine gets the most benefit from ablations, while the criticality of the emergency vehicle pre-emption robustness provides the basis for CVaR modelling.
Keywords:	Deep Reinforcement Learning, Traffic Delay, Transformer Encoder, Optimal Execution Engine, FinEdge
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	COMMUNICATION-EFFICIENT DECENTRALISED MULTI-AGENT REINFORCEMENT LEARNING FOR COMMUNITY-LEVEL ENERGY MANAGEMENT USING TOPOLOGY-AWARE ADAPTIVE STATE COMPRESSION
Author:	DR A BRAHMANANDA REDDY, MOSURI SANTHOSH KUMAR, D. CHAKRADHAR, DR K. SURYANARAYANA, MS Y. LAKSHMI, DR GANTA JACOB VICTOR, ELANGOVAN MUNIYANDY, DR KISHORE KUMAR KAMARAJUGADDA
Abstract:	Decentralised energy management in residential communities is challenging. Agents need to organise the use of distributed energy resources while also following the rules of the shared network. The topology of a distribution network can be shown by graph-based multi-agent reinforcement learning (MARL), but big state representations may make it more difficult for communities to communicate and compute as they grow. To minimise shared state information while maintaining characteristics important for energy-management choices, this study suggests the Topology-Aware Graph Transformer with Adaptive State Compression (TAGT-ASC) framework. The technique combines proximal policy optimisation (PPO)-based multi-agent learning with a Graph Transformer encoder and a learnt gating mechanism. Incorporating the IEEE 33-bus distribution-network benchmark and real-world OpenCEM microgrid observations, the assessment framework is complete. With an operational cost of ₹143.60/day, 317.90 kWh/day of network losses, 0.0069 p.u. of mean voltage deviation, 184.50 kWh/day of PV curtailment, and 96.40 kWh/day of battery throughput, TAGT-ASC outperforms GAT-MAPPO, Graph-MASAC, and Graph-TD3. There is a 62.5% drop in communication overhead when adaptive compression is used compared to the uncompressed setup. There is also a drop in median inference latency from 4.80ms to 3.40ms and a drop in peak memory from 420 MB to 325 MB. These findings suggest that adaptive state compression may be used to strike a balance between computational efficiency and energy-management performance. To demonstrate resilience and scalability in the face of disruptions and growing agent counts, more quantified tests are needed.
Keywords:	Decentralised Multi-Agent Reinforcement Learning; Topology-Aware Graph Transformer; Adaptive State Compression; Community-Level Energy Management; Distributed Energy Resources; Microgrid Optimisation; Communication-Efficient Control.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	PRIVACY PRESERVING FEDERATED ANALYTICS OF CAPSULE ENDOSCOPY FOR GASTROINTESTINAL DISEASE DETECTION
Author:	Dr. SIVA CHAKRA AVINASH BIKKINA, DR. V. PRAVEEN KUMAR, DR S V N SREENIVASU, DR. SESHA BHARGAVI VELAGALETI, MASARATH SABA, VAMSIDHAR TALASILA, ELANGOVAN MUNIYANDY, MR.SAI PRASANTH KANUPARTHY
Abstract:	Capsule endoscopy enables the detection of gastrointestinal abnormalities, but privacy constraints, heterogeneous data distributions, and communication overhead remain major challenges for collaborative artificial intelligence in healthcare. This study presents FedCapsule, a privacy-preserving federated learning framework for gastrointestinal disease detection that integrates parameter-efficient foundation-model adaptation, patient-level differential privacy, secure aggregation, and heterogeneity-aware optimization. The framework was tested with a total of 37,607 training frames and 16,132 validation frames from three public capsule endoscopy data sources along with an external test set of 4,385 annotated frames from the All India Institute of Medical Sciences (AIIMS), New Delhi, India. The eight simulated clients are used for training, and a non-independent and identically distributed partition based on the Dirichlet distribution technique is used to distribute the training data. FedCapsule employs a frozen DINOv2 ViT-S/14 backbone adapted through low-rank adaptation (LoRA), logit-adjusted cross-entropy, prototype alignment, differentially private federated analytics, and drift-aware aggregation. At a privacy budget of ε = 8, the reported results show a macro-F1 score of 66.8% and a macro-AUROC of 0.936 on the external test set. The framework uses 151,306 trainable parameters and an estimated communication cost of 0.62 MB per round. These results suggest that parameter-efficient adaptation combined with privacy-preserving optimization may support gastrointestinal image classification under simulated federated conditions. The proposed framework needs to be further validated in independent clinical institutions that they are robust, generalizable and clinically applicable.
Keywords:	Gastrointestinal, Capsule Endoscopy, Patient, Disease Detection, FedCapsule
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	AN EXPLAINABLE CLUSTER-GUIDED CNN–GNN FRAMEWORK FOR RELATIONAL LEARNING AND CONGENITAL HEART DISEASE CLASSIFICATION FROM ECHOCARDIOGRAPHIC IMAGES
Author:	ABHIGNA JALLI, N. SRINIVASU NULAKA
Abstract:	The Congenital Heart Disease (CHD) is still one of the most significant reasons for infant morbidity and mortality rates, hence the need for an early and accurate diagnosis that would lead to proper clinical treatment. Although recent advancements in artificial intelligence techniques have significantly contributed to the automatic analysis of echocardiography images, the current methods still face the problem of failing to effectively model the complex anatomical relationships between cardiac structures while being difficult to interpret and less applicable in clinical practice. In light of the stated limitations of current AI technologies in this area, this research suggests an Explainable Cluster-Guided Convolutional Neural Network–Graph Neural Network (CNN–GNN) architecture for automated CHD classification using echocardiographic images. The suggested architecture uses deep feature extraction followed by cluster-guided graph construction to construct a graph with anatomically meaningful cardiac region representation, followed by graph-based relational learning through a Graph Neural Network (GNN) for improved disease classification. At the same time, SHAP and Grad-CAM are used as complementary explainability methods to further improve model transparency, interpretability of the results, and support clinically interpretable decision-making. According to experimental results obtained for the CAMUS echocardiographic dataset, the proposed method was able to classify the images with an accuracy of 98.7%, precision of 98.5%, recall of 98.3%, F1 score of 98.4%, specificity of 99.1% and AUC of 99.1%, thus outperforming not only conventional deep learning but also the current graph-based baseline models. In addition to superior classification performance, the suggested approach is also able to effectively and accurately highlight diagnostically relevant cardiac regions, which improves the reliability, trustworthiness, and interpretability of its automated predictions. Overall, it can be said that the proposed method shows the effectiveness of combining graph representation learning guided by clustering with explainable artificial intelligence in improving diagnostic accuracy as well as making the model more transparent.
Keywords:	Congenital Heart Disease, Echocardiography, Convolutional Neural Network, Graph Neural Network, K-Means Clustering, Explainable Artificial Intelligence, Grad-CAM, SHAP, Medical Image Analysis.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	AN ADAPTIVE MULTI-SOURCE SPATIO-TEMPORAL GRAPH TRANSFORMER FRAMEWORK WITH CONTEXT-AWARE FEATURE FUSION FOR AGRICULTURAL YIELD FORECASTING AND FOOD SECURITY
Author:	JAGADISH A N, DR. AKKU MADHUSUDHAN, MRS. KARUNA VALISAMMAGARI, DR. POTNURI SURIBABU, RAMYA R, VAKAIMALAR ELAMARAN
Abstract:	Food security planning accurate forecasts of crop yields are critical. Yet, combining heterogeneous agricultural information and modelling complex spatial and temporal dependencies are still great challenges. The current study suggests a Context-Aware Adaptive Spatio-Temporal Graph Transformer (CA-ASTGT) that enhances crop-yield prediction by adaptively combining multi-source information, integrating spatio-temporal information, and fusing features with context. The framework is based on a mixture of source-specific encoders and learned source weighting, a graph-based spatial message passing, a temporal Transformer encoder, and a context-conditioned fusion gate. A regression head predicts crop yield, and a relative yield-shortfall indicator is used in the production-risk assessment. The model is tested against the CY-Bench regional benchmark and is field-level YieldSAT independent test. The comparison is based on the root mean squared error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and normalised RMSE (nRMSE) with the performance of ConvLSTM, spatio-temporal graph convolutional networks (STGCN), and a spatio-temporal graph Transformer (STGT). When compared with baselines, CA-ASTGT performs best on CY-Bench with an RMSE of 0.61 t/ha, an MAE of 0.44 t/ha, an R² of 0.89, and an nRMSE of 13.7%. On YieldSAT, it has an RMSE of 0.72 t/ha and an R² of 0.84. Ablation results show that all major components are contributing to predictive performance. The results here show the potential of the CA-ASTGT for crop-yield forecasting and for screening for production shortfalls, but more testing is needed to determine the generalisability and operational efficiency of the CA-ASTGT.
Keywords:	Agricultural Yield Forecasting; Spatio-Temporal Graph Transformer; Multi-Source Data Integration; Context-Aware Feature Fusion; Food Security; Production-Risk Assessment.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A GENERALIZED HYBRID CNN TRANSFORMER FRAMEWORK WITH BIOINSPIRED AOHHO TUNING FOR HETEROGENEOUS MEDICAL IMAGE CLASSIFICATION APER TITLE HERE
Author:	ANJANEYULU GUDLA,G MAHENDER, DR KATAM NAGA LAKSHMAN, DR VOOTLA SRIDHAR, PRAVEENA NUTHAKKI
Abstract:	Clinical diagnosis and treatment planning depend on Magnetic Resonance Imaging (MRI) is an accurate method of classifying brain tumors. However, conventional convolutional neural networks often have difficulty dealing with complex tumor patterns, interclass similarity, and suboptimal parameter settings despite the remarkable Intensification of deep learning techniques in the field of medical image analysis. Using an optimized multi-feature learning strategy and convolutional neural networks, this study develops a hybrid deep learning framework that addresses these challenges. As part of the proposed architecture, spatial representations are extracted using convolutional layers, while far-range contextual dependencies across MRI slices are captured by transformer modules. Moreover, a multi-feature fusion component improves discriminative feature learning by combining spatial and frequency-domain representations. The hyperparameters learning rate, convolutional kernel size, and attention weights are automatically tuned using a hybrid optimization strategy based on Aquila optimizer and Harris Hawks optimization. By combining adaptive feature fusion and evolutionary optimization, classification robustness is improved while overfitting is reduced. A comparison of the proposed framework with conventional CNN-based models and recent hybrid architectures based on benchmark MRI datasets shows superior performance. During training, the accuracy, sensitivity, and F1-score of classifications improved significantly. Moreover, the framework proved its versatility by achieving 97.13% training convergence and 84.50% test accuracy on the PathMNIST histological dataset. A hybrid metaheuristic optimization approach in conjunction with transformer-based attention and multi-feature representation learning provides an effective approach for reliably detecting brain tumors in medical imaging
Keywords:	Brain Tumor Classification, Magnetic Resonance Imaging (MRI), Convolutional Neural Networks (CNN), Vision Transformer, Cross-Dataset Validation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A FLEXIBLE PAPER-BASED WEARABLE SYSTEM FOR CONTINUOUS SWEAT BIOMARKER AND ECG MONITORING: QUALITY-AWARE DELAY-COMPENSATED MULTIMODAL FUSION
Author:	DR. V. V. R. MAHESWARA RAO, MRS. DEEPTHI KAMIDI, DR. SUDHEER KUMAR NAGOTHU, NALLA AKHILA4, MOLIGI SANGEETHA, DR GANTA JACOB , ELANGOVAN MUNIYANDY, DR. RAVINDRANADH JAMMALAMADUGU
Abstract:	Wearable patches on paper are a low-cost way to perform continuous sweat analysis in addition to electrocardiography (ECG); however, integrated systems will have to deal with the challenges of intermittent sweat sampling, delays in sweat collection, fluidic interference, and motion induced ECG interference. In this study, we propose a processing framework called Quality-Aware Delay-Compensated Multimodal Fusion (QADMF) to achieve the joint estimation of ECG signal quality, sweat-sample validity, and channel-specific sensor delay using a cross-modal attention mechanism that is computed based on the channel's reliability. It was tested on 48 half-hour recordings from the MIT-BIH Arrhythmia Database contaminated with MIT-BIH Noise Stress Test recordings, externally tested on 998 PhysioNet/CinC 2011 recordings, and integrated with a physics-based paper-fluidic sweat transport simulator to produce four channels of biomarker data, two auxiliary fluidic channels, and ground truth delays and validity labels. The accuracy of ECG quality classification was 95.2% ± 0.3% for QADMF across five independent runs, which was significantly higher than five baselines (p = 0.002 record-level test), with a mean absolute error of 1.78 s ± 0.09. The accuracy of invalid-sample detection was 94.1% ± 0.4% for QADMF across five independent runs, which was significantly higher than five baselines (p = 0.002 record-level test). The reconstruction error of biomarkers increased by only 2.2 percentage points with simulated sensor drift of 10% per hour, while it was 9.5 for the best baseline. The results demonstrate that explicit quality and delay modelling increase the reliability of the integrated paper-based monitoring data significantly.
Keywords:	Paper-Based Wearable Sensors, Sweat Biomarkers, Electrocardiography, Multimodal Fusion, Signal Quality Assessment, Delay Compensation, Sensor Drift
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	FLOWPULSE-HSAGNET: MICROFLOW-GRANULARITY SPECTRAL ANALYSIS FOR EARLY-STAGE VOLUMETRIC AND APPLICATION-LAYER DDOS SURGE IDENTIFICATION
Author:	GONE VIVEK SAI, A. MAHENDAR, MAHESWARI K, NUTHANAKANTI BHASKAR, K. SHILPA, K. RANJITH REDDY
Abstract:	Flow-level aggregate detectors may recognize distributed denial-of-service (DDoS) attacks only after traffic volume rises. FlowPulse-HSAGNet detects repeating traffic patterns that precede surges by dividing traffic into 100 ms micro-windows and constructing victim-specific rate signals. A hybrid spectral-attention network and gradient-boosted ensemble combine traffic magnitude, spectral descriptors, and burst growth. Training, validation, testing, and figure generation run on one consumer-grade graphics processing unit. Evaluation across four public datasets containing 9,703,976 flows achieved 99.99000% pooled accuracy on 1,940,793 held-out flows, with 194 errors and 99.94466% macro accuracy. All sixteen predefined criteria passed, fully addressing research gaps G1–G5 within the evaluated scope. Microflow granularity improved F1 by 1.4238 percentage points and early detection by 12.5%; mean early-warning lead time was 5.2994 s. Unified temporal-spectral scoring improved surge detection by 12% and F1 by 1.25 percentage points. Target-adapted transfer retained 96.2552% of within-dataset F1, while inference latency was 0.0083 s. Purged contiguous-block splits protected series-level evaluation from overlapping-window leakage. Spectral features provided no steady-state classification benefit, zero-shot cross-corpus transfer required target adaptation, and the accuracy improvement over Random Forest was not significant at the 5% level.
Keywords:	DDoS Detection, Microflow Analysis, Spectral Analysis, Early Warning, Attention Networks, Gradient Boosting, GPU Acceleration, Reproducible Evaluation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	ONCOGRAPH-RLNET: BIOLOGICALLY CONSTRAINED GRAPH REINFORCEMENT LEARNING FOR REPRODUCIBLE BIOMARKER DISCOVERY, RADIOTHERAPY PLAN ERROR DETECTION AND RISK-AWARE TREATMENT OPTIMISATION
Author:	CH SAI LAKSHMI, DR. CH. VENKATA RAMANA REDDY, DR L RAGHU KUMAR
Abstract:	Background: Treatment decisions in oncology depend on both tumour biology and the quality of the radiotherapy plan that delivers treatment, yet these are modelled separately, without biological network structure or sequential, risk-aware decision-making. Methods: OncoGraph-RLNet combines (i) gated multi-omics fusion tolerant to missing modalities, (ii) a graph attention network whose logits are biased by a gene–pathway knowledge graph, (iii) LP-Stability, a resampling-based score that selects reproducible gene signatures, (iv) a patient-conditioned CNN–Transformer for plan-error detection and dose prediction, and (v) a conservative offline distributional RL agent optimising a mean–CVaR objective under a toxicity budget. Evaluation uses TCGA, TCIA and OpenKBP with leakage-safe splits. Results: OncoGraph-RLNet reached 96.2 ± 0.8 % stratification accuracy, a C-index of 0.892 and a signature stability index of 0.872 (best baseline, GAT: 93.4 %). It detected injected plan errors with AUROC 0.981 and lowered the dose score to 0.087 Gy (best baseline 0.118 Gy). The learned policy reduced cost violations to 4.2 % (7.4 % for quantile CQL) at 97.1 % guideline concordance. Removing LP-Stability, the graph branch, the biological prior or CVaR each degraded the corresponding metric. Conclusions: Coupling biologically constrained graph learning, plan-error detection and risk-aware offline RL yields more stable signatures, more sensitive plan screening and safer policies than single-purpose models; clinical use requires prospective validation.
Keywords:	Precision Oncology, Multi-Omics, Graph Attention Network, Offline Reinforcement Learning, Radiotherapy Quality Assurance
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	ARTIFICIAL INTELLIGENCE IN DETECTION OF DISINFORMATION IN THE DIGITAL ENVIRONMENT IN THE CONTEXT OF LEGAL REGULATION OF FREEDOM OF SPEECH
Author:	ANNA KOVALOVA, IVO SVOBODA, OKSANA YALOVA, VALERII NIKITCHENKO, YURII VYKHODETS
Abstract:	Disinformation is becoming a powerful tool for influencing public opinion and political processes In the modern digital environment, posing a threat to democratic institutions. The use of artificial intelligence (AI) to automatically detect fake news is gaining particular importance, while posing a challenge to ensure freedom of speech and compliance with the legal proportionality principles. The aim of this study was to develop and investigate the effectiveness of a multimodal model for disinformation detection that combines technical methods of AI with legally oriented mechanisms of transparency and explainability. The proposed model was based on a combination of transformer architectures for text data, computer vision methods for images, and graph neural networks for modelling content distribution. Progressive multimodal fusion, contrastive prior learning, and explainability approaches were used. Empirical testing was carried out on several representative datasets (LIAR, FakeNewsNet, COVID-19, multimodal datasets). The proposed architecture achieved high accuracy rates (Acc ≈ 0.938, F1 ≈ 0.925, AUC ≈ 0.965), outperforming both single-modal (BERT, GNN) and bimodal solutions. Cross-domain experiments demonstrated the generalization ability of the multimodal model, although the results indicated the difficulty of transferring models between different information environments. The research hypothesis was confirmed: combining multimodal models with graph structures and legally oriented design significantly increases the efficiency and transparency of disinformation detection systems. The proposed approach helps to reduce the risk of excessive content blocking and provides a balance between combating disinformation and protecting freedom of speech. Further research should focus on developing more representative multilingual data sets, as well as optimizing the proposed model for real-time operation and integrating algorithms into regulatory practices of digital platforms, taking into account legal and social aspects.
Keywords:	Disinformation, Artificial Intelligence, Multimodal Models, Freedom of Speech, Legal Regulation, Explainability, Graph Neural Networks, Governance, Human Rights, Legal System.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	ADAPTIVE CYBER-RESILIENCE FRAMEWORK FOR INTELLIGENT INTRUSION DETECTION IN 6G-ENABLED INTERNET OF MEDICAL THINGS NETWORKS
Author:	BASIL BABY K , Dr. A. NITHYA RANI
Abstract:	Network security has become increasingly critical in 6G-enabled Internet of Medical Things (IoMT) networks, where massive device connectivity, heterogeneous communication, and low-latency healthcare services expand the potential attack surface. The resource-constrained and sensitive nature of IoMT requires intrusion detection mechanisms that are adaptive, resilient, and computationally efficient. Network Intrusion Detection Systems (NIDS) provide an essential layer of protection by monitoring network traffic and identifying anomalous activities and potential attacks. This study proposes an Adaptive Cyber-Resilience Framework that integrates bio-inspired optimization with adaptive anomaly detection, dynamic rule adjustment, automated response, and continuous feedback to strengthen NIDS performance. Experimental evaluation using a network intrusion dataset demonstrates that the proposed framework achieves 77.512% detection accuracy, 77.941% F-measure, 79.698% precision, and 76.259% recall, outperforming DI-NIDS and DMAN. The framework also achieves an attack detection rate of 76.259% and a missed detection rate of 23.741%, demonstrating its potential to provide adaptive and resilient intrusion detection for next-generation IoMT networks.
Keywords:	Network Intrusion Detection Systems, 6G, Internet of Medical Things, Intrusion Detection, Cybersecurity in Healthcare, Oranguttan Optimization.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	QUANTUM CONTRASTIVE LEARNING FOR RARE DISEASE DIAGNOSIS USING MULTIMODAL ELECTRONIC HEALTH RECORDS
Author:	MAYA B. DHONE, JILLELLA VENKATESWARA RAO, MADDHI SUNITHA, V. LAKSHMI SAILAJA, M. HARSHAVARDHAN, ANNAMARAJU THANUJA, JOSE. P, TALE NAGAMANI, DARA RAJU
Abstract:	Rare disease diagnosis from electronic health records (EHRs) remains difficult due to the limited number of disease samples, heterogeneous clinical information, imbalanced datasets, and missing modalities. This paper proposed a novel method of Quantum Contrastive Learning for Rare Disease Diagnosis (QCL-RDD), which combined demographic data, laboratory results, clinical events, and clinical notes of the patients. The framework employed modality-specific feature extraction, patient-level contrastive learning, attention-based multimodal fusion, and an eight-qubit variational quantum circuit for compact feature learning. The proposed model resulted in an accuracy of 93.26%, balanced accuracy of 90.18%, macro-F1 of 89.74%, and Area Under the Receiver Operating Characteristic curve (AUROC) of 95.12%, outperforming all classical and multimodal methods used for comparison. The results of the ablation study demonstrated that contrastive learning and quantum feature processing contributed to the diagnostic performance of the model, while maintaining an accuracy of 86.42% even with the availability of only two modalities. The results of the experiment confirmed that QCL-RDD offered a robust and reliable approach for handling heterogeneous and incomplete EHR data and enhanced rare disease classification. The proposed framework could support reliable and efficient AI-assisted rare disease screening in clinical decision-support systems.
Keywords:	Quantum Contrastive Learning; Rare Disease Diagnosis; Electronic Health Records; Multimodal Learning; Variational Quantum Circuit; Clinical Decision Support.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	LONGITUDINAL CRACK SEVERITY CLASSIFICATION BASED ON CONVNEXT CRACK WIDTH ESTIMATION USING TRANSFER LEARNING
Author:	RATU NURMALIKA, SARIFUDDIN MADENDA, NAHDALINA, RODIAH
Abstract:	Artificial Intelligence (AI) has become an important technology for improving road infrastructure management, particularly in the assessment of pavement damage. Longitudinal cracks are among the most common forms of road deterioration and, if not identified and treated promptly, can reduce transportation safety and increase maintenance costs. Conventional inspection methods rely on manual visual observations, which are often time-consuming, labor-intensive, and prone to subjective errors. This study proposes a transfer learning framework for longitudinal crack severity classification based on crack width estimation using a modified ConvNeXt architecture. The proposed framework consists of dataset preparation, image preprocessing, data augmentation, dataset partitioning, crack width estimation, severity classification, and performance evaluation stages. The ConvNeXt model is modified by replacing the original classification head with a regression layer to predict crack width as a continuous value. The predicted crack width is subsequently mapped into three severity levels, namely light, medium, and heavy cracks. To improve adaptation to the target dataset while preserving pretrained knowledge, a fine-tuning strategy with a low learning rate and Cosine Learning Rate Decay scheduling is employed. Experimental results demonstrate that the proposed model achieves the best performance under a 70:20:10 train-validation-test ratio, obtaining a Test Loss of 625.4404 and a Test R² score of 0.5050. Furthermore, the average prediction error between manual crack width measurements and model predictions was below 1.1% across all severity categories, indicating strong agreement between estimated and reference measurements. The integration of transfer learning, adaptive learning rate scheduling, and crack width-based severity classification provides an effective and reliable approach for intelligent road condition assessment and decision support in road maintenance management.
Keywords:	ConvXNet, Fine Tuning, Learning Rate, Longitudinal Crack, Pipeline
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	MALWAREXAI-SECNET: AN EXPLAINABLE SEVERITY-AWARE MULTI-TASK DEEP LEARNING FRAMEWORK FOR INTELLIGENT MALWARE DETECTION AND SECURITY ANALYTICS
Author:	ANURADHA ANUMOLU, ARUN SINGH CHOUHAN, S SRINIVASA RAO
Abstract:	Modern malware is becoming increasingly obfuscated, polymorphic and behaved, posing challenges for reliable malware detection and security analysis. Current deep learning methods are mostly task-specific and lack the ability to jointly address the following problems in a single framework: malware detection, malware family classification, identification of security-related behavioral capabilities, assessment of the severity of the threat, uncertainty prediction, explanation and security decision support. To overcome these drawbacks, the authors have introduced a novel, explainable, severity-aware multi-task deep learning framework called MalwareXAI-SecNet for intelligent malware detection and security analytics.For this purpose, authors have presented an innovative explainable and severity-aware multi-task deep learning framework, dubbed MalwareXAI-SecNet, to overcome these drawbacks for intelligent malware detection and security analytics. The framework proposes a dataset-adaptive feature encoding that involves a gated residual multilayer perceptron (grid-MLP), a multi-scale one-dimensional convolutional neural network (mdCNN), and a transformer encoder, followed by adaptive gated-attention fusion. The missing label aware multi-task optimisation allows for learning from diverse datasets and reducing the conflicting predictions across the different tasks (detection, family classification, behavioral capabilities, and ordinal threat severity). The uncertainty estimation, probability calibration, and interpretable predictions are facilitated by Monte Carlo dropout, temperature scaling, SHAP, integrated gradients, and attention-based evidence. A detection F1-score of 99.46%, 98.69% on CIC-MalMem-2022 and EMBER2024, respectively, and family-classification Macro-F1 score of 97.31% on BIG 2015 are reported from experiments performed on CIC-MalMem-2022, EMBER2024 and BIG 2015, respectively. The framework also provides a family-classification Macro-F1 score of 98.51% and a behavioural-capability Macro-F1 score of 96.31% and a threat-severity Macro-F1 score of 96.72% on CIC-MalMem-2022. The results show the potential of the proposed framework in providing a comprehensive malware intelligence from predictive performance, behavioural interpretation, severity-based risk prioritisation and confidence-aware security recommendations. The framework can help security analysts prioritise suspicious files and direct files in doubt or high-risk cases for further analysis.
Keywords:	Malware Detection, Multi-Task Deep Learning, Explainable Artificial Intelligence, Threat Severity Prediction, Security Analytics
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A FREQUENCY-AWARE TRANSFORMER FRAMEWORK FOR VEHICLE SPEED VIOLATION DETECTION WITH ROBUST LICENCE PLATE RECOGNITION
Author:	M AYYAPPA CHAKRAVARTHI, KAMATAM NAGESWARA REDDY, PASUPULETI NAGALAKSHMIPRIYANKA, NAGA MALLIKHARJUNARAO BILLA, VEJENDLA BALA NAGA AJAY, T SUBBA REDDY, MEENA ABARNA K.T, NIMMAGADDA CHANDRA SEKHAR
Abstract:	This paper presents a frequency-aware transformer-based framework for vehicle speed violation detection and automatic licence plate recognition using traffic surveillance video. The proposed framework integrates vehicle detection, multi-object tracking, speed estimation, violation classification, and licence plate recognition into a unified pipeline. Spatial and frequency-domain feature representations are intended to improve vehicle detection under challenging illumination and image-quality conditions, while transformer-based feature fusion models contextual relationships in traffic scenes. Vehicle speed is estimated from tracked trajectories using camera calibration and temporal displacement, and vehicles exceeding the applicable speed limit are processed by a deep learning-based licence plate recognition module. The manuscript reports an experimental evaluation using traffic surveillance videos, vehicle image frames, and licence plate images. The reported results include 97% vehicle detection accuracy, 95% speed-estimation accuracy, 96% licence plate recognition accuracy, and a 3% false-violation rate. The reported classification results also include a precision of 96.8%, recall of 97.8%, and F1-score of 97.3%. These findings indicate the potential of integrating vehicle detection, trajectory-based speed estimation, and licence plate recognition into a unified traffic-monitoring system. Further validation using clearly documented datasets, calibrated speed measurements, and reproducible comparative experiments is necessary to establish the framework's generalizability and real-world effectiveness.
Keywords:	Vehicle Speed Violation Detection, Frequency-Aware Deep Learning, Transformer-Based Computer Vision, Automatic License Plate Recognition, Intelligent Transportation Systems.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	CTWGAN-GP: SEQUENTIAL INTEGRATION OF CTGAN AND WCGAN-GP FOR PRACTICAL MACHINE LEARNING UTILITY OF SYNTHETIC TABULAR DATA
Author:	RAWAN F. ALSHEHRI, SABAH M. ALZAHRANI
Abstract:	Generating high-quality synthetic tabular data is essential for machine learning applications in data-scarce and privacy-sensitive domains. This study presents CTWGAN-GP, a sequential generative framework that integrates Conditional Tabular GAN (CTGAN) and Wasserstein Conditional GAN with Gradient Penalty (WCGAN-GP) to examine whether multi-stage model integration can improve the quality and utility of synthetic tabular data. Existing approaches typically rely on a single GAN model, overlooking the potential benefits of sequential integration. Experiments were conducted using two benchmark datasets: the Adult dataset, which includes mixed data types, and the Credit Card Fraud dataset, which contains only numerical features. Two sequential configurations were evaluated using machine learning utility metrics, including F1-score and ROC AUC. Results show that CTWGAN-GP improves performance on the Adult dataset, particularly when CTGAN was applied after WCGAN-GP, demonstrating better downstream classification outcomes. Specifically, CTWGAN-GP achieves an F1-score of 0.6454, outperforming CTGAN (0.6067) by approximately 6.4% and WCGAN-GP (0.4195) by approximately 54%. Notably, the strongest sequential configuration achieved its best performance without requiring preprocessing or postprocessing, suggesting that its effectiveness is attributable to the model design rather than external corrections. In contrast, for the numerical-only Credit dataset, WCGAN-GP alone achieves the highest performance, while incorporating CTGAN—whether before or after WCGAN-GP—reduces downstream classification performance. These findings highlight that the effectiveness of sequential GAN frameworks depends strongly on dataset characteristics and that no single architecture consistently generalizes across datasets.
Keywords:	Synthetic Tabular Data; CTGAN; WCGAN-GP; Sequential GAN; Machine Learning Applications.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	ECDF-NET: AN EXPLAINABLE CONTINUAL DEEPFAKE FORENSIC NETWORK WITH ADAPTIVE SPATIAL-FREQUENCY FUSION
Author:	G.V. ASHRITHA, Dr. ANNALAKSHMI GOVINDARAJ
Abstract:	The rapid evolution of DeepFake and diffusion-based face generation methods has made the reliable forensic analysis of facial media increasingly difficult, because detectors trained on known manipulations degrade under cross-domain distribution shifts, common media transformations, and newly emerging forgery techniques. Existing studies address spatial-frequency learning, continual adaptation, forgery localization, and explainability largely in isolation, and rarely within a single model. This paper addresses this gap by proposing the Explainable Continual DeepFake Forensic Network (ECDF-Net). ECDF-Net adaptively fuses spatial and frequency-domain features to capture both visual and spectral manipulation artifacts, maintains a continual prototype memory that is updated only by reliable samples so that new forgery characteristics are learned while previous knowledge is retained, and jointly produces a detection decision, a forgery localization map, and multi-source forensic evidence. Robustness, consistency, and explanation consistency learning are further introduced to stabilize predictions and explanations under common media transformations. The framework is evaluated on FaceForensics++, while Celeb-DF v2 is used for cross-dataset evaluation and continual adaptation. On FaceForensics++, ECDF-Net achieves 98.42% accuracy, a 98.39% F1-score, and an AUC of 0.993. On Celeb-DF v2, it achieves 93.76% accuracy in direct cross-dataset evaluation, which rises to 97.21% after continual adaptation, while accuracy on FaceForensics++ decreases by only 0.46 percentage points. The ablation study confirms the contribution of adaptive fusion, prototype memory, robustness consistency, and explanation consistency learning. The principal contribution of this work is new evidence that detection, continual adaptation, localization, and stable explanation can be achieved from a single spatial-frequency representation, with reliability-gated prototype updates reducing the cross-dataset accuracy gap from 4.66 to 0.75 percentage points without retraining from scratch.
Keywords:	DeepFake detection, spatial-frequency learning, adaptive feature fusion, continual learning, prototype memory, forgery localization, explainable artificial intelligence, forensic analysis.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	MALGUARD-IoT: A COST-AWARE HIERARCHICAL AND EXPLAINABLE EDGE FRAMEWORK FOR MIRAI MALWARE DETECTION
Author:	HANI AL-BALASMEH
Abstract:	This study develops and evaluates MALGUARD-IoT, a cost-aware, hierarchical edge framework for detecting Mirai malware from identity-excluded network-flow statistics. The method selects 24 behavioral features, fits a histogram-gradient-boosting detector with five-fold isotonic calibration, and applies a prespecified decision threshold of 0.25 reflecting a 3:1 false-negative-to-false-positive cost ratio. A second classifier assigns detected malicious flows to reconnaissance, infection, or attack action. The experiments use 40,893 training and 9,335 held-out test flows and compare five baseline models. MALGUARD-IoT achieved test F1 = 1.000 and MCC = 1.000 (0 false positives and 0 false negatives), while the strongest uncalibrated boosting baseline missed three malware flows. Stage attribution achieved macro-F1 = 0.9873. The binary and stage artifacts required approximately 189.4 kB and 41.9 kB, with host-measured inference times of 5.93 and 10.64 ms per 1,000 flows, respectively. Robustness was conditional: F1 decreased to 0.9063 under 10% interquartile-range Gaussian perturbation and to 0.8338 under 20% feature dropout. These results support compact, auditable decisions on the evaluated Mirai benchmark but do not demonstrate cross-family or live-gateway generalization.
Keywords:	Internet Of Things, Mirai Malware, Edge Intrusion Detection, Probability Calibration, Explainable Learning.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	HYBRID QUANTUM CONVOLUTIONAL NETWORKS FOR PARKINSON'S DISEASE STAGE PREDICTION USING WEARABLE SENSOR SIGNALS
Author:	GUDEME JAYA RAO, SAI GEETHA PAMIDIMUKKALA², MADDHI SUNITHA, DURGA BHAVANI KANCHAPU, P SAI ARUN MIHIR, MEDISHETTY SWAPNA, SEEMAKURTI RAVI SANKAR, CHERUKUPALLI RAMESH KUMAR⁸, DARA RAJU
Abstract:	It is difficult to perform automated Parkinson’s disease (PD) stage prediction based on signals obtained from wearables due to the individuality of gait patterns across different patients and the similarity of movement characteristics between adjacent PD stages. In this study, we proposed a novel Hybrid Quantum-Convolutional Network (HQCN) for automated PD stage prediction based on vertical ground reaction force (VGRF) signals measured from both feet through 16 channels. Specifically, our HQCN framework adopted temporal segmentation, multi-scale convolutional feature extraction, a six-qubit variational quantum circuit, temporal attention, and ordinal stage classification. The HQCN was evaluated based on accuracy, precision, recall, macro-F1, Area Under the Curve (AUC), and stage mean absolute error (stage-MAE). Our proposed HQCN achieved an accuracy of 94.26%, macro-F1 of 93.62%, AUC of 0.972, and stage-MAE of 0.184, which outperformed the compared SVM, Random Forest, CNN, CNN-LSTM, and CNN-Attention models. The above results showed that quantum-enhanced latent representation coupled with temporal attention and ordinal stage prediction improved the discrimination of PD stages and reduced large stage-prediction errors. The proposed framework demonstrated the potential of combining wearable sensing and hybrid quantum-classical learning for objective and automated monitoring of Parkinson’s disease progression.
Keywords:	Parkinsons Disease, Wearable Sensors, VGRF Signals, Quantum Machine Learning, Convolutional Neural Network, Stage Prediction
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	COMPARATIVE MACHINE LEARNING MODELS FOR DATA-DRIVEN VESSEL TURNAROUND TIME PREDICTION AT TANGER MED PORT
Author:	MOSTAFA AL UAHABI , HICHAM ATTARIUAS
Abstract:	Maritime shipping underpins global trade, handling more than 90% of global goods movement and links markets across continents. Ports sit at the core of this system and their effectiveness greatly impacts the smooth functioning of worldwide supply chains. Even though larger ships have boosted efficiency per call, they require more docking time. This strain on terminal infrastructure makes Vessel Turnaround Time (VTT) a crucial metric for container terminals (TCs). Existing research predicting VTT often suffers from limited timeframes and a lack of comprehensive variables. This research creates and contrasts eight machine learning regression models to predict VTT at the second TC in Tanger Med Port, representing one of the initial studies to forecast VTT at this specific location using 12 years of Terminal Operating System data. Following a detailed preprocessing phase and feature selection using RFECV, seven significant predictive variables were isolated from an initial set of 114, concentrating on cargo handling volume, crane productivity and operational delays within the terminal. Hyperparameters for each model underwent tuning with a grid search incorporating 5-Fold cross-validation. The best-performing model was BPNN (R² = 0.9741, RMSE = 2.40 h, MAE = 1.61 h). These findings illustrate that a streamlined, data-driven feature set markedly enhances VTT prediction accuracy, offering terminal operators a dependable decision-support tool for operational planning.
Keywords:	Vessel Turnaround Time, Machine Learning, Data-Driven Prediction, Feature Selection, Tanger Med Port.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	SPECTFT-DSDNET: SPECTRAL MODULATED TEMPORAL FUSION TRANSFORMER ENHANCED FOR EMOTION RECOGNITION
Author:	PAGADALA VEERANJANEYULU, PADMAJA USHA RANI DUDA, SINDHU BOIANAPALLI, SRI LAKSHMI PARVATHI SITANABOINA, RAMYA KRISHNA DHULIPALLA
Abstract:	Physiological monitoring for emotion recognition provides objective affective state assessments resistant to social masking. However, decoding non-stationary Electroencephalography (EEG) and Electrocardiogram (ECG) signals across subjects using existing deep neural networks remains challenging due to high spectral blurring, amplitude scale mismatches between microvolt and millivolt signals, and temporal information loss. To address this, the Spectral Modulated Temporal Fusion Transformer Enhanced Dynamic Spectrum Driven Network (SpecTFT-DSDNet) is introduced. The proposed architecture utilizes Synchrosqueezed Dynamic Short-Time Fractional Fourier Transform (SSQ-DSTFFT) spectral mapping, enhanced by a novel, learnable spectral modulation block and a Temporal Fusion Transformer (TFT) to address inter-subject variability. These representations are encoded through an Efficient Convolutional Neural Network (CNN) and processed through a Temporal Fusion Transformer (TFT), which is equipped with a Gated Residual Network (GRN) and Variable Selection Network (VSN) to suppress irrelevant noise. A unique learnable spectral modulation block is integrated between the Long Short Term Memory (LSTM) layers and masked attention to ensure that model remains frequency aware. SpecTFT-DSDNet is evaluated under a subject-independent 5-fold cross-validation protocol, by achieving classification accuracy of 98.89% for valence and 98.88% for arousal on DEAP dataset, 99.28% for valence and 99.41% for arousal on DREAMER dataset, and 98.92% for valence 99.58% for arousal on AMIGOS dataset. SpecTFT-DSDNet enhances spectral energy concentration by 68.6% over existing time-frequency transforms while maintaining an inference latency of 12.4 ms per segment. These ensure that using integrated frequency-aware spectral modulation with gated temporal transformer context effectively, which resolves inter-subject signal variance, introducing scalable and computationally viable solution for cross-subject physiological affective computing.
Keywords:	Efficient Convolutional Neural Network, Emotion Recognition, Gated Residual Network, Synchrosqueezed Dynamic Short-Time Fractional Fourier Transform, Temporal Fusion Transformer, Variable Selection Network
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	QUANTUM GRAPH LEARNING FOR PERSONALIZED CANCER TREATMENT RECOMMENDATION
Author:	M. SRAVANI, MEDI SANDHYA RANI, JAYAMMA RODDA³, JUVVALA SAILAJA, BINODINI KAR, BOBBA SAHITHI, B. KESAVA RAO, K. BAGYALAKSHMI, DARA RAJU
Abstract:	Personalized cancer treatment recommendation involves the integration of heterogeneous data pertaining to genomics, cancer, and drug response. Furthermore, it entails the modeling of complex relationships between various biological entities and treatment options. Conventional machine learning methods might fail to capture such relationships, and the application of quantum-enhanced graph learning to personalized treatment recommendation remains underexplored. In this paper, we propose the framework called Quantum Graph Learning for Personalized Cancer Treatment Recommendation (QGL-PCTR). QGL-PCTR incorporates several classical components, including heterogeneous graph neural networks, relation-aware attention, variational quantum processing, and quantum–classical feature fusion, to develop personalized cancer treatment recommendations. We build a heterogeneous graph representing cancer samples, genes, cancer types, and drugs using the dataset from Genomics of Drug Sensitivity in Cancer (GDSC). A relation-aware graph neural network produces 128-dimensional embeddings, which are then compressed into six features and encoded into six qubits. These qubits undergo processing in a three-layer variational quantum circuit, and the resulting quantum representation is fused with the graph embedding to calculate treatment-compatibility scores and rank candidate treatments. QGL-PCTR yields an overall accuracy of 94.80%, precision of 93.90%, recall of 94.20%, and an F1-score of 94.00%. In addition, the accuracy for recommending the top-5 treatments is 94.30%. Comparative and ablation analyses of the proposed approach demonstrate improvements over the evaluated classical machine-learning and graph-learning baselines. Moreover, ablation tests reveal the complementary contributions of relation-aware attention, quantum circuit processing, and feature fusion. The suggested method provides a computational approach for personalized treatment-response prediction and ranking potential treatment candidates. Yet, validation using patient-level clinical data is needed to establish the suitability of QGL-PCTR for clinical decision support.
Keywords:	Quantum Graph Learning; Personalized Cancer Treatment; Heterogeneous Graph Neural Network; Quantum Machine Learning; Treatment Response Prediction; Drug Recommendation.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A SYSTEMATIC REVIEW OF TEXT-IMAGE MULTIMODAL SARCASM DETECTION: MODELING APPROACHES, BENCHMARK RELIABILITY, AND EVALUATION CHALLENGES
Author:	CHAO BING, AZREEN AZMAN, NURUL AMELINA NASHARUDDIN, MASRAH AZRIFAH AZMI MURAD
Abstract:	The increasing use of images and text to convey ideas in social media has drawn researchers' attention to the task of multimodal sarcasm detection. Multimodal sarcasm detection is different from the traditional text-only sarcasm detection in that the incongruities between the modalities have to be captured as well as the meaning of each modality. While recent years have seen some advances, there are currently no uniform standards for modelling strategies, dataset construction and evaluation. Hence this paper conducts a systematic review of the available literature in the social media context. Specifically, this research focuses on induction and classification of the current literature on fusion-based architectures, incongruity modeling strategies, knowledge guided methods and task variants other than binary classification. Moreover, we discuss popular datasets and their drawbacks like annotation noise, shortcut learning bias and lack of diversity. Meanwhile, unfair comparisons due to differences in evaluation metrics and experimental setting is discussed. The results of this study further highlight the major challenges of multimodal sarcasm detection such as subjectivity, context dependency, cross-platform variability, etc., and provide suggestions for future research on the development of more robust, interpretable, and generalizable multimodal sarcasm detection models. The overall purpose of this literature review is to structure the reference for research in this area and standardize the evaluation practice.
Keywords:	Multimodal Sarcasm Detection, Cross-Modal Interaction, Incongruity Modeling, Social Media Analytic, Systematic Literature Review
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	RANSOM-XAI: AN EXPLAINABLE STATIC-BEHAVIORAL FUSION FRAMEWORK FOR FAMILY-DISJOINT RANSOMWARE DETECTION
Author:	HANI AL-BALASMEH
Abstract:	Ransomware detection remains difficult when previously unseen malware families resemble known threats and a classifier provides insufficient evidence for security decisions. This study proposes RANSOM-XAI, a static–behavioral fusion framework that separates Portable Executable attributes from sandbox behavior, calibrates each LightGBM branch, and combines the class probabilities using a logistic stacker. A family-disjoint research design evaluates transfer to four withheld ransomware families and four withheld other-malware families. On 21,752 samples, the family-disjoint test achieved 97.44% accuracy, 0.9716 macro-F1, 0.9606 Matthews correlation coefficient, and 0.9647 ransomware F1. Early feature fusion achieved a slightly higher macro-F1 (0.9735), whereas the proposed method retained interpretable branch evidence. SHAP masking gave a 5.54-fold greater ransomware-probability reduction than random masking; top-10 explanation stability was 0.9636. Requiring branch agreement and confidence of at least 0.85 allowed 66.24% of decisions to be automated at 98.82% accuracy. However, behavioral-branch failure reduced macro-F1 to 0.1539, and slight behavioral-feature noise caused substantial deterioration. The contribution is thus an auditable, family-disjoint decision framework combining calibrated fusion, quantitative explanation checks, and selective review rather than a claim of universal detection or the highest raw predictive score. Reliable runtime telemetry and external validation remain prerequisites for deployment.
Keywords:	Ransomware Detection, Explainable Machine Learning, Family-Disjoint Validation, Static–Behavioral Fusion, Selective Classification.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	AN EXPLAINABLE DEEP LEARNING FRAMEWORK FOR MULTI-CLASS BONE CANCER HISTOPATHOLOGICAL IMAGE CLASSIFICATION USING GRID-OPTIMIZED EFFICIENTNET BONE
Author:	GUNTOJU KALPANA DEVI, MAHESH BABU ARRAMA
Abstract:	Bone cancer is a rare but aggressive malignancy that primarily affects the skeletal system and poses significant challenges for accurate diagnosis and treatment planning. Among primary bone cancers, osteosarcoma is the most prevalent subtype, requiring precise histopathological evaluation to distinguish between healthy tissue, viable tumor regions, and necrotic tumor areas. However, manual examination of histopathological images is labour-intensive, time-consuming, and prone to inter-observer variability. This study therefore addresses the research problem of whether a lightweight transfer-learning model, when systematically tuned and made explainable, can classify osteosarcoma tissue accurately, stably, and transparently. To address these challenges, this study proposes GO-EfficientNetBone, a Grid-Optimized EfficientNetBone framework for automated multi-class bone cancer tissue classification. The proposed methodology integrates image preprocessing, data augmentation, EfficientNetB0-based deep feature extraction, Global Average Pooling, Grid Search hyperparameter optimization, stratified 10-fold cross-validation, and Grad-CAM explainability within a unified architecture. Histopathological images obtained from the TCIA Osteosarcoma Tumor Assessment Dataset were categorized into Non-Tumor, Viable Tumor, and Necrotic Tumor classes. Experimental evaluation demonstrated that the proposed framework achieved an overall accuracy of 99.04%, precision of 98.97%, recall of 98.91%, F1-score of 98.94%, MCC of 0.985, and a macro-average AUC of 0.998. Furthermore, Grad-CAM localization achieved an average accuracy of 97.33%, providing visual explanations that enhance model transparency and clinical interpretability. Comparative analysis and ablation studies confirmed the superiority of the proposed framework over conventional machine learning and deep learning approaches. The results indicate that GO-EfficientNetBone provides an accurate, robust, and explainable solution for computer-aided bone cancer diagnosis, offering significant potential for supporting pathologists in clinical decision-making and treatment response assessment. The study is limited to a single public dataset analysed at the patch level; future work will therefore focus on patient-level, multi-centre, and whole-slide validation.
Keywords:	Bone Cancer, Osteosarcoma, Histopathological Image Classification, Deep Learning, EfficientNetB0, Transfer Learning, Grid Search Optimization, Explainable Artificial Intelligence, Grad-CAM, Computer-Aided Diagnosis.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	HUMAN-CENTRIC DATA GOVERNANCE FOR TRUSTWORTHY DIGITAL TRANSFORMATION: DEVELOPING AND EMPIRICALLY VALIDATING A FRAMEWORK FOR DIGITAL TRUST
Author:	RAJAT VERMA, SHASHI KANT GUPTA
Abstract:	While digital transformation initiatives rely more on the intensive mobilization, circulation and algorithmic processing of personal and organizational data, the governance systems that frame these data are largely compliance-based and technology-centric. This orientation addresses the issue of why investments in technology are often not matched by ongoing stakeholder acceptance. This study builds and tests, through a simulation-based analysis, a Human-Centric Data Governance (HCDG) framework, where digital trust serves as a vehicle to bring governance practices to a transformation result. The framework includes five elements of governance which are as follows: Transparency of the Data Practices, Individual Data Agency, Algorithmic Accountability, Ethical Stewardship Capability and Participatory Oversight, while perceived data risk serves as a countervailing factor. A 41-item multi-item reflective instrument was identified in the literature, and a synthetic dataset was created with 612 respondents (involving managerial and professional profiles in manufacturing, financial services, healthcare and public-sector organizations) from an explicit latent-variable model that can be used for Internal Coherence. The measurement and structural models were assessed using Structural Equation Modelling in the partial least squares tradition (equal-weight composites), together with a bootstrapped mediation analysis, a permutation-based multi-group analysis, and an importance–performance map analysis. Construct reliability and discriminant validity were met for all constructs (Composite Reliability ranged from 0.908 to 0.924 and the average variance extracted ranged from 0.701 to 0.752). Together the five governance dimensions and the perceived data risk account for 54.5% of the variance in digital trust. Algorithmic accountability has the largest positive influence (β = 0.202), and perceived data risk has a significant negative influence (β = −0.110). Digital trust is statistically significantly mediating all fifteen governance-outcome relationships, while the permutation-based multi-group analysis was unable to detect any significant sector differences for the governance pathways at the 5% level. The simulation-tested framework provides a concise and quantifiable foundation for designing governance arrangements that contribute to sustaining trustworthy digital transformation.
Keywords:	Human-Centric Data Governance, Digital Trust, Digital Transformation, Algorithmic Accountability, Data Ethics, Structural Equation Modelling, Simulation-Based Validation, Stakeholder Commitment.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	HYBRID TF-IDF - WORD2VEC FEATURE FUSION WITH STACKING ENSEMBLE LEARNING FOR URDU FAKE NEWS DETECTION
Author:	ZAHIR ABBAS KHAN, REKHA V
Abstract:	The rapid dissemination of misleading information online has become a major social issue, particularly in low-resource languages like Urdu. The ability of existing false news detection techniques to recognize semantic connections in textual data is limited because they often solely employ lexical representations. This study proposes a hybrid fake news detection method that combines Word2Vec embeddings with Term Frequency–Inverse Document Frequency (TF-IDF) to leverage lexical and semantic information present in Urdu news articles. The gathered features are fused and classified using a stacking ensemble that includes Support Vector Machine (SVM), Logistic Regression (LR), and Extreme Gradient Boosting (XGBoost) as foundation learners. For the studies, a combined Hook-and-Bait Urdu fake news dataset including 78,406 news articles 42,179 real and 36,230 fake news samples—was utilized. The proposed framework had an accuracy of 95.17%. Comparative studies show that the hybrid representation and ensemble learning approach works better than conventional machine learning techniques based solely on TF-IDF features. The findings demonstrate that integrating lexical and semantic representations can greatly improve Urdu false news detection ability.
Keywords:	Fake News Detection, Machine Learning, Natural Language Processing, Stacking Ensemble, TF-IDF, Word2Vec
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	PROOFER: INTEGRATING DIGITAL GAME-BASED ACTIVITY PACKAGES AND DELIBERATE PRACTICE FOR PROFESSIONAL PROOFREADING SKILL DEVELOPMENT
Author:	PARWAPUN KAMTAB, NAPHEE PORNNUAM
Abstract:	Professional proofreading requires repeated, standards-based practice in language, document presentation, and scholarly conventions, yet technology-enhanced training models integrating game-based practice and deliberate practice remain limited. This study developed and preliminarily evaluated PROOFER, a seven-stage professional training model integrating digital game-based activity packages with deliberate practice for proofreading skill development. A three-phase research-and-development process comprised model synthesis and expert review; development and quality evaluation of the digital game-based activity package, training website, digital manual, training activity plans, and research instruments; and field implementation using a one-group posttest-only design. Five experts reviewed the model, other expert panels examined the supporting resources and instruments, and the package was piloted with 10 individuals with characteristics similar to those of the target group. Fifteen entry-level assistant editors from Sukhothai Thammathirat Open University Press voluntarily completed an 18-hour program over three weeks. The PROOFER sequence comprises Prepare and Diagnose; Review Standards and Modeling; Objective Goal Setting; Online Game-Based Focused Practice; Feedback and Error Analysis; Extended Repetition and Progressive Challenge; and Real-Task Transfer and Reflection. Expert evaluation of the model was very high (M = 4.61, SD = 0.51), and the digital game-based activity package was rated very high in quality (M = 4.74, SD = 0.48). Participants achieved an overall post-training proofreading performance score of 86.87%, with technical proofreading highest (90.56%), followed by academic proofreading (85.83%) and language proofreading (84.22%); overall satisfaction was very high (M = 4.78, SD = 0.42). The findings support the feasibility of PROOFER as a structured architecture for professional proofreading practice, although the posttest-only design limits interpretation to post-training performance rather than causal effectiveness.
Keywords:	Deliberate Practice, Digital Game-Based Activity Package, Digital Game-Based Learning, Professional Proofreading, Workplace Training
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A FEDERATED DEEP REINFORCEMENT LEARNING FRAMEWORK FOR AUTONOMOUS NETWORK TRAFFIC OPTIMIZATION IN 6G NETWORKS
Author:	RUDRA YAMINI RANI, MADDELA YELLAMMA, ADDANKI MEGHANA SAI SREE, K.V.Y RADHA SRI, K DEVIPRIYA5, M. LENIN KUMAR, R. CHANDRAMOHAN, SRILAKSHMI RAMYA SAKAMUDI
Abstract:	Ultra-dense connectivity, massive Internet of Things (IoT) deployment, holographic communications, latency-sensitive intelligent applications, and digital twins are anticipated to be enabled by the rapid adoption of 6G networks. Yet, due to the variety and diversity of 6G traffic, traditional traffic optimization techniques, such as load balancers, face numerous challenges, including limited adaptability and scalability, as well as unresolved privacy concerns. To address these issues, a Digital Twin-Assisted Federated Multi-Agent Deep Reinforcement Learning Framework (DT-FMDRL) is proposed for 6G networks to optimize autonomous network traffic management. The proposed framework integrates Multi-Agent Deep Reinforcement Learning (MADRL), Digital Twin-assisted training, Adaptive Performance-Aware Aggregation (APAA), and Traffic Volatility-Aware Reward Learning to enable distributed, intelligent traffic management without sharing raw network data while preserving privacy. To train and evaluate the scheme, we generated 2.5 M traffic samples in the 6G-TrafficNet digital twin simulation environment. Experimental analysis shows that DT-FMDRL outperforms baselines such as Heuristic Traffic Engineering (HTE), SDN-DRL, MADRL, and classical FedAvg-DRL. After 245 rounds of communication, the scheme achieves a throughput of 14.0 Gbps, an average latency of 6.23ms, a packet loss of 1.8%, and a congestion index of 0.28, which guarantees convergence. The results show that this multi-agent RL approach, together with a digital twin and a federated learning (FL) algorithm, can substantially improve efficiency when optimizing traffic patterns and address data privacy concerns. The proposed DT-FMDRL framework is scalable, energy-efficient, and independent, suitable for future intelligent communications and 6G networks.
Keywords:	6G Networks, Federated Deep Reinforcement Learning, Network Traffic Optimization, Digital Twin, Multi-Agent Reinforcement Learning, Energy-Efficient Network Management
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	REAL-TIME MULTIMODAL PROGNOSIS IN LIVER TRANSPLANTATION USING AN ITERATIVE CROSS-ATTENTION FUSION NETWORK
Author:	DIVYA MIDHUNCHAKKARAVARTHY, CHANUMOLU KIRAN KUMAR, V LAKSHMAN NARAYANA
Abstract:	The correct and timely diagnosis following liver transplantation depends on the collation of disparate clinical, radiologic and microscopic data. In this paper, we present the Iterative Cross-Attention Fusion Network (ICAF-Net), a multimodal deep-learning based approach to address three key challenges for graft survival prediction, post-transplant complication detection, and readmission risk assessment. It consists of five components: Spatio-Temporal Cross-Modality Attribution Encoder (STCMAE), Radiopathological Graph Attention Alignment (RP-GAA), Multi-Resolution Cross-Entropy Disentangler (MR-CED), Adaptive Causal Inference Layer for Prognostic Reasoning (ACI-PROG), and Latency-Controlled Prognostic Transformer (LC-ProTrans). The framework aims to learn the temporal dependency, to synchronize the radiological and histopathological representations, to remove the modality-specific noise, to estimate the effects of the treatment and to control the latency of the inference process. The reported evaluation is conducted using the clinical time-series data that exists in MIMIC-III, the liver computed tomography (CT) data that exists in LiTS, and the histopathological images that exist in TCGA-LIHC. The accuracy of predicting graft survival, detecting post-transplant complications, and predicting readmission risk was 0.91, 0.88 and 0.83 respectively for the ICAF-Net. The average and maximum inference speed were 210 ms and 250 ms, respectively. The results show that these classifiers have a good predictive accuracy and computational efficiency in the test environment. The results, however, are subject to limitations in direct clinical conclusions as differences exist between the source datasets, and there is a lack of evidence of prospective, transplant specific, external validation. Additional validation work will be required before clinical deployment, with patient linked transplant cohorts.
Keywords:	Liver Transplantation, Multimodal Fusion, Cross-Attention, Prognostic Modeling, Causal Inference, Process.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	ENERGY-EFFICIENT RESOURCE ACCOUNTING FOR IOT-DRIVEN AMI USING FEDERATED RL AND TINYML OPTIMISATION
Author:	SAHANA EDWIN, DR. BALAKRISHNAN S, MRS. MODUGULA SIVAJYOTHI, DR. G. SHANMUGASUNDAR, DR. K. LAKSHMI KALA, VAMSIDHAR TALASILA, ELANGOVAN MUNIYANDY, K. V. NARASIMHA REDD
Abstract:	IoT-based smart meters are being used more and more in Advanced Metering Infrastructure (AMI). Reducing energy consumption during distributed learning, communication, and data processing is still difficult, though. Energy-aware framework is used in this paper. It integrates TinyML optimisation, actor-critic-based resource allocation, and unified energy accounting. The framework handles client participation, local training, communication scheduling, and model compression based on resources. Evaluation on two public Zenodo datasets is conducted to compare the proposed method with FedAvg, local TinyML and reinforcement learning without TinyML methods. The proposed framework's total energy consumption 18.4 kJ, energy per task 9.2 mJ, communication overhead 142 MB, inference latency 3.1 ms, peak memory 86 KB, and mean absolute error 0.0384. Energy consumption reduced by 42.1% compared to FedAvg with comparable predictive performance. The contributions of adaptive resource allocation, TinyML optimisation, and communication scheduling are also reflected in the results of the ablation. The results suggest that the concept of integrated resource optimisation for energy-efficient AMI analytics is promising and needs to be validated in real deployments beyond the simulated evaluation setting.
Keywords:	Advanced Metering Infrastructure; Energy Accounting; Federated Learning; Actor–Critic Reinforcement Learning; Tinyml; Energy Optimisation.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	QUANTUM DEEP LEARNING FOR EARLY DETECTION OF ACUTE RESPIRATORY DISTRESS SYNDROME USING IOT SENSORS
Author:	NAM VASUNDHARA, BANANA LAKSHMANA RAO, PRASUNA GRANDHI, MARISETTI SRI DURGA, ROSHINI L, KURAGANTI SANTHI, K PRASANTHI, R. KIRUBA BURI
Abstract:	Acute Respiratory Distress Syndrome (ARDS) is an important respiratory disorder where early risk assessment is essential for timely clinical intervention. In this research work, an IoT-LSTM-PCA-Variational Quantum Classifier (VQC) hybrid model was suggested for early risk prediction of ARDS by employing physiological sensor data. For the analysis of the physiological data, six physiological parameters consisting of oxygen saturation (SpO₂), respiratory rate (RR), heart rate (HR), body temperature (BT), systolic blood pressure (SBP), and diastolic blood pressure (DBP) were processed using normalization and temporal feature extraction. An LSTM neural network was used for learning the sequential physiological patterns, whereas Principal Component Analysis (PCA) technique was used to reduce the dimensionality of learned features followed by quantum encoding. Finally, the reduced features were processed by VQC for the purpose of ARDS risk prediction. Experiments carried out on the MIMIC-IV dataset produced an accuracy of 96.8%, precision of 96.4%, sensitivity of 97.1%, specificity of 96.2%, F1-score of 96.7%, and AUROC of 0.982. Comparative and ablation studies revealed that the combination of temporal features, PCA, and VQC resulted in better prediction performance in comparison with conventional machine learning methods and standalone LSTM models. The proposed framework provides a compact hybrid deep learning–quantum approach for continuous physiological monitoring and early ARDS risk identification, with potential applicability to IoT-enabled clinical early-warning systems. Further prospective and multi-center validation is required before clinical deployment.
Keywords:	Acute Respiratory Distress Syndrome, Internet of Things, Long Short-Term Memory, Principal Component Analysis, Variational Quantum Classifier, Early Risk Prediction.
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	GS-BERT: GRAPH-STRUCTURED BERT FOR SENTENCE-LEVEL EXTRACTIVE DOCUMENT SUMMARIZATION
Author:	BARAA A. ELHADY, HELAL A. SULEIMAN2 OSAMA E. EMAM
Abstract:	Automated document summarization is essential for condensing voluminous textual data into concise, informative summaries. In domains where factual reliability is mandatory, extractive summarization is widely preferred over abstractive generation because directly extracting source sentences prevents factual inconsistencies and generative hallucinations. We present GS-BERT (Graph-Structured BERT), a hybrid neural model designed for sentence-level extractive summarization. GS-BERT integrates contextual sentence representations from BERT with multi-layer Graph Convolutional Networks (GCNs) and introduces a trainable Sentence-level Positional Encoding (SPE) layer. Each sentence is mapped to a contextual vector via BERT, enriched with learnable positional information, and linked through an attention-driven dynamic adjacency matrix. A three-layer GCN conducts relational message passing across sentences, after which a multi-layer perceptron classifies sentence salience. The architecture was evaluated on 287,113 training articles from the CNN/DailyMail benchmark and tested on 1,000 unseen articles under a single-epoch training regimen. GS-BERT attained ROUGE-1, ROUGE-2, and ROUGE-L F1 scores of 42.1, 19.5, and 40.2, respectively, markedly exceeding the conventional Lead-3 baseline and demonstrating superior balance in sentence selection across document sections. The findings confirm that coupling pretrained transformer representations with topological graph convolutions effectively captures non-local document relationships while retaining vital discourse order, establishing an effective architectural framework for extractive text summarization.
Keywords:	Extractive Summarization, Transformer Encoders, Graph Convolutional Networks, Positional Encoding, ROUGE Evaluation
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	A ROBUST GAN FRAMEWORK WITH PERCEPTUAL MULTI-LOSS LEARNING FOR APPLE LEAF IMAGE ENHANCEMENT
Author:	SHIVANI GUPTA, POOJA RANI
Abstract:	High-quality images are essential for improving the accuracy of plant disease identification. Images captured in agricultural fields under natural conditions are often affected by various types of noise due to illumination changes, environmental disturbances and sensor limitations. This noise reduces the effectiveness of deep learning models. In this study, a perceptual GAN-based framework is developed to enhance the quality of apple leaf images obtained from PlantVillage dataset. In this proposed framework, adversarial loss, L1 loss, perceptual loss using a pre-trained VGG16 network, and structural similarity (SSIM) loss are combined so that both visual quality and structural information can be preserved. An encoder–residual–decoder architecture is used for the generator, while the PatchGAN is adopted as discriminator to enforce local texture realism. To make the training data more realistic, Gaussian, salt-and-pepper, and speckle noise were added during training. The performance of the proposed approach is compared with the Gaussian filter, Median filter, DnCNN, and UNet, and the best results are obtained in terms of PSNR and SSIM. The proposed perceptual multi-loss GAN achieved an average PSNR of 31.29 dB and an SSIM of 0.82 on images affected by Gaussian, salt-and-pepper, and speckle noise. The enhanced images exhibit improved structural clarity and are suitable for preprocessing in automated plant disease detection systems.
Keywords:	Apple Leaf Image Enhancement, Image Denoising, Generative Adversarial Network, Perceptual Loss, Deep Learning
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	AN INTELLIGENT MULTI-AGENT AGENTIC AI FRAMEWORK FOR RETAIL SUPPLY CHAIN OPTIMIZATION AND LSTM-BASED DEMAND PREDICTION
Author:	MALLANNAGARI SUNITHA, THONDEPUADILAKSHMI, DR.B. JYOSHNA, DR.R. SATEESH KUMAR, DR.B. KIRANMAI
Abstract:	Traditional retail inventory evaluation and control systems rely heavily on static threshold criteria and predetermined lot-sizing rules that lack adaptability during compound supply chain disruptions and fail to incorporate multi-modal contextual information. This paper proposes a comprehensive five-layer Agentic artificial intelligence framework that synergistically integrates four independently developed technologies into a unified, deployable retail supply chain simulation environment: Large Language Model (LLM) ReAct reasoning, Long Short-Term Memory (LSTM) primary demand forecasting, ChromaDB episodic memory retrieval, and LangGraph-orchestrated state machine workflow control. The multi-agent architecture comprises four collaborating autonomous actors Demand, Warehouse, Supplier, and Logistics agents executing on top of a discrete-event Mesa simulation scaffold. Each agent generates structured, auditable decision records (Decision Record) directly at inference time. In a rigorous 30-day stress simulation bench marking against classical rule-based control policies under compound multi-stage disruptions, the proposed agentic system achieves an approximate 9 percentage point improvement in order fill rate (86%to95%), a 71% reduction in stock out days (21days down to 6 days), and a35% decline into supply chain operational cost ($57k down to $37k), driven primarily by eliminating catastrophic stock out penalties. Component ablation demonstrates that LangGraph conditional state routing and Message Bus peer-to-peer communication contribute most significantly to system’s ability. Comprehensive mathematical modeling, workflow state machines, and operational counter measures are detailed to support enterprise deployment.
Keywords:	Multi-Agent Systems, Retail Supply Chain, LLM Agents, LSTM Forecasting, LangGraph
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	H3D-TNET: SURFACE-CANONICALIZED GEOMETRY–LANGUAGE LEARNING WITH 3D-TEXT METRICS FOR JOINT DETECTION, RECOGNITION, AND UNDERSTANDING OF 3D TEXT
Author:	AVINASH BURADAGUNTA, DR M VIJAY REDDY, DR. G. RAMESH CHANDRA
Abstract:	Reading text from 3D objects is important for industrial inspection, robotics, augmented reality, and heritage object digitisation; however, current 2D pipeline approaches to optical character recognition (OCR) do not work when text appears on a non-planar and arbitrarily oriented surface. In this paper, we propose H3D-TNet (Hybrid 3D Text Network), a new hybrid deep learning model that allows for text detection, recognition, and semantic understanding using RGB-D and point cloud data. The proposed approach unifies a surface-aware canonicalization layer, two parallel encoders for 3D spatial features (sparse 3D CNN and geometry-aware point-transformer branch), OCR using a transformer-based layer, and a language reasoning module. All these components are trained in a multi-task way with loss functions accounting for text detection, localisation, recognition, semantic classification, and geometric alignment tasks. Since regular 2D detection metrics do not measure 3D localisation and semantic correctness, we also introduce five 3D text metrics: 3D Localisation Accuracy Score (3D-LAS), Geometric Alignment Score (GAS), Semantic Understanding Score (SUS), Robustness Evaluation Index (REI), and an Overall 3D-Text Intelligence Score (3D-TIS). The proposed H3D-TNet is compared with eight baselines, namely PointNet++, Point Transformer, PV-RCNN, Text4Point, DBNet+TrOCR, CRAFT+TrOCR, a vision-language model built based on CLIP, and a simple OCR plus depth model, on a collection of both public (ScanNet, SUN RGB-D, ICDAR, Total-Text) and synthetic 3D-text data. In contrast to the baselines, H3D-TNet achieves up to 6-10% improvements in terms of 3D-text detection F1-score and 8-12% improvements in terms of semantic understanding. Moreover, ablation and robustness studies (rotation, point dropout, and Gaussian noise) confirm the contribution of each component to the overall performance and robustness of the proposed framework. Also, the analysis of the rank correlation between different evaluation metrics and perturbation conditions reveals the consistent superiority of the method across different comparison conditions instead of having a single point of advantage. Furthermore, qualitative case studies also illustrate the effectiveness of the proposed framework in detecting curved industrial markings, degraded inscriptions, and augmented reality text overlays. The principal scientific contribution is a single end-to-end formulation in which surface canonicalization, 3D geometric encoding, transformer OCR, and language reasoning are learned jointly, together with a dedicated set of metrics for scoring text on 3D surfaces. The framework offers practical value for industrial inspection, robotics, augmented reality, and heritage digitisation, while open issues remain in real-world validation, multi-seed statistical confirmation, and on-device efficiency.
Keywords:	3D Text Detection, Point Cloud Learning, Vision–Language Networks, Transformer OCR, RGB-D Fusion, Multimodal Deep Learning, Surface Text Understanding, Geometric Alignment
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text

Title:	QUANTUM-ENHANCED DEEP LEARNING FRAMEWORK FOR AUTOMATED IDENTIFICATION OF ENDOMETRIOSIS FROM PELVIC MEDICAL IMAGES
Author:	K DEVIPRIYA, SREELALITHA TALLURI, K. S. S. NARAYANA, SAPE CHITTIBABULU, M. SUMALATHA, HARI KISHAN CHAPALA, SESHA BHARGAVI VELAGALETI, GARIGIPATI RAMA KRISHNA
Abstract:	Additionally, lesions can manifest in many ways, and diagnostic algorithms for identifying them in chronic disorders can be complex, depending on anatomical complexity, imaging acquisition methods, and lesion presentation. In medical image analysis, deep learning has advanced significantly; however, classic models may suffer from limited data, noise, and prediction inaccuracies. This work, QEDL-Endo, is a quantum-enhanced deep learning method for automatically diagnosing endometriosis from pelvic-region magnetic resonance imaging (MRI). It uses an 8-dimensional latent representation, classical–quantum feature fusion, an 8–qubit variational quantum circuit, an uncertainty module and a modified classifier. The goal is to determine whether this approach can improve classification accuracy while keeping the quantum architecture's resource complexity low by using human anatomy to learn a representation during training and a quantum feature map to map the chosen features into resource-efficient representations. On this reported test set, QEDL-Endo reached 95.1% accuracy, an AUROC of 0.978 and an AUPRC of 0.974. Additionally, it reported a Brier (probability) score of 0.061 and an expected calibration error of 0.029 for its probabilistic predictions. Performance improved after introducing attention, amplitude transformation, and feature fusion. These data suggest that this proposed architecture may be a computer-aided imaging method, though it may not offer a quantum advantage or be clinically ready. A second constraint is that not all units are available and independent; therefore, validation requires multi-center evaluation of real-world performance. Future studies should focus on the study's external generalizability, the use of validated diagnostic labels, and whether a quantum advantage was realized, regardless of statistical significance over an equivalently optimized classical model.
Keywords:	Endometriosis, Pelvic MRI, Deep Learning, Quantum Machine Learning, Anatomy-Guided Attention, Medical Image Classification
DOI:
Source:	Journal of Theoretical and Applied Information Technology 31^st October 2026 -- Vol. 104. No. 20 -- 2026
Full Text