|
|
|
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
August 2026 | Vol. 104
No.15 |
|
Title: |
TOPOLOGY-AWARE DEEP LEARNING FOR ROBUST 3D MRI BRAIN TUMOR SEGMENTATION |
|
Author: |
DR.ARADDHANA, DR SIVANEASAN BALA KRISHNAN, DR. SHRIKANT KULKARNI , DR. PRASUN
CHAKRABARTI, JINU SOPHIA J, SILAS STEPHEN D, S.B.G.TILAK BABU |
|
Abstract: |
The high morphological variability, indistinct tumor boundaries, and severe
class imbalance in clinical imaging data pose challenges to accurate 3D brain
tumor segmentation from multi-modal MRI. These issues are further exacerbated
within the heterogeneous, multi-vendor, multi-field-strength imaging settings
common in Indian tertiary-care hospitals, which is essential for
neuro-oncological diagnosis and treatment planning. TopoSegNet is a novel
Topology-aware Encoder-Decoder network that, for the first time, combines Betti
number regularization based on persistent homology (PH) and a convolutional
block attention module (CBAM)-based decoder to perform volumetric brain tumor
segmentation. The persistent homology bottleneck explicitly filters the
predicted segmentation to ensure that it preserves the topological invariants –
connected components (β₀), loops (β₁) and voids (β₂) – of the ground-truth tumor
mask, thereby decreasing the number of spurious connected components and hollow
predictions by 38.7% compared to the best baseline model. Researcher also
suggest a boundary-aware composite loss that combines Dice, cross-entropy loss,
Hausdorff distance and topological persistence losses in a learnable manner.
TopoSegNet is benchmarked on a novel 450-subject multi-institutional Indian
dataset acquired across five premier hospitals, including five tumor grades and
1.5T and 3T scanner protocols. TopoSegNet outperforms three strong contemporary
baselines, nnU-Net v2, SwinUNETR, and TransBTS, with respect to Dice score, with
statistically significant improvements obtained on whole tumor, tumor core, and
enhancing tumor sub-regions respectively. Compared to nnU-Net v2, the accuracy
of the model is improved by 19.4% for the HD95 metric with the value reduced
from 8.2 mm to 7.9 mm (WT). |
|
Keywords: |
Brain Tumor, Imaging, Diagnosis, Persistent Homology, Dice Score, Segmentation |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
BEYONF ACCURACY: ACOST-ESNSITIVE CLINICAL FRAMEWORK FOR MINMIZING MISSED
DIABETES DIAGNOSES WITH MULTI-DATASET VALIDATION |
|
Author: |
A.M.M.MADBOULY, MONA.R.ELHEFNAWY,RAMADAN BABERS |
|
Abstract: |
This study proposes a cost-sensitive machine learning framework specifically
designed to minimize false negatives in diabetes screening—a critical clinical
imperative often overlooked in traditional accuracy-focused approaches. By
integrating advanced resampling techniques (SMOTE, ADASYN, SMOTE-Tomek) with
multiple machine learning classifiers and systematic threshold optimization, we
created a framework that prioritizes sensitivity over traditional metrics.
Validation was performed on two independent datasets: the Pima Indians Diabetes
Database (768 samples) and a symptom-based diabetes dataset (520 samples). Our
framework achieved recall rates exceeding 0.94 across all model configurations
on the Pima dataset, with perfect recall (1.0) demonstrated on the symptom-based
dataset. Crucially, threshold optimization analysis revealed that an optimal
decision threshold of 0.11-0.15 (rather than the conventional 0.5) minimizes
clinical misclassification costs while maintaining high sensitivity. Sensitivity
analysis across cost ratios (3:1 to 10:1) confirmed that optimal thresholds
consistently remain below 0.2, demonstrating robustness across clinically
relevant scenarios. This work represents the first systematic integration of
advanced algorithms, modern resampling techniques, and cost-driven threshold
optimization for clinical-safe diabetes screening. Our findings establish a
methodological framework applicable to other safety-critical disease screening
scenarios where the cost of missed diagnoses substantially exceeds the cost of
false alarms. |
|
Keywords: |
Diabetes Prediction; Cost-Sensitive Learning; Imbalanced Data; SMOTE; Recall
Optimization; Clinical Decision Support; Machine Learning. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
CLUSTER-GUIDED ROBUST HYBRID LEARNING FOR IMBALANCED SOFTWARE DEFECT PREDICTION |
|
Author: |
DR SAMBASIVARAO BARAGADA, P.SUNITHA DEVI, BAGAM LAXMAIAH, DR V RAMALATHA,
JAYAVELU S, PUPPALA RAMYA, ELANGOVAN MUNIYANDY, DR. SURESH KUMAR PITTALA |
|
Abstract: |
Software defect prediction datasets typically exhibit missing values, noisy
metrics, severe class imbalance, and latent subgroup structures that existing
models fail to exploit. Current methods deal with these issues in isolation,
leading to overall information loss, predictions that are different across
folds, and poor recall for the minority class in industrial situations. The
paper proposes RCIE-SDP (Robust Clustering-Informed Ensemble for Software Defect
Prediction), which is a unified hybrid framework that jointly performs robust
preprocessing, uncovers latent defect subgroups through unsupervised clustering,
and trains a cluster-guided imbalance-aware ensemble classifier with calibrated
probability outputs. The proposed model is evaluated on the 63,586 instance
SQuaD benchmark. Moreover, the experiment uses 725 heterogeneous features from
450 projects. It reports an F1-score of 0.835, AUC of 0.934, and AUPRC of 0.871.
Further, the score is obtainable better than Random Forest and XGBoost. In other
words, the result is better by 16.6% and 8.7%, respectively, for these datasets.
Other effort-aware metrics yield Recall@20%LOC of 0.734. Calibration analysis
shows a Brier score of 0.089. The framework exhibits increased stability
(standard deviation less than 0.029 for all metrics) and generates accurate
calibrated probabilities, which is a requirement for risk-based triage. The
findings of this paper demonstrate that incorporating clustering guidance in an
imbalance-aware ensemble learning framework leads to practical improvement in
defect detection in the presence of noise and imbalance. |
|
Keywords: |
Software Defect Prediction; Imbalanced Learning; Clustering-Guided Ensemble;
Robust Preprocessing; Effort-Aware Metrics; Model Calibration |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
TOWARD TRUSTWORTHY RANSOMWARE DETECTION: AN EXPLAINABLE AND IMBALANCE-AWARE
TRANSFORMER-BASED FRAMEWORK |
|
Author: |
ARUNA RAO S L , DR. NAGARATNA P HEGDE , DEVI , DR V RAMALATHA , JAYAVELU S ,
PUPPALA RAMYA , ELANGOVAN MUNIYANDY, DR. SURESH KUMAR PITTALA |
|
Abstract: |
The use of encryption, data theft, and disruptive destruction has made
ransomware a major cybersecurity threat, and traditional signature-based
defences are often ineffective against novel and rebranded ransomware. The
purpose of this study was to construct an explicable, imbalance-conscious deep
learning model of detecting ransomware based on handcrafted behavioural features
and transformer-based representation learning. The experimental machine-learning
design was implemented on a new public ransomware corpus of 21,752 samples (26
malware families). The training was skewed intentionally to represent the class
imbalance that was present. The proposed pipeline extracted file-system,
process, registry, API, and network indicators and then learned higher-order
dependencies using a transformer encoder that is optimised with focal loss,
class weighting, and SMOTE. The proposed model had a better accuracy of 97.18%
with a 0.962 precision, 0.951 recall, 0.956 F1-score, and 0.986 ROC-AUC,
compared to the transformer baseline and other classical and deep learning
models. Analysis of SHAP and attention further demonstrated that the most
noteworthy indicators were file rename rate, write bursts that appear
encryption-like, frequency of registry modification, and suspicious PowerShell
activity. In general, the framework provides precise, low-false-positive, and
interpretable ransomware detection that would be used in security operations. |
|
Keywords: |
Ransomware Detection, Explainable AI, Transformer Networks, Behavioural
Analysis, Class Imbalance, Deep Learning, Malware Classification, Cybersecurity |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
ADAPTIVE DENOISING MULTI-VIEW TRANSFORMER FOR SOFTWARE DEFECT PREDICTION WITH
CLASS-BALANCED CROSS-ATTENTION FUSION |
|
Author: |
SRINIVASARAO DHARMIREDDI , DR. M KRISHNA PRASAD , M. KIRAN KUMAR , ARUNA RAO S L
, GNANA DEEPTHI B , A. BAKIYA , ELANGOVAN MUNIYANDY , DR THALAKOLA
SYAMSUNDARARAO |
|
Abstract: |
Existing software defect prediction (SDP) models continue to suffer from noisy
inputs, missing values, skewed class distributions, and weak generalisation on
unseen modules or projects. Conventional classifiers fail to capture minority
defect instances. Moreover, many deep models improve accuracy without jointly
considering quality and imbalance. This paper proposes a novel Adaptive
Denoising Multi-View Transformer that comes with Imbalance-Aware Cross-Attention
Fusion called ADMT-IBF. The system incorporates a preprocessing phase capable of
robust missing-value imputation, outlier suppression, and family-aware
normalisation. It has a multi-view representation that learns complementary
descriptors from raw features, statistical summaries and entropy-based views. It
also has a cross-attention fusion with gated residual connections and an
imbalance-aware output that combines focal loss, class-balanced reweighting and
probability calibration. ADMT-IBF has been evaluated on a publicly available
dataset of 60000-module 22+ SDPs, which includes a code and process metrics set.
The evaluation results include F1=0.868, AUC=0.941 and G-mean=0.889. This result
has outperformed the XGBoost and the CNN-BiLSTM baselines by a margin of
12.0-16.0% in F1 and 4.4-6.0% in AUC. Tests show that using adaptive
preprocessing gives the largest individual gain (+4.1% F1) of all components,
while the proposed focal-class-balanced loss helps recall for the minority class
by 8.4%, without collapsing precision overall. Experiments on nine unseen
systems show strong generalisation, as median F1 > 0.82 on all targeted ones.
The results imply that future SDP systems should jointly model data cleanliness,
feature diversity, and class imbalance to enable more reliable defect triage in
real-world settings of software engineering. |
|
Keywords: |
Software Defect Prediction, Transformer, Multi-View Learning, Cross-Attention,
Class Imbalance, Focal Loss, Data Denoising, Deep Learning |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
ATTACK-AWARE MODEL PREDICTIVE CONTROL FOR SECURE OPERATION OF ISLANDED
MICROGRIDS |
|
Author: |
DR. P. ROHINI, S BASKARAN, NAGADEVI BALA NAGARAM, SRINIVAS BABU N, VASUJADEVI
MIDASALA, NIDAL AL SAID, S. B. G. TILAK BABU |
|
Abstract: |
Distributed energy resources (DERs), power electronic converters and digital
communication infrastructure have made islanded microgrids tightly coupled
cyber-physical systems vulnerable to a variety of malicious cyber-attacks such
as false data injection attacks (FDIA), denial-of-service (DoS) attacks and
replay attacks. In traditional Model Predictive Control (MPC) based frequency
and voltage control for islanded microgrids, it is assumed that the measurements
taken by the sensors and the communication link are reliable, which makes them
highly susceptible to coordinated cyber-attacks that can lead to instability and
cascading failure in the microgrid. A new Attack-Aware Model Predictive Control
(AA-MPC) framework is proposed in this paper, which integrates a real time
residual based anomaly detector and a robust state estimator, allowing
simultaneous detection, isolation and accommodation of attacks. The proposed
formulation extends the classical MPC cost function and constraints in a min-max
robust optimization structure, which explicitly considers the impact of
corrupted measurements, and a sequential detector based on the Cumulative Sum
(CUSUM) identifies the anomalous residuals with low latency. To validate the
proposed approach to FDIA and DoS attack scenarios, a reduced order microgrid
model consisting of three distributed generators, an energy storage system,
critical load, non-critical load, and a communication layer was developed for
the simulation environment of MATLAB/Simulink. The results of the simulation
show that the proposed AA-MPC significantly affected the frequency deviation
peak from 0.42 Hz to 0.075 Hz and the voltage deviation peak from 0.118 p.u. to
0.021 p.u., with a shorter delay time for attack detection compared to the
conventional MPC controller without attack awareness. The proposed scheme also
achieves the true positive rate of above 96% and false positive rate of below 4%
for the tested attack intensities. The results demonstrate that by incorporating
attack detection and resilient estimation into the predictive control
formulation, security and dynamic performance of the islanded microgrid are
significantly improved and no extra hardware investment is necessary.
|
|
Keywords: |
Islanded Microgrid, Model Predictive Control, Cyber-Physical Security, False
Data Injection Attack, Attack Detection, Resilient State Estimation, Robust
Optimization |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
SLGT-NET: AN EXPLAINABLE STACKED TABNET–LIGHTGBM FRAMEWORK FOR LUNG NODULE
CLASSIFICATION FROM CT AND MRI IMAGES |
|
Author: |
SHEIK JAMIL AHMED, VISHWANATH Y., SAIRA BANU ATHAM |
|
Abstract: |
Early and accurate classification of lung nodules plays a crucial role in
improving lung cancer prognosis and significantly reducing patient mortality.
Delays or inaccuracies in diagnosis often lead to late-stage detection, limiting
treatment options and survival rates. To address this challenge, the present
study introduces a robust and explainable radiomics-driven framework named
Stacked Light Gradient TabNet (SLGT-Net) for reliable differentiation between
benign and malignant lung nodules.The proposed framework begins with advanced
image preprocessing techniques designed to enhance image quality, normalize
intensity variations, and suppress noise artifacts commonly present in medical
imaging data. From the preprocessed images, a comprehensive set of handcrafted
radiomic features is extracted, capturing morphological properties, texture
patterns, and micro-texture characteristics that are closely associated with
tumor heterogeneity. These features form an interpretable and clinically
meaningful feature space.SLGT-Net leverages a stacking-based ensemble strategy
that integrates the complementary strengths of TabNet and Light Gradient
Boosting Machine (LightGBM). TabNet employs sequential attention mechanisms to
dynamically select the most informative features, enhancing interpretability and
reducing redundancy. In parallel, LightGBM efficiently models complex nonlinear
relationships within high-dimensional radiomic features. A meta-learner is
subsequently employed to fuse the predictions from both models, further
improving classification robustness and generalization.To promote clinical trust
and transparency, explainable artificial intelligence (XAI) techniques are
incorporated through TabNet attention masks and SHapley Additive exPlanations
(SHAP) analysis. Experimental evaluation on publicly available datasets
demonstrates superior performance, achieving an accuracy of 97.91%, precision of
96.23%, recall of 96.80%, F1-score of 96.51%, and specificity of 96.41%. These
results confirm that SLGT-Net provides a reliable, interpretable, and
high-performing solution for lung cancer detection and clinical decision
support. |
|
Keywords: |
Lung cancer detection, Magnetic resonance imaging, Chromated tomography,
Explainable artificial intelligence, Handcrafted radiomic features |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID QUANTUM CLASSICAL MACHINE LEARNING FOR REAL TIME DIABETIC RETINOPATHY
DETECTION AND TRIAGE |
|
Author: |
B. DEEVENA RAJU, M. ARCHANA, KANAKALA SRINIVAS, NALLA AKHILA, GOWRIPUSHPA
GEDDAM, SRILAKSHMI RAMYA SAKAMUDI, M. CHAITANYA KUMARI, R. PRATHAP KUMAR |
|
Abstract: |
Diabetic retinopathy (DR) continues to be one of the main causes of vision loss
and blindness among people with diabetes, therefore requiring early detection in
order to achieve better clinical outcomes. However, current state-of-the-art DR
detection approaches mainly depend on classical deep learning models, which
suffer from high computational complexity and provide limited support for
automated real-time patient triaging in clinical settings. To address these
challenges, this study presents a Hybrid Quantum–Classical Machine Learning
(HQCML) model for real-time detection of diabetic retinopathy and patient
prioritization with respect to disease severity using retinal fundus images. The
proposed HQCML framework is based on deep convolutional neural network-based
feature extraction and the subsequent classification of the disease using a
variational quantum classifier. Moreover, this study introduces a novel Quantum
Priority Triage Mechanism (QPTM), which divides patients according to the
severity of their condition into three categories—routine, moderate, and urgent
priority levels—thereby enabling efficient clinical decision support. The
experimental results, which have been obtained from analysis of the APTOS 2019
and EyePACS datasets, show that the proposed framework has achieved 97.2%
classification accuracy, outperforming traditional machine learning and deep
learning approaches, such as support vector machines, random forests,
convolutional neural networks, and vision transformers. Furthermore, the
proposed framework has also reached 96.4% precision, 95.8% recall, 96.1%
F1-score, and an AUC of 0.98, demonstrating robust diagnostic performance. Along
with improvements in classification accuracy, the main contribution of this
study is demonstrating the feasibility and effectiveness of integrating hybrid
quantum–classical learning with automated patient triage in a unified framework
for diabetic retinopathy screening. The proposed approach also provides a
scalable foundation for next-generation intelligent eye-screening systems and
highlights the potential of quantum-enhanced artificial intelligence for
real-time clinical decision support. |
|
Keywords: |
Diabetic Retinopathy Detection, Hybrid Quantum–Classical Machine Learning,
Retinal Fundus Image Analysis, Variational Quantum Circuits, Deep Learning,
Automated Clinical Triage |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING CLINICAL NAMED ENTITY RECOGNITION VIA FEW SHOT PROMPT TUNING OF LARGE
LANGUAGE MODELS USING MIMIC III DATA |
|
Author: |
DR. SELVANI DEEPTHI KAVILA, MANIKUMARI ILLA, V N V L S SWATHI, NALLA AKHILA,
ERUKALA MAHENDER, SRIKANTH CHERUKUVADA, JOHN T MESIA DHAS, DR. EDIGA CH. RAMA
TULASI |
|
Abstract: |
In clinical Named Entity Recognition (Clinical NER), structured medical
information can be extracted from unstructured clinical text, but often the lack
of annotated clinical data and the high per-document processing cost of
traditional model fine-tuning approaches hamper effectiveness. To efficiently
and economically solve Clinical NER in low-resource settings, this study
proposes a framework for a Few-Shot Prompt-Tuned Large Language Model
(FSPT-LLM). Its proposed approach combines parameter-efficient soft prompt
tuning, Semantic Diversity Sampling (SDS), and Prompt Consistency Regularization
(PCR) to fine-tune pretrained LLMs on limited labeled clinical text while
keeping the backbone model frozen. Experiments conducted on the MIMIC-III
dataset using 5-, 10-, and 20-shot settings demonstrate that the proposed
framework achieves a precision of 95.6%, recall of 94.8%, F1-score of 95.2%, and
accuracy of 95.5% in the 20-shot setting. The FSPT-LLM outperforms ClinicalBERT
by 2.5 percentage points in F1-score while requiring fewer than 1% of the model
parameters to be trainable. Entity-wise evaluation further demonstrates
consistent performance across disease, medication, procedure, and laboratory
test entities. These findings indicate that combining few-shot learning with
parameter-efficient prompt tuning can provide strong generalization and
computational efficiency for Clinical NER with limited annotated data. The
proposed framework therefore offers a scalable and data-efficient approach for
bridging large language models and practical clinical information extraction
systems. |
|
Keywords: |
Clinical Named Entity Recognition, Few-Shot Learning, Prompt Tuning, Large
Language Models, MIMIC-III, Clinical NLP |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
SYNTHESIS OF A SELF-ORGANIZING NONLINEAR DISSIPATIVE SYSTEM AND
INVESTIGATION OF ITS STABILITY BY THE LYAPUNOV FUNCTION METHOD |
|
Author: |
A. ALIYA MURATKALIYEVA, BEKET KENZHEGULOV, GULZHUZIM MENDIGALIYEVA, MAKHATOVA
VALENTINA, AMANGELDI ARDAK, BAKYTGUL TURSYNOVA, ZHANARGUL KABYLKHAMIT, MAXOT
RAKHMETOV |
|
Abstract: |
This study proposes a novel analytical approach for the synthesis and stability
analysis of a self-organizing nonlinear dissipative dynamical system based on
the Lyapunov function method. A nonlinear mathematical model incorporating
dissipative mechanisms and nonlinear feedback is developed to describe the
formation of stable self-organizing dynamic regimes. An analytical Lyapunov
function is constructed, and sufficient conditions for asymptotic stability are
derived, demonstrating that system stability is guaranteed when the dissipative
coefficients dominate the nonlinear interaction parameters. To validate the
proposed approach, numerical simulations were performed using the fourth-order
Runge–Kutta method in MATLAB/Python. Three system configurations were
investigated: the original model (M0) and two modified models (M1 and M2) with
enhanced dissipation and nonlinear feedback. The simulation results are in good
agreement with the analytical stability conditions derived from the Lyapunov
method. Among the investigated configurations, the M2 model demonstrated the
best dynamic performance, achieving the fastest stabilization time (21-time
units), the lowest stationary Lyapunov function value (0.0007), and the most
pronounced self-organizing behavior through rapid convergence to a stable
attractor. The relative error between the analytical predictions and numerical
simulations remained below 1%, confirming the validity of the proposed
mathematical model. The proposed synthesis methodology provides an effective
analytical framework for designing stable self-organizing nonlinear dissipative
systems with guaranteed Lyapunov stability. The developed approach can be
applied to intelligent control systems, robotics, cyber-physical systems,
adaptive dynamic systems, and intelligent energy networks, where robust
nonlinear stability is a fundamental requirement. |
|
Keywords: |
Nonlinear Dynamical Systems; Lyapunov Stability; Dissipative Systems;
Self-Organization; Nonlinear Differential Equations; Attractor Dynamics;
Nonlinear Control Systems; Phase Space Analysis. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
THE VISTA TRAINING MODEL: A MISSION-BASED VIRTUAL WORLD APPROACH TO 3D SCENE
CREATION |
|
Author: |
PARWAPUN KAMTAB, WEERAKRON CHAREANSUK |
|
Abstract: |
This study developed and evaluated the VISTA Training Model, a mission-based
virtual-world approach to 3D scene creation. A three-phase research and
development process covered model design, development of mission-based
activities and a learning environment implemented in Spatial, and implementation
with learners. The five stages were Virtual Orientation, Identify the Mission,
Survey and Strategize, Three-dimensional Creation, and Assess, Adjust and Apply.
Five experts evaluated the model, activity plans, and virtual environment. The
implementation used a one-group pretest-posttest pre-experimental design with 15
graduate students. The principal investigator and the second author
independently rated participants' pretest performance, posttest performance, and
completed projects using analytic rubrics. Expert ratings were at the highest
interpretation level for the model (M = 4.71), activity plans (M = 4.75), and
virtual environment (M = 4.68). The overall competency score was higher at
posttest (M = 4.45, SD = 0.21) than at pretest (M = 2.85, SD = 0.49), t(14) =
13.48, p < .001, Cohen's dz = 3.48. Inter-rater ICCs ranged from .82 to .91.
Completed-project quality (M = 4.53, SD = 0.28) was strongly associated with
posttest competency, r = .78, 95% CI [.45, .92], p < .001. Learner satisfaction
was rated at the highest level (M = 4.80, SD = 0.40). Within this one-group
sample, posttest competency scores were higher, completed projects were rated
highly, and participants reported high satisfaction. These findings provide
preliminary evidence of the model's feasibility as a mission-based virtual-world
approach to 3D scene creation; they do not establish causal effectiveness. |
|
Keywords: |
3D scene creation, immersive learning technology, mission-based learning,
virtual world learning environment, 3D content production |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
A TRANSFER LEARNING APPROACH FOR EARLY DETECTION OF FEBRILE SEIZURES USING EEG
AND CLINICAL FEATURES |
|
Author: |
KANAKALA SRINIVAS, M ARCHANA, SAJJA SUSHUMNA, DIDDY VEERA CHAKRAM, CH SIVA RAMA
KRISHNA, P. LAKSHMI PRASANNA, B. NANCHARAIAH,
MALAPATI NARESH |
|
Abstract: |
Predicting febrile seizures in pediatric patients remains challenging due to the
scarcity of annotated pediatric electroencephalogram (EEG) data, the complex
temporal characteristics of EEG signals, and the limited integration of
electrophysiological and clinical information in existing prediction approaches.
This paper presents a novel deep learning framework, termed the Hybrid Transfer
Learning Multimodal Attention Network (HTMAN), for the early prediction of
seizure-related preictal patterns in the context of febrile-seizure risk
assessment. The proposed framework employs an ImageNet-pretrained ResNet-18
backbone, fine-tuned on Short-Time Fourier Transform (STFT)-based EEG
spectrograms, to extract high-level spatial and spectral representations. A
Bidirectional Long Short-Term Memory (Bi-LSTM) network subsequently models
temporal dependencies in the extracted EEG features, while an attention-based
multimodal fusion mechanism integrates EEG representations with structured
febrile-seizure-related clinical features. Experimental evaluation demonstrates
that HTMAN achieves an accuracy of 93.2%, an F1-score of 0.92, and an area under
the receiver operating characteristic curve (AUC) of 0.95 for distinguishing
seizure-related preictal and interictal patterns. The framework further
demonstrates the ability to discriminate preictal EEG activity at prediction
horizons of up to 5 minutes before annotated seizure onset. The principal
contribution of this work is a unified transfer-learning-based multimodal
architecture that combines EEG feature learning, temporal sequence modeling,
clinical-feature integration, and attention-based feature weighting under
limited annotated EEG conditions. Importantly, the present study constitutes a
proof-of-concept investigation: the EEG recordings are used to learn
seizure-related preictal representations and are not assumed to represent
clinically confirmed febrile-seizure events. Therefore, the reported findings
demonstrate the feasibility of the proposed transfer-learning and multimodal
framework rather than direct clinical validation of febrile-seizure prediction.
Future validation using synchronized EEG recordings and clinical measurements
from pediatric patients with clinically confirmed febrile seizures is required
to establish the clinical effectiveness and generalizability of the proposed
framework. |
|
Keywords: |
Febrile Seizure Prediction, EEG Analysis, Transfer Learning, Multimodal Deep
Learning, Attention Mechanism, Early Detection |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
GENERATIVE ARTIFICIAL INTELLIGENCE FOR DESIGNING NON-FUNGIBLE TOKENS (NFTS)
BASED ON BRAND IDENTITY FOR DIGITAL ASSETS IN HIGHER EDUCATION |
|
Author: |
SAOWALUCK BOONRAWD, KANYAMAAT SANGOUNSAK , PANITA WANNAPIROON , PRACHYANUN
NILSOOK |
|
Abstract: |
The rapid advancement of generative artificial intelligence (AI) and
blockchain technologies has transformed digital content creation and value
generation in the contemporary digital economy. In higher education, these
technologies create new opportunities to enhance student competencies through
experiential learning and digital asset development. This study aimed to
synthesize theoretical concepts and related research, develop a conceptual
framework, and evaluate the suitability of a generative AI–based framework for
designing non-fungible tokens (NFTs) grounded in brand identity for digital
assets in higher education. The research was conducted in three phases: (1)
synthesis of literature and theoretical foundations related to generative AI,
brand identity, NFTs, and student competencies; (2) development of a conceptual
framework integrating technological processes, creative design, and branding
strategies; and (3) evaluation of the framework’s suitability by seven experts
in artificial intelligence, computer art, and branding using purposive sampling.
The proposed framework comprises five interrelated components: brand identity
input, generative AI processes, AI-generated outputs and student competencies,
evaluation mechanisms, and NFT marketplace application with feedback loops. The
findings indicate that the framework provides a systematic guideline for
integrating generative AI and NFT design into higher education, supporting
digital creativity, AI literacy, strategic branding, and digital value creation. |
|
Keywords: |
Generative artificial intelligence, Non-fungible tokens (NFTs), Brand identity,
Student competencies, Higher education |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
XDPQ-PTS-AODV: AN EXPLAINABLE XGBOOST AND DEEP Q-NETWORK ENABLED POST-QUANTUM
TRUST-BASED SECURE ROUTING FRAMEWORK FOR MOBILE AD HOC NETWORKS |
|
Author: |
SINGIREDDY SATEESH REDDY , E.ARAVIND |
|
Abstract: |
Mobile Ad Hoc Networks (MANETs) are highly vulnerable to routing attacks due to
their decentralized architecture, dynamic topology, and lack of centralized
administration. Traditional trust-based routing approaches often suffer from
static trust evaluation, limited adaptability to dynamic attacks, and lack of
explainability. To address these limitations, this paper proposes XDPQ-PTS-AODV,
an Explainable XGBoost and Deep Q-Network Enabled Post-Quantum Trust-Based
Secure Routing Framework for MANETs. The proposed framework integrates
multi-metric trust computation, XGBoost-based malicious node detection, SHAP
explainability, DQN-based priority time-slot scheduling, and post-quantum route
authentication using CRYSTALS-Kyber and CRYSTALS-Dilithium. The protocol was
implemented and evaluated in NS-3 under Blackhole, Grayhole, and Combined attack
scenarios with network sizes ranging from 20 to 200 nodes. Experimental results
demonstrate that XDPQ-PTS-AODV achieves up to 95.0% Packet Delivery Ratio (PDR),
162.55 kbps kbps throughput, 90.8% packet loss reduction, and 79.9% delay
reduction compared with conventional AODV under combined attack conditions. The
XGBoost classifier achieves 98.5% accuracy, 98.1% precision, 97.8% recall, and
97.9% F1-score for malicious node detection. The results confirm that
integrating explainable AI, reinforcement learning, trust management, and
post-quantum security significantly enhances routing reliability and security in
MANET environments. |
|
Keywords: |
Post-Quantum Security, Trust Routing, Explainable AI, MANET, AODV, XGBoost, SHAP
and DQN |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
CROSS-DOMAIN EMOTION DETECTION USING LIGHTWEIGHT LEXICON-GUIDED FEATURE FUSION:
A THREE-DATASET STUDY WITH EXPLAINABILITY AND CONTRASTIVE ALIGNMENT ANALYSIS |
|
Author: |
. HEMAKIRTHIGA ,J. ARUNADEVI |
|
Abstract: |
Emotion detection models trained on one social media platform generalize poorly
to another, yet this cross-domain generalization problem receives comparatively
little attention relative to within-dataset benchmarking. We study it directly:
we propose a lightweight, fully CPU-trainable architecture that fuses TF-IDF/SVD
text embeddings with NRC Emotion Lexicon association features through a learned
gate, and use it as the common architecture for a systematic three-dataset
cross-domain generalization study spanning tweets (SemEval-2018 E-c), Reddit
comments (GoEmotions), and personal narrative essays (ISEAR). Our central,
statistically robust finding is that simple joint multi-domain training closes
most of the cross-domain generalization gap relative to zero-shot single-domain
transfer (bootstrap 95% CI for the Macro-F1 gain on GoEmotions: [0.095, 0.129]).
As one specific, rigorously tested question within this study, we ask whether
supervised contrastive alignment can improve on this strong, simple
joint-training baseline; despite a dedicated hyperparameter search, a
memory-bank negative pool, a stricter Jaccard-based positive-pair criterion,
higher representation capacity, and evaluation across all three test domains, we
find no configuration that statistically significantly outperforms naive pooling
a considered negative result for this specific mechanism, not a limitation of
the cross-domain generalization finding itself, which stands independently. We
additionally report within-domain results against five baselines (our fusion
model is competitive with, but does not exceed, a strong TF-IDF+SVM baseline we
position the architecture explicitly as a vehicle for generalization and
interpretability rather than raw within-domain accuracy), a real measured
computational-efficiency comparison, an ablation testing whether the learned
gate's near-constant empirical behavior (std = 0.012) means its adaptivity is
dispensable (it is not: a fixed-scalar-gate ablation performs measurably worse,
p = 0.0625), a multi-head gating variant, model-intrinsic explainability (gate
reliance, SHAP over the interpretable lexicon channel), and a systematic
label-confusion error analysis. |
|
Keywords: |
Cross-Domain Generalization; Emotion Detection; Social Media NLP; Lexicon
Feature Fusion; Supervised Contrastive Learning; Multi-Label Classification;
Explainable AI; Error Analysis |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
MEDGUARD: A GRAPH-AUGMENTED, FEDERATED AND EXPLAINABLE AI FRAMEWORK ON
BLOCKCHAIN FOR HEALTHCARE INSURANCE FRAUD DETECTION |
|
Author: |
SWAPNA DONEPUDI, NIRMALA DEVI K , K VENU GOPAL3,PATHAPATI SAROJA, SWATHI
VODDI5,PRASAD DEVARASETTY,DESHINTA ARROVA DEVI |
|
Abstract: |
Healthcare insurance remains burdened by claim fraud, slow manual adjudication,
and limited trust in centralized data handling. This paper presents an updated
MedGuard architecture that moves beyond tabular ensemble classifiers to a
graph-augmented, federated, and explainable fraud-detection pipeline anchored on
a permissioned blockchain. Claims are represented as a patient–provider network;
centrality and community-detection features extracted from this graph are fused
with claim-level features and fed into a hybrid SVM + Gradient-Boosted-Trees
classifier. On a 5,000-record synthetic claims benchmark built for this study,
the graph-augmented model reaches 93.6% accuracy, 75.7% precision, 70.7% recall,
and an AUC of 0.885 which shows a substantial improvement over the same hybrid
classifier without graph features (85.0% accuracy, AUC 0.782) and over
standalone Logistic Regression, Decision Tree, and Random Forest baselines
(71.5–87.7% accuracy). A federated-averaging simulation across three synthetic
hospital clients reaches 73.1% accuracy without any client sharing raw claim
records, closely tracking a centralized model trained on the pooled data
(71.4%), supporting federated learning as a viable privacy-preserving
alternative. Claim cost estimation via a Random Forest regressor achieves R² =
0.810 (RMSE ≈ ₹14,742). Per-claim Shapley-value explanations are computed
directly to justify individual fraud flags, and a zero-knowledge-proof
verification step is proposed to let smart contracts confirm claim eligibility
without exposing medical details on-chain. Verified claims and documents
continue to be anchored via smart contracts on a permissioned blockchain, with
supporting files encrypted and stored on IPFS. |
|
Keywords: |
Blockchain, Healthcare Data Security, Graph-Based Fraud Detection, Federated
Learning, Explainable AI, Process Innovation, Financial Access, IPFS,
Zero-Knowledge Proofs |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
BLOCKCHAIN-ENABLED-DEEP-LEARNING-FOR-REAL-TIME-FDI-DETECTION-IN-VEHICULAR-CLOUD-NETWORKS |
|
Author: |
HANI AL-BALASMEH |
|
Abstract: |
Vehicular Cloud Networks (VCNs) support supercritical services through
continuous data exchange among vehicles, Roadside Units (RSUs), edge nodes, and
cloud services, but their openness makes them vulnerable to False Data Injection
(FDI) attacks. Existing studies typically emphasize either intelligent attack
detection or trustworthy evidence management; consequently, they do not fully
address the combined requirements of real-time classification, tamper-resistant
verification, and adaptive response. This paper proposes a blockchain-enabled
CNN-LSTM framework that analyzes mobility, temporal, communication, and
consistency-based features and then records compact message hashes and detection
metadata on a distributed ledger. Smart contracts validate the transaction
structure, verify message integrity, update vehicle trust scores, and trigger
restriction actions when repeated malicious behavior is observed. The
methodological contribution is a single detection-verification-trust pipeline
that connects feature engineering, sequence-based deep learning, hash-only
on-chain evidence, and trust-based response. On the reported held-out test set,
the framework achieved 98.42% accuracy, 98.11% precision, 98.73% recall, a
98.42% F1-score, and a 1.39% false positive rate. The CNN-LSTM classifier
required 12.6 ms per inference, while the complete preprocessing-to-confirmation
path required 73.4 ms; 10,000 compact blockchain transactions required 8.2 MB.
Comparative and ablation analyses show that engineered temporal features improve
detection, the hybrid CNN-LSTM outperforms conventional and standalone models,
and blockchain adds traceability and evidence integrity with measurable latency.
The findings provide operational evidence that an integrated, metadata-oriented
design can strengthen cyber-resilience and data trustworthiness in real-time VCN
security, subject to the stated simulation and deployment limitations. |
|
Keywords: |
Vehicular Cloud Networks, False Data Injection, CNN-LSTM, Blockchain
Verification, Trust Management |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
RESOLVING STATIC ROLE BOTTLENECKS IN MULTI-AGENT WORKFLOW ORCHESTRATION THROUGH
DYNAMIC ALLOCATION: THE CODA FRAMEWORK |
|
Author: |
HICHAM SADIKI , RAJAE ZRIAA , SAID AMALI |
|
Abstract: |
Multi-agent workflow orchestration suffers from a fundamental limitation: agents
are assigned fixed roles at initialization, creating bottlenecks under variable
load and paralysis when agents fail. This paper proposes CoDA, a fully
decentralized framework that enables autonomous real-time role negotiation
through a hybrid auction-consensus protocol. A utility function integrates
capability affinity, execution time, load balance, and communication latency to
guide allocation. Experiments on the Mathematical Workflow Benchmark 2026 (200
tasks, five domains) demonstrate that CoDA reduces workflow completion time by
26.8 percent over static baselines, recovers from failures in under 5.2 seconds,
and keeps communication overhead below 5.2 percent. These results establish that
dynamic role allocation is both practical and essential for scalable, resilient
multi-agent orchestration in enterprise and scientific computing. |
|
Keywords: |
Multi-Agent Systems; Dynamic Role Allocation; Workflow Orchestration;
Distributed Consensus; Mathematical Computing; Contract Net Protocol |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
AN ADAPTIVE WHALE OPTIMIZATION AND BAYESIAN CBIGRU-BL FRAMEWORK FOR IOT -BASED
DIABETICS AND CARDIOVASCULAR RISK PREDICTION |
|
Author: |
SHIVAKUMAR DARIPALLY , Dr. N. KRISHNA KUMAR |
|
Abstract: |
Diabetes and cardiovascular disease together account for a large share of
preventable deaths worldwide, and both are increasingly tracked through Internet
of Things (IoT) wearable and point-of-care devices that stream physiological and
biochemical readings continuously. Converting these streams into actionable risk
predictions requires a model that can narrow noisy, high-dimensional sensor data
to the indicators that matter and report a calibrated confidence alongside its
prediction, since a confidently wrong output carries real cost in healthcare.
This paper proposes an IoT-oriented risk-prediction framework built on two
contributions. The first, an Adaptive Diversity-guided Whale Optimization
Algorithm (ADWOA), replaces the fixed, iteration-only convergence schedule of
standard WOA with an Adaptive Diversity Factor computed from the spread of
population fitness values, letting the exploration-exploitation balance track
the actual search state rather than a pre-set linear decay; ADWOA performs
weighted feature selection across the clinical/IoT indicator set. The second, a
Deep Convolutional Bidirectional Gated Recurrent Unit with Bayesian Learning
(CBiGRU-BL) network, chains 1D-CNN feature extraction, BiGRU sequence modelling,
and a variational Bayesian output layer yielding a risk probability with an
epistemic uncertainty estimate, both stages tuned by ADWOA. The pipeline was
evaluated on the Pima Indians Diabetes dataset (768 records, 8 indicators) and
the UCI Cleveland Heart Disease dataset (303 records, 13 indicators), using
5-fold cross-validation with 15 independent optimizer runs. ADWOA converged
faster than standard WOA and surpassed its mean/best fitness on Diabetes (0.4839
vs. 0.4816 mean; 0.4961 vs. 0.4895 best), closely matched it on Heart, and
showed lower run-to-run variance than PSO and GWO on both datasets. With
ADWOA-selected features, CBiGRU-BL reached 71.2% accuracy/0.800 AUC-ROC on
Diabetes and 78.6% accuracy/0.862 AUC-ROC on Heart, competitive with a
non-Bayesian CNN-BiGRU counterpart while supplying a per-prediction uncertainty
estimate unavailable from any deterministic baseline. Robustness under feature
noise, computational cost, and deployment considerations are also examined. |
|
Keywords: |
IoT-Based Healthcare, Diabetes Risk Prediction, Whale Optimization Algorithm,
Bayesian Deep Learning, Bidirectional Gated Recurrent Unit |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
HYBRID DEEP LEARNING ARCHITECTURES FOR TEXT CLASSIFICATION: A TAXONOMY AND
CRITICAL SURVEY |
|
Author: |
D.JASMINE GUNASUNDARI, J. ARUNADEVI |
|
Abstract: |
Text classification underlies applications such as sentiment analysis, news and
topic categorization, misinformation detection, biomedical document triage, and
spam or smishing filtering. Convolutional neural networks (CNNs), recurrent
architectures (LSTM/GRU), attention mechanisms, and pretrained transformer
language models such as BERT and RoBERTa have each individually advanced the
state of the art, yet no single family dominates across accuracy, robustness to
short or noisy text, computational cost, and interpretability. A substantial and
fast-growing line of research instead combines two or more of these components
into hybrid deep learning architectures. This survey applies a PRISMA-informed
search and screening protocol across arXiv, the ACL Anthology, IEEE Xplore,
ScienceDirect, SpringerLink, PubMed Central, PLOS ONE, and Nature Scientific
Reports, yielding 63 hybrid-architecture studies (2014-2025) that meet explicit
inclusion criteria, alongside 20 background and methodological references (83
references in total). From these we construct a five-family taxonomy — CNN-RNN
hybrids, attention-augmented hybrids, pretrained-language-model hybrids,
graph-augmented/ensemble hybrids, and LLM-era/parameter-efficient hybrids — and
we compare them along reported accuracy, computational overhead, robustness, and
interpretability rather than cataloguing architectures in isolation. We report a
quantitative synthesis of accuracy deltas and dataset dependencies drawn
directly from the included studies; a dedicated treatment of explainability
methods (attention analysis, LIME, SHAP, Integrated Gradients) and the ongoing
debate over attention's faithfulness as an explanation; and a discussion of the
computational and energy cost that hybridization adds on top of an already large
pretrained backbone. We identify recurring weaknesses — inconsistent baselines,
limited robustness and cross-lingual evaluation, underreported computational
cost, and interpretability claims that are rarely evaluated as rigorously as
accuracy — and propose a research agenda centred on parameter-efficient hybrid
fine-tuning, standardized cost-aware benchmarking, and explainability-by-design.
We close with an explicit statement of this survey's own methodological
limitations. |
|
Keywords: |
Text classification, Hybrid deep learning, Taxonomy, BERT, Convolutional neural
networks, Attention mechanism, Parameter-efficient fine-tuning, Explainability,
Natural language |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
NEXUS: EVENT-DRIVEN DEPENDENCY-AWARE SCHEDULER FOR BURSTY MICROSERVICE WORKLOADS |
|
Author: |
MITULKUMAR RAJ , PARAG SANGHANI, NIRAJ SHAH |
|
Abstract: |
Microservices workloads experience 10-20x traffic spikes within seconds, and
hopping between nodes slows response times, so the Kubernetes scheduler must
place pods quickly while keeping related services nearby. Existing schedulers
address only one side of this tradeoff. Gang schedulers like Volcano incur
continuous CPU overhead when idle, but network-aware frameworks lack runtime
dependency detection and adaptive activation. In this paper, we introduce NEXUS,
an event-driven Kubernetes scheduler extender that combines a spike detection
engine integrated with Prometheus, a dependency graph engine that creates
runtime DAGs from pod annotations, a gang lifecycle manager enforcing the
temporary coordination, and a locality-aware dynamic node scoring system.
Evaluating a 6-node AWS EKS cluster using the Google Online Boutique benchmark
with 20x Flash, NEXUS reduces P95 latency by 8.1% compared to the default
scheduler, 46.8% compared to YuniKorn, achieves a colocation ratio of 0.75,
eliminates scheduling errors, reduces control plan overhead 22x compared to
Volcano. |
|
Keywords: |
Kubernetes Scheduling, Container Orchestration, Microservice Architecture,
Cloud-native System, Gang Scheduling, Burst workloads. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
A ROBUST AND ADAPTIVE ROUTING ALGORITHMS FOR ENHANCING COMMUNICATION EFFICIENCY
IN DENSE AND SPARSE VANET ENVIRONMENTS |
|
Author: |
SUDHEER NIDAMANURI, C. RAMESH KUMAR |
|
Abstract: |
Vehicular Ad hoc Networks (VANETs) form the communication backbone of
Intelligent Transportation Systems (ITS), enabling vehicle-to-vehicle (V2V) and
vehicle-to-infrastructure (V2I) exchange of safety, traffic, and infotainment
data. Routing performance in VANETs is highly sensitive to network density:
dense urban traffic causes broadcast storms and packet collisions, while sparse
rural or highway scenarios suffer frequent link breakages and network
partitioning. Existing topology-based (AODV, DSR) and position-based (GPSR)
protocols are optimized for a single density regime and degrade sharply outside
it. This paper proposes a robust, density-adaptive routing framework that
dynamically switches between a cluster-based greedy forwarding strategy for
dense regions and a delay-tolerant store-carry-forward strategy for sparse
regions, guided by a lightweight density estimation module and a
machine-learning-based link-quality predictor. Using two Kaggle datasets (VANET
Real-Time Route Optimization and VANET Traffic Congestion), the framework was
benchmarked against AODV, DSR, GPSR, and ten recent adaptive routing approaches.
Results show the proposed framework achieves a Packet Delivery Ratio of 95.8%
and 93.9%, end-to-end delay of 36 ms and 41 ms, routing overhead of 15% and 16%,
and throughput of 985 kbps and 945 kbps on two datasets respectively —
consistently outperforming all baselines across both dense and sparse regimes. |
|
Keywords: |
Vehicular Ad hoc Networks, Routing Protocols, Dense Network, Sparse Network,
Density-Adaptive Routing, Intelligent Transportation Systems, Packet Delivery
Ratio |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
DEEP CONVOLUTIONAL AUTOENCODERS WITH ADAPTIVE GATING FOR HIGH-FIDELITY ECG
DENOISING UNDER SEVERE NOISE SATURATION: A COMPARATIVE STUDY |
|
Author: |
EIHAB BASHIER, TOUFIK BEN JABEUR, HASSAN AL REASI, HAMAD AL REESI, TAISER TAHA |
|
Abstract: |
Electrocardiogram (ECG) signals are frequently corrupted by strong clinical
artifacts like muscle noise and baseline wander, rendering automated diagnostic
systems ineffective. While deep learning models filter signals successfully,
preserving critical physiological features like R-peaks under high noise
saturation remains a major challenge. We propose a new data-driven, two-stage
deep learning framework for robust ECG denoising across a wide range of
signal-to-noise ratios (SNRs), from $-20$ dB to $+20$ dB. The system features a
built-in engine for morphological parameter minimization using analytical
inverse-variance scaling to balance mean squared error against first- and
second-order derivative penalties. Additionally, an empirical-adaptive shortcut
gating mechanism dynamically regulates internal layer transitions based on the
real-time detected noise floor. We present a comprehensive comparative study
evaluating three architectural configurations combined with a Stage 2 Deep
Residual Refinement network: Stacked Convolutional Autoencoders (Stacked
ConvAE), Base Convolutional Autoencoders (Base ConvAE), and U-Net Convolutional
Autoencoders (U-Net ConvAE). Models were trained on true ECG data from the
MIT-BIH Arrhythmia Database synthesized with authentic clinical noise profiles
from the Noise Stress Test Database (NSTDB). Experimental results demonstrate
strong performance across all noise regimes. In an extreme $-20$ dB noise
environment, the U-Net ConvAE framework recovers the signal to a post-refined
SNR of $+10.16$ dB and a morphological correlation coefficient of 0.93.
Crucially, all tested autoencoders successfully restore the cardiac peak
detection rate from a failing 0.53 to a highly reliable 0.90. |
|
Keywords: |
ECG Signal Denoising, Convolutional Autoencoders, Deep Residual Networks,
Clinical Feature Preservation, Automated QRS Detection. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING CONTEXT-AWARE DIALOGUE SYSTEMS VIA MULTI-MODAL FUSION AND HIERARCHICAL
SELF-ATTENTION |
|
Author: |
Dr CHINTAKINDI SRINIVAS , VIJAYAGANTH R , Dr B.RANGA SWAMY , Dr SRAVANI
KOPPURAVURI , Dr HARI JYOTHULA , Dr SUBBA RAO POLAMURI, Dr N. NEELIMA , KADIYALA
SUDHAKAR |
|
Abstract: |
This paper introduces a novel approach to context-aware dialogue systems through
the integration of multi-modal fusion architecture and hierarchical
self-attention mechanisms. The proposed model, Multi-Modal Hierarchical
Attention Network (MMHAN), addresses critical challenges in modern
conversational AI by effectively incorporating visual, textual, and historical
context cues to generate more coherent and contextually appropriate responses.
Unlike prior works that treat modalities independently or apply flat attention
across dialogue history, MMHAN introduces a three-level hierarchical attention
mechanism combined with a cross-modal fusion component. Experiments on three
benchmark datasets (MultiWOZ 2.1, Ubuntu Dialogue Corpus, and Visual Dialog)
demonstrate that MMHAN outperforms state-of-the-art baselines by 7.2% on
response relevance and 9.1% on context retention. Human evaluation shows MMHAN
responses were preferred in 76.3% of comparisons. A novel metric, Cross-Modal
Coherence Score (CMCS), correlates strongly with human judgments (r=0.82). With
only 15% additional parameters over text-only models, MMHAN is efficient and
suitable for real-world deployment. |
|
Keywords: |
Multi-Modal Fusion, Hierarchical Self-Attention, Context-Aware Dialogue,
Cross-Modal Reasoning, Transformer Architecture, Response Generation,
Conversational AI |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
AN EXPLAINABLE AI-ASSISTED NLP FRAMEWORK FOR PERSONALIZED ENGLISH GRAMMAR
INSTRUCTION |
|
Author: |
PASUPULETI VENKATA RAMANA, B. NEELAMBARAM, DR. D. VIJAYALAKSHMI, PHANIMALA
THIRAGATI, DR. K. K. SUNALINI, DR. BOLIGARLA MURALIKRISHNA,
DR. BHUVANESWARI
PAGIDIPATI, A. RAMAPRATHAP REDDY |
|
Abstract: |
The process of correcting the grammar of the English language is another vital
aspect that requires the provision of accurate English grammar teaching and
training for improvement in learners' writing skills. Nonetheless, traditional
approaches for correcting grammatical errors are not always efficient in
comprehending the context of the texts and providing interpretability. In this
study, we have proposed an Explainable Artificial Intelligence (XAI)-assisted
Natural Language Processing (NLP) approach for personalizing English grammar
training. The approach under investigation includes the First Certificate in
English (FCE) dataset along with the following stages: text preprocessing,
contextual feature extraction with the help of BERT, grammar error detection,
personalized feedback generation, and the attention-based explainability module.
The accuracy of the proposed approach has been calculated on the basis of such
metrics as Accuracy, Precision, Recall, F1-score, and Area Under the Curve
(AUC). As it has been demonstrated by experimental results, the proposed
approach shows excellent performance because it achieves Accuracy of 97.48%,
Precision of 97.21%, Recall of 96.93%, F1-score of 97.07%, and AUC of 99.00%,
significantly outperforming CNN, LSTM, Bi-LSTM, and Transformer-based
approaches. The explanation module demonstrates the linguistic features
responsible for prediction decisions and therefore increases the transparency of
the model and learner confidence. Thus, the proposed approach offers
interpretable, personalized, and accurate grammar feedback and can be used as an
intelligent language learning framework, automated writing assistance system,
and next-generation educational technology solution. |
|
Keywords: |
Explainable Artificial Intelligence (XAI); Natural Language Processing (NLP);
English Grammar Error Detection; BERT; Personalized Learning; Grammar Feedback. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
UUC: UNIFORMLY UTILISED LAST LEVEL CACHE OF MODERN MULTICORE PROCESSORS |
|
Author: |
PURNENDU DAS, BISHWA RANJAN ROY, BARGA DEORI |
|
Abstract: |
A set-associative last-level cache (LLC) can have free or weakly reused lines
and still suffer frequent misses. The reason is simple: each address is confined
to one set, whereas application footprints are rarely distributed uniformly over
all sets. Existing flexible-placement caches improve this situation, but
typically require extra tags, relocation chains or fixed set partnerships. This
paper presents Uniformly Utilised Cache (UCC), a two-choice LLC organisation for
multicore processors. Every block has a conventional home set and one
tag-derived remote set. A fill is redirected only when the home set is under
sustained pressure and the remote set is lightly occupied. The normal home
lookup is unchanged; the remote set is searched sequentially and only when a
displacement counter indicates that a remote copy may exist. Per-core insertion
budgets prevent one application from occupying an excessive number of borrowed
lines. At a 2 MB/core LLC budget, UCC improves single-core IPC by 6.42% on
average and four-core weighted speedup by 8.50% over a conventional SRRIP cache.
It outperforms modeled V-Way, ZCache, SBC, and FS-DAM configurations while
requiring 2.72% metadata for a 4 MB, 16-way LLC. |
|
Keywords: |
Last-Level Cache, Multicore Processor, Cache Utilisation, Alternative Indexing,
Conflict Miss. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
DEEP LEARNING DRIVEN CLIMATE CHANGE PREDICTION USING SPATIO TEMPORAL SATELLITE
DATA FUSION FOR REAL TIME ENVIRONMENTAL MONITORING |
|
Author: |
G. BALU NARASIMHA RAO, ALURI LAKSHMI², U. SADHANA, Y. Y. S. MANOGNA, DEVADI
GANESH, P. LAKSHMI PRASANNA, KARNATAPU LEELA KRISHNA, MANGALAGIRI SRIKANTH KUMAR |
|
Abstract: |
It is essential to predict climate change in real-time and with high accuracy to
minimize environmental risks and facilitate decision-making towards
sustainability. Nevertheless, current models used to predict climate changes
struggle to process heterogeneous spatio-temporal satellite data in order to
ensure high prediction accuracy and low latency needed for real-time
environmental monitoring. In this study, the new Deep Learning-driven
Spatio-Temporal Satellite Data Fusion (DL-STF) approach will be developed by
integrating various sources of satellite data, including MODIS, Sentinel-2, and
ERA5, into a hybrid architecture combining Convolutional Neural Networks (CNNs)
for spatial feature extraction, Bidirectional Long Short-Term Memory (Bi-LSTM)
networks for temporal modeling, and a multi-head attention-based fusion
mechanism for adaptive feature weighting. The novel approach described above can
be regarded as the main research contribution due to its ability to jointly
learn spatial and temporal representations through an attention-driven fusion
strategy to improve climate prediction performance. The experimental analysis
demonstrated the accuracy of the model equal to 96.3% while the error rate was
relatively low (RMSE: 0.18, MAE: 0.12), and the latency was minimal (95 ms)
compared to conventional CNN, LSTM, and CNN-LSTM models. As a result, it can be
concluded that the novel attention-based spatio-temporal fusion framework
significantly increases predictive accuracy, robustness, and computational
efficiency, enabling scalable real-time environmental monitoring for early
warning systems, climate risk assessment, and data-driven policy planning to
strengthen climate resilience. |
|
Keywords: |
Climate Change Prediction, Spatio-Temporal Data Fusion, Deep Learning, Satellite
Remote Sensing, Attention Mechanism, Real-Time Environmental Monitoring |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
BIOFUSION-YOLO: AN INTELLIGENT DEEP LEARNING FRAMEWORK FOR MORPHOLOGICAL AND
COLOR-BASED DETECTION OF NUTRIENT DEFICIENCIES IN BANANA LEAVES |
|
Author: |
N. LAKSHMI KALYANI , KOLLA BHANU PRAKASH |
|
Abstract: |
Properly diagnosing nutrient deficiencies in the crop is essential in order to
maximize fertilizer use, improve agricultural productivity and to sustain
farming. Most of these methods used for soil nutrient assessment (primarily
chemical laboratory-based) are more time consuming, expensive, and can't be
conducted real-time. This paper proposes a new method named MorpBioFusion-YOLO
that addresses these problems using the morphological and color-based deep
learning in RGB images obtained by the field. To achieve this, we have proposed
a novel method, called MorpBioFusion-YOLO, which at the same time tackles the
problem of autonomous classification of nutrient deficiency, based on the deep
learning in both morphological and color components from RGB images captured in
the field. The suggested framework includes feature extraction at multiple
scales, adaptive morphological representation learning using colour-sensitized
spatial analysis functions to find nutrient stress pattern of Boron, Calcium,
Iron, Magnesium, Manganese, Zinc, Potassium, Sulphur deficiency and Healthy leaf
conditions. From a technical perspective, it predominantly uses an adapted
detection backbone composed of YOLO and allows for a path aggregation system
that performs detailed fusion operations in the various features,
notwithstanding varying illumination, orientation and background conditions
across the entire image. Augmentation strategies and texture preserving
transformations are employed during training to increase the ability of
generalization. Our proposed model performance on the banana nutrient deficiency
plants with labels yields high overall mean Average Precision (mAP@0.5): 81.7%
with the extreme accuracy in identifying Healthy and Sulphur classes 0.995 and
good performance in other classes as Boron deficiency, Potassium deficiency and
Manganese deficiency. Moreover, the presented framework is able to provide
effective real-time inference performance, ideal for edge-supported smart
farming applications. Experimental studies showed that morphological structures
of leaves, venation characteristics, chlorosis pattern and variation in the
level of pigmentation are reliable traits for the estimation of nutrient stress.
The proposed BioFusion-YOLO framework is intelligent agricultural solution for
realizing early nutrient diagnosis and precision crop management that is
effective, low in cost and scalable. |
|
Keywords: |
Precision Agriculture, Nutrient Deficiency Detection, Banana Leaf Analysis,
Morphological Feature Fusion, Explainable Deep Learning, RGB Plant Phenotyping,
Multi-Scale Object Detection, Smart Farming, Leaf Color Analysis, Agricultural
Computer Vision and Edge AI in Agriculture |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
ENHANCING CLOUD NETWORK INTRUSION DETECTION USING SMOTE, TRANSFORMER ENCODER AND
ADVERSARIAL HARDENING |
|
Author: |
SAMRAT KRISHNA GADDAM, A. SRINAGESH |
|
Abstract: |
As the utilization of Cloud Computing is growing rapidly it encounters more
challenging and evolving cybersecurity threats and zero-day attacks. In order to
detect intrusion detection more Intelligent and adaptive deep learning models
are crucial. With the goal to identify anomalies in cloud network traffic, this
study compares three sequential learning architectures the conventional LSTM,
its bidirectional extension (Bi LSTM), and Projected LSTM Classifier (PLC) and
proposes a novel Transformer Encoder with Adversarial Hardening
(Transformer+AT). Two popular datasets, CICIDS2018 and CIC-IoT2023, were used
for the experiments. During preprocessing, the Synthetic Minority Over-sampling
Technique (SMOTE) was applied to concentrate on the extreme class imbalance that
can be prevalent in real-world cloud data. Amid the baselines models PLC has
acheieved an accuracy of 96% on CICIDS2018 but has significantly dropped to 82%
on CIC-IoT2023. Show cases a overview gap which is addressed by the proposed
model.
The proposed Transformer+AT achieve 96.05% on CICIDS2018 and 96.1% on
CIC-IoT2023 shows significant stable performance among the two datasets
different from PLC as it has dropped from 96% to 82% on CIC-IoT2023. Decisively
Transformer+AT shows outstanding adversarial robustness by 0.06% drop of
accuracy under FGSM attack on CICIDS2018 and 8.39% on CIC-IoT2023 when compared
to the 20-25% drop for baseline models. These findings disclose that when
Transformer architectures with SMOTE and adversarial hardening are combined,
they improve generalization and robustness among varied environments of Cloud
and IoT. |
|
Keywords: |
LSTM, Bi-LSTM, Projected LSTM, Transformer, Adversarial Training, FGSM, Cloud
Security, Intrusion Detection |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
A HYBRID QUANTUM CLASSICAL FRAMEWORK FOR CROSS DOMAIN TRANSFER LEARNING WITH
REAL TIME ADAPTIVE MULTI TASK LEARNING |
|
Author: |
CHEEMALADINNE VENGAIAH, ROOPADEVI BOSUKONDA, DR. B. RANGA SWAMY, RAGHUNATH
MANDIPUDI, D. VIJAYA SRI, DR. ATHMAKURI SATISH KUMAR,
DR. M. SRIDHAR, JOHN T
MESIA DHAS |
|
Abstract: |
The growing need for intelligent systems, which can operate in heterogeneous
areas, indicates the limitations of traditional transfer learning methods,
especially with regard to processing high-dimensional data and operating in
real-time conditions. This paper presents a new cross-domain transfer learning
system based on Quantum Neural Networks (QNNs) to be used in multi-task
applications with real-time adaptation, to facilitate scalable and efficient
knowledge transfer across domains, including, but not limited to, healthcare,
IoT, and cybersecurity. The suggested architecture is based on a hybrid
quantum-classical architecture, where parameterized quantum circuits are used to
represent the features, domain-invariant learning and multi-task optimization
strategies are applied, and a reinforcement learning-based module is used to
provide continuous real-time adaptation by using streaming data. Experimental
findings show that the model has 94.8% accuracy, which is significantly higher
than the current classical and transformer-based methods, and that the
adaptation latency of the model is 95 ms, and the transfer efficiency is more
than 10 times higher, which means that generalization and robustness are higher.
The results prove that quantum-enhanced transfer learning is a powerful and
scalable solution in next-generation intelligent systems, with significant
practical impact in real-time systems, such as smart healthcare, cybersecurity,
and IoT systems. |
|
Keywords: |
Quantum Neural Networks, Cross-Domain Transfer Learning, Multi-Task Learning,
Real-Time Adaptation, Hybrid Quantum-Classical Models, Intelligent Systems. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
THE IMPACT OF MANAGERIAL INNOVATION AND DIGITAL TECHNOLOGY ON THE FINANCIAL
PERFORMANCE OF SMES: EVIDENCE FROM AN EMPIRICAL STUDY |
|
Author: |
SARA SAIL, HANANE AAMOUM, IKRAM EL HACHIMI, YASSIN SELOUANI, NAJOUA EL ABBAS EL
GHALEB, IMAD AIT LHASSAN |
|
Abstract: |
Although the resource-based view predicts that managerial innovation should
enhance firm performance, empirical evidence remains contradictory and is drawn
overwhelmingly from large firms in developed economies, leaving the financial
payoff of managerial innovation for small and medium-sized enterprises (SMEs) in
emerging markets such as Morocco poorly understood. The primary objective of
this research is to investigate the impact of managerial innovation and digital
technology on the financial performance of SMEs from 2020 to 2023, aiming to
construct a model that elucidates the influence of managerial innovation
dimensions, with particular attention to the technological dimension, on
financial performance. The research framework developed in this study was tested
on 260 Moroccan SMEs operating across various sectors between 2020 and 2023.
Adopting a positivist epistemological stance, the study employed a
hypothetico-deductive reasoning approach and a quantitative research design. The
empirical validation of the proposed model provided evidence of a partially
significant relationship between managerial innovation dimensions and firm
financial performance. The results revealed that strategic planning, employee
development and motivation, inter-organizational relations and partnerships, and
information and communication technologies have a statistically significant
impact on financial performance. In contrast, the structural dimension of
managerial innovation did not exhibit a statistically significant relationship
with financial performance. This finding suggests that structural changes may
not necessarily translate into immediate financial benefits for SMEs,
particularly in contexts characterized by economic uncertainty and resource
constraints. This study enables SME managers to identify the organizational
changes required to foster managerial innovation and improve performance. It
also contributes to the existing literature by providing empirical evidence on
managerial innovation within the context of Moroccan SMEs. By identifying which
dimensions of managerial innovation reliably translate into financial gains and
which do not, the study offers SME managers and policymakers an evidence-based
basis for prioritizing investment under conditions of resource scarcity. |
|
Keywords: |
Managerial innovation, Financial performance, SMEs, Strategic Dimension,
Structural Dimension, Employee Motivation and Development Dimension,
Inter-organizational Relations and Partnerships Dimension, Information and
Communication Technologies Dimension. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
A HYBRID MULTIPLE ENSEMBLE LOAD BALANCING FRAMEWORK FOR DECISION-MAKING IN
REAL-TIME MULTI-SERVER, MULTI-TASK SCHEDULING IN CLOUD ENVIRONMENTS |
|
Author: |
GUTTA SRIDEVI, TUMMA SRINIVASA RAO, GUNTAPALLI MINNI, HARISH VUNDAVALLI,
RAGAVAMSI DAVULURI5, SAMEENA B |
|
Abstract: |
As the number of tasks and virtual machines increases for task scheduling in
cloud computing environments, predicting an efficient load balancing model for
multi-server, multi-task scheduling becomes challenging. Many conventional
models rely on static multi-task scheduling algorithms for decision-making.
Additionally, these models utilize traditional classification algorithms for
generating multi-task patterns, often working with limited data sizes. This work
presents a hybrid load balancing framework designed to enhance decision pattern
mining in the multi-server, multi-task scheduling process. The framework
integrates Particle Swarm Optimization (PSO) and Ant Colony Optimization (ACO)
into the ensemble load balancing approach within a real-time cloud computing
environment. Furthermore, a hybrid decision tree classifier is proposed to
facilitate decision-making in the multi-task scheduling process. Experimental
results demonstrate that the proposed model outperforms conventional models in
terms of multi-task, multi-server scheduling in real-time cloud computing
environments. |
|
Keywords: |
Multi-task, Multi-server, Load balancing, Cloud computing, Classification
Algorithm. |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
AN INTELLIGENT MULTIMODAL BRAIN TUMOR DISEASE DETECTION FRAMEWORK USING
MULTIHEAD CROSS-COVARIANCE ATTENTION FUSION-BASED ADAPTIVE EXPLAINABLE PYRAMID
DILATED RES-DENSENET |
|
Author: |
B.N. GARI KALAVATHI, UMADEVI RAMAMOORTHY |
|
Abstract: |
Brain tumor is among the most life-threatening neurological disorders, arising
from the uncontrolled proliferation of abnormal cells within the brain and its
surrounding tissue. Precise and early identification of the tumor region is
essential for timely clinical intervention, yet single-modality imaging such as
Magnetic Resonance Imaging (MRI) alone often struggles to capture both the
structural and the metabolic characteristics of a lesion. Positron Emission
Tomography (PET) provides complementary functional information that, when fused
with MRI, improves the reliability of tumor localization. However, existing
multimodal frameworks frequently lose high-frequency structural detail during
fusion, align cross-modal features poorly, and depend on manually tuned
hyperparameters that limit generalization across patients. To address these
shortcomings, this study designs a Multihead Cross-Covariance Attention
Fusion-based Adaptive Explainable Pyramid Dilated Res-DenseNet
(MCCAF-AExPDRDNet) for multimodal brain tumor segmentation from paired MRI and
PET slices. The proposed dual-branch network extracts structural and metabolic
features independently through pyramid dilated dense blocks and residual dense
blocks, and fuses the resulting bottleneck representations through a multihead
cross-covariance attention mechanism before reconstructing the tumor mask
through a skip-connected decoder. To remove the burden of manual hyperparameter
selection, a Best-fit Guided Learning based Apiary Organizational-based
Optimization algorithm (BGL-AOO) is proposed to adaptively tune the hidden
neuron count, learning rate, and steps per epoch of the segmentation network.
The framework is validated on a real, paired MRI-PET brain dataset against
U-Net, ResUNet, TransUNet, GARU-Net, MBTC-Net and ResNet18 baselines, and
against Educational Competition Optimization (ECO), Quokka Swarm Optimization
(QSO), the Supercell Thunderstorm Algorithm (STA) and standard Apiary
Organizational-based Optimization (AOO). Across six activation functions, the
proposed BGL-AOO-MCAF-XPDN model attains a mean accuracy of up to 95.58%, a Dice
coefficient of up to 95.60%, an IoU of up to 91.57%, a recall of up to 95.66%,
and a PSNR of up to 61.70 dB, consistently outperforming every comparator model
and optimizer. These outcomes indicate that the designed framework offers a
scalable and clinically relevant tool for computer-aided brain tumor diagnosis. |
|
Keywords: |
Apiary Organizational-Based Optimization, Brain Tumor Disease Diagnosis,
Multihead Cross-Covariance Attention Fusion, Multimodal MRI-PET Segmentation,
Pyramid Dilated Res-Densenet |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
A COMPARATIVE MACHINE LEARNING FRAMEWORK FOR EARLY DETECTION OF ALZHEIMER’S
DISEASE USING STRUCTURED CLINICAL DATA |
|
Author: |
TENG LE XIN, WAN NOOR HAMIZA WAN ALI, SAHARUDIN ISMAIL, NORMAISHARAH MAMAT, NOOR
JANNAH ZAKARIA, RABATUL ADUNI SULAIMAN, FAIRUZ AMALINA |
|
Abstract: |
Despite more than a decade of machine learning (ML) research on Alzheimer's
disease (AD), early detection remains largely unsolved in practice:
high-accuracy models are concentrated in neuroimaging-based deep learning, which
is costly and inaccessible in resource-constrained settings, while studies using
cheaper, structured clinical data rarely combine rigorous preprocessing,
systematic hyperparameter optimisation, class-imbalance handling, and a
validated deployment pathway within a single, reproducible framework. This paper
addresses that gap. It proposes and experimentally validates a comparative ML
framework for early AD detection using structured clinical, demographic,
lifestyle, and cognitive assessment data. A publicly available Kaggle dataset
comprising 2,149 patient records and 35 attributes was used. Three supervised
classifiers—Support Vector Machine (SVM), Random Forest (RF), and Logistic
Regression (LR) were developed within a reproducible preprocessing pipeline,
optimised using GridSearchCV with stratified cross-validation, and evaluated
under two experimental conditions: with and without Synthetic Minority
Oversampling Technique (SMOTE). The RF classifier achieved the best performance
on the held-out test set, with 94.42% accuracy, 91.95% F1-score and 93.93%
ROC-AUC, and it further exhibited strong robustness to class imbalance, with
only marginal performance variation across the SMOTE and non-SMOTE conditions.
In contrast, LR benefited substantially from SMOTE (F1-score improving from
73.58% to 76.70%), while SVM performance declined slightly. Compared against a
closely related baseline study evaluating SVM, RF, and CNN on a similarly
structured Kaggle dataset, the proposed RF model outperformed the baseline's
traditional ML results while remaining competitive with its CNN-based results,
without requiring neuroimaging data. The validated model was further deployed as
an interactive Streamlit-based prediction application, with an embedded
Developer Mode empirically confirming prediction consistency between offline
experimentation and real-time deployment. These findings demonstrate that a
carefully optimised, ensemble-based ML framework operating on non-invasive
structured data can achieve potentially useful screening performance while
offering limited global interpretability through feature-importance estimates,
being computationally lightweight, and deployable, making it well-suited to
resource-constrained healthcare screening contexts such as Malaysia. |
|
Keywords: |
Alzheimers disease; Random Forest; Clinical data; Class imbalance; SMOTE;
Machine learning deployment |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
Title: |
C&D SENTINEL: AN ENTERPRISE-GRADE MULTIMODAL CONTENT MODERATION SYSTEM USING
HYBRID TWO-STAGE NLP AND COMPUTER VISION |
|
Author: |
K. SHAILAJA, T. JALAJA, M. CHAITANYA, V. DARPAD SA |
|
Abstract: |
The rapid increase of user-generated content through digital means has resulted
in an urgent requirement for scalable, accurate and relevant content moderation
systems. The C&D Sentinel content moderation pipeline solves this issue with an
enterprise-class hybrid architecture consisting of locally deployed machine
learning models and cloud-based large language model (LLM) reasoning. The system
identifies toxic, hateful and unsafe content across a range of formats,
including text, still images, and video, while providing low latency.
Traditional moderation systems rely primarily on manual keyword filtering or
opaque "black box" classifiers that cannot interpret contextual differences
between words; these methods often mislabel harmless items – such as satire,
slang, and artistic forms – as dangerous, resulting in numerous false positive
instances. This paper demonstrates that using a Hybrid Two-Stage Architecture
combining light weight local models and advanced reasoning capabilities provided
by cutting-edge LLMs will resolve these limitations. DistilBERT, an example of
such an efficient model, can perform quick text classification, while BLIP
(Bootstrapping Language-Image Pretraining) will create captions to image content
effectively converting visual data into an appropriate textual form for semantic
analysis.
The Llama 3 70B model is utilized within the system to provide fast
inference with low latency with integration onto the Groq LPU (host machine)
allowing for improved detection of malicious behaviour and amending false
positives compared to previous methodologies for detection of malicious
behaviour in context. One of the implementations of this system is its video
moderation, which is processed in a dual-stream framework where images are
captured using OpenCV for visual features and where audio is converted to text
using Whisper. These two streams are analysed together to allow a greater
understanding of what the video is and therefore whether or not the video
contains any HA. A second component of the system is an RLHF (reinforcement
learning from human feedback) feedback loop, which allows for continual
improvement of the content detection capabilities through a human-in-the-loop
process. RLHF will allow for continual enhancement of detecting new types of
language and harmful content as they emerge. The system was evaluated against
the Facebook Hateful Memes Challenge data set, where it achieved 90% accuracy,
far better than a zero-shot baseline using CLIP that achieved only 14% accuracy.
The system is also deployed to Groq LPU hardware which ensures that the system
can perform real-time analyse with less than 200 milliseconds of inference
latency, making the system well-suited for enterprise content moderation.
Overall, the C&D Sentinel pipeline has demonstrated a significant improvement in
terms of accuracy, lower false positive rates and real-time performance
capabilities to develop a robust and scalable platform for next-generation
content moderation. |
|
Keywords: |
Content Moderation, Retrieval-Augmented Generation, DistilBERT, Multimodal AI,
Reinforcement Learning from Human Feedback, Groq LPU, BLIP, Toxicity Detection |
|
DOI: |
|
|
Source: |
Journal of Theoretical and Applied Information Technology
15th August 2026 -- Vol. 104. No. 15-- 2026 |
|
Full
Text |
|
|
|
