Paper Submission - NPS Australia Submission System

Regime-Aware Temporal Convolutional–Gated Recurrent Forecasting of Hydrogen Production in Wave-Driven PEM Electrolysis

This research introduces RA-TCGN, a compact regime-aware temporal convolutional-gated recurrent model for multi-step hydrogen-production forecasting in wave-powered PEM electrolysis. It establishes a leakage-safe, experiment-level evaluation showing that one-step accuracy is misleading and demonstrates that operating-regime conditioning reduces pooled LOEO RMSE by 28.6%. External testing confirms improved forecasting and substantially lower extreme errors while revealing regime-specific deployment limitations.

A Comparative Assessment of Near- and Far-Field EM Radiation from Daily-Use Wireless Devices

This research provides a comparative assessment of near-field and far-field electromagnetic (EM) radiation emitted by commonly used wireless devices. The study analyzes radiation characteristics at different distances and operating conditions, helping to improve understanding of human exposure to EM radiation and the variation of radiation intensity in practical daily-use scenarios

Beyond Internal Validation: Cross-Dataset Generalization of Machine Learning and Deep Learning Models for Oral Cancer Image Classification

The contribution of this study is the systematic evaluation of the cross-dataset generalization of oral cancer image classification models, moving beyond conventional within-dataset validation. By evaluating ML and DL models across two independently collected datasets using bidirectional external validation, the study quantifies the generalization gap and demonstrates that high internal performance may not translate to unseen datasets. The study further integrates the selected model into a web-based screening prototype, connecting model evaluation with practical accessibility.

A Retrieval-Augmented Generation Framework for Cancer Health Question Answering Using Semantic Filtering and GPT-4o-mini

This work presents a lightweight, domain-specific RAG framework for cancer QA that uses semantic embedding-based topic filtering to construct a focused 752-pair corpus from MedQuAD. It demonstrates substantial performance gains over BM25 (64.1% MRR improvement) and non-RAG GPT-4o-mini (10.4% BERTScore improvement), with expert evaluation confirming higher safety and accuracy. The ablation studies systematically identify optimal retrieval depth (k=4) and embedding model, while the framework is designed for scalable, non-diagnostic consumer health information access.

Q-Learning-Based Degradation-Aware BESS Management for Agricultural Microgrids

Battery Energy Storage System (BESS) integration
improves reliability and sustainability in agricultural microgrids,
but inefficient control accelerates degradation and can itself
compromise reliability. We propose a reliability-constrained,
degradation-aware BESS management strategy using Q-learning.
During development we found that an unconstrained action space
lets a Q-learning agent inflate its apparent cycling reduction
by withholding discharge from demand– a reward-hacking
failure mode leaving 4.7–14.9 percentage points more demand
unserved than the baseline. We resolve this by guaranteeing
discharge whenever the battery can physically supply it, and
training the agent only to control charging intensity during
solar surplus. Evaluated on one year of real 15-minute field
data from an agricultural microgrid in Jashore, Bangladesh,
across three agricultural seasons with a chronological, leakage
free 80/20 train/test split and validated across 3 independent
training seeds, the controller matches a same-dataset greedy
baseline’s service reliability to within 0.03% of total demand
while reducing cumulative battery cycling by 44.4–54.1% and
cumulative degradation damage by 66.2–69.3% (mean ± std) on
held-out data. A companion ablation confirms the unconstrained
variant inflates its apparent reduction to 77.8–95.2% only by
sacrificing reliability. A complementary multi-objective (NSGA
II) analysis and a benchmark of four supervised SoC/RUL fore
casting models (best R2=0.915, LSTM) are also reported. Results
indicate reliability-first reward design is essential for trustworthy
degradation-aware reinforcement learning in agricultural BESS
applications.

An Explainable and Probability-Calibrated AdaBoost Framework for Type-2 Diabetes Risk Prediction Using SMOTETomek

Type-2 Diabetes Mellitus (T2DM) is a growing
public health challenge, particularly in low- and middle-income
countries where early risk identification remains critical. This
study proposes an explainable and probability-calibrated Ada
Boost framework for T2DM prediction using a Bangladesh-based
clinical dataset of 1,065 women. Six machine learning classifiers
were evaluated with four imbalance-handling strategies under
stratified 10-fold cross-validation. AdaBoost with SMOTETomek
achieved the best performance and was further optimized using
isotonic probability calibration. The calibrated model achieved
an accuracy of 0.939, recall of 0.999, F1-score of 0.961, and
AUC of 0.905 on an independent test set. Calibration significantly
improved probability reliability, reducing the Brier score from
0.1618 to 0.0620 and Expected Calibration Error from 0.3194 to
0.0542. SHAP analysis identified fasting glucose, systolic blood
pressure, age, and number of pregnancies as the most influential
predictors, while Decision Curve Analysis demonstrated positive
clinical net benefit across relevant threshold ranges. The proposed
framework integrates imbalance correction, calibrated predic
tion, explainability, and clinical utility assessment, providing a
transparent approach for diabetes risk stratification. External
validation on larger multi-center cohorts is required before
clinical deployment.

Explainable Lightweight Deep Learning for Plant Leaf Disease Detection Using MobileNet and Grad-CAM

This research proposes an explainable lightweight deep learning framework for plant leaf disease detection using MobileNetV2 and Grad-CAM. The proposed approach achieves high classification performance while reducing computational complexity, making it suitable for mobile and resource-limited agricultural applications. MobileNetV2 is compared with EfficientNetB0 and ResNet50 based on accuracy, precision, recall, F1-score, model parameters, and inference time. Grad-CAM is integrated to provide visual explanations by highlighting disease-relevant regions of leaf images, improving model transparency and reliability.

A Trust-Weighted Blockchain Framework Coupling AI-Driven Risk Scoring with Bounded Sybil-Resistant Stake Economics

1.A source-degeneracy diagnostic for fake-news benchmarks. We show that reported detection accuracies above 99% are largely an artifact of benchmark construction: on ISOT the source field alone predicts the label with 100% accuracy, a bag-of-words model matches deep encoders, and a source-disjoint split is therefore impossible in principle.
2.A rigorous four-protocol, leakage-controlled evaluation. On 61,199 items we progress from an in-domain split (97.6%) through lexical scrubbing and cross-corpus transfer (down to 61.7%) to leave-one-publisher-out cross-validation, where the mean per-publisher AUC of the label-mixed hyperpartisan outlets is near chance (≈0.52), reported with Hanley–McNeil confidence intervals.
3.Evidence that calibration does not transfer under distribution shift. A source-fitted temperature removes far less calibration error than a target-fitted one, and the divergent GossipCop temperature collapses predictions toward 0.5, showing that single-scalar temperature scaling cannot repair severe shift but correctly detects that the score is uninformative.
4.A remediation architecture whose safety does not depend on the classifier. The calibrated score allocates scarce human adjudication rather than delivering verdicts; an oracle-review upper bound and an imperfect/adversarial five-voter jury simulation both lift a below-baseline classifier (65.3%) to a safe system (82.8–97.3%).
5.A bounded, Sybil-resistant bonding curve. A logarithmically bounded Integrity Bond compresses the novice-to-elite entry-cost ratio from 10,001× to 14.0×, invariant to the risk multiplier, while preserving the security invariant that every fresh identity posts the maximum bond.
6.A deployed, security-tested ERC-1155 contract with a burst-detection correction. We identify and fix an ordering flaw in which writing the probation flag inside the reverting transaction silently discards it, and report on-chain gas, batching savings (35.5% per item), and slashing-funded verification, with deployment claims scoped to a permissioned/Layer-2 setting.

History-based Improvement of the Co-change Suggestion Capability of FLeCCS

FLeCCS is a technique that suggests code fragments likely to need a matching change (co-change candidates), based only on how similar two fragments look on the page. It never checks whether fragments have actually changed together in the past, so it misses real candidates that happen to be worded differently. This paper introduces RT-FLeCCS, which mines a system’s own revision history as method-level association rules using Tarmaq, keeps only the rules whose support and confidence clear chosen thresholds, and merges the result into FLeCCS’s suggestions. Across four open-source systems (774 to 1700 revisions each, in C and Java), RT-FLeCCS raises recall over plain FLeCCS by 53% to 170% and F-score by 33% to 90%, while giving up almost nothing in precision. A support threshold of 1 paired with a confidence threshold of 0.5 gives a reliable general-purpose setting across all four systems.

An Explainable Ensemble Machine-Learning Framework for Predicting ICU Mortality in Critically Ill Patients

1. A 71-feature clinical feature-engineering pipeline built from 22 raw ICU variables.
2. An ensemble architecture combining XGBoost,LightGBM, CatBoost, HistGradientBoosting,ExtraTrees, and Random Forest.
3. An optimized weighted ensemble with logisticregressionstacking to fuse predictions from complementary base learners.
4. A preprocessing and class-balancing pipeline usingmedian imputation, SMOTE,RobustScaler, and F1-driven threshold tuning.

An Intelligent Multi-Scale Lesion Attention Transformer Framework with Retrieval-Augmented Farmer Guidance for Onion Leaf Disease Diagnosis

Onion diseases have been a significant hurdle for crop production, resulting in huge crop losses. It is necessary to detect and identify diseases quickly and correctly for effective disease management. This paper proposes a deep learning framework for identifying onion leaf diseases automatically. Our data collection methodology involves capturing images of onions grown in the field under various conditions in Faridpur, Bangladesh. We used our collected images for classifying diseases through the proposed method.Our proposed framework utilizes EfficientNetB4 along with a lesion attention mechanism and multi-scale feature fusion by transformers to increase feature extraction and classification capability of the network. To make our model more transparent, we have used the XAI approach to explain our prediction using a Grad-CAM visualization technique to help in understanding which region is important for predicting the infection. Furthermore, our model suggests a treatment strategy based on RAG guidance module. As per experimental results, our proposed model performs with 98.65% classification accuracy.

Robust Federated IIoT Intrusion Detection Under Byzantine Label-Poisoning Attacks: Class-Wise and Client-Wise Analysis on DataSense

The main contribution is empirical, not methodological. The paper shows that global metrics can hide a targeted attack completely. Under Bruteforce→Benign label poisoning, accuracy stays near 94% for all three aggregators. But Bruteforce recall drops to 4.11% under FedAvg. Trimmed Mean recovers it to 47.24%, and FedCARR to 84.65%. The paper backs this with a reproducible 10-client, 8-class non-IID DataSense setup, and with class-wise, client-wise, ROC, and paired statistical diagnostics on one locked experimental grid. The practical takeaway is that robust FL for IIoT intrusion detection must be evaluated per class and per client, since aggregate scores alone will not reveal the failure.

PP-GBD: Privacy-Preserving Graph Neural Botnet Detection in Encrypted Traffic

This work proposes PP-GBD, a unified privacy-preserving framework for botnet detection in encrypted traffic using graph neural networks and payload-independent metadata. Its key contribution is a formally calibrated fixed-slot graph-level differential privacy mechanism with explicit add/remove adjacency and sensitivity, integrated with federated learning and an HE-oriented polynomial GCN design. The study also provides a reproducible controlled benchmark and quantifies the privacy–utility trade-offs across non-private, differentially private, polynomial, and federated models.

LAUC-RiskNet: Leakage-Aware and Interpretable Learning for Rule-Driven Mental-Health Labels

The significant contribution of this research is an audit-then-ablate framework that identifies and quantifies deterministic target leakage in behavioral mental-health datasets. LAUC-RiskNet combines interpretable threshold learning, contextual encoding, adaptive fusion, probability calibration, and uncertainty estimation. Crucially, ablation shows that the apparent predictive performance arises mainly from the label-generation rule rather than independent clinical signals, enabling more transparent and trustworthy evaluation.

Evaluating Arterial Network Resilience Post-Mass Transit Integration: A Hybrid Simulation and Empirical Corridor Study

Widening roads offers only a partial solution to the traffic gridlock that plagues rapidly developing megacities. We evaluate a 3.18-km (1.98-mi) section of Kazi Nazrul Islam Avenue, Dhaka, before and after the integration of MRT Line 6, AI-enforced access management, and semi-automatic signal control, and estimate the associated change in corridor delay and CO₂ emissions. We investigate the impact of integrating high-capacity rail with technology-driven surface management. We use a hybrid approach, comparing a validated 2018 microscopic simulation that documented severe delays under manual police control with new 2026 field trajectory data. Recent developments include the introduction of Mass Rapid Transit Line 6, deployment of Artificial Intelligence (AI) cameras for lane monitoring, and implementation of a semi-automatic Intelligent Transportation System (ITS) developed by the Bangladesh University of Engineering and Technology (BUET). Despite permanent lane narrowing from MRT viaduct pillars, observed peak-hour car speeds rose from a 5.2 km/h simulated baseline to 12.3 ± 0.5 km/h (n = 15; +136%). Because classified volume counts were unavailable, the relative contributions of operations management and MRT-induced demand shift cannot be separated and are discussed in §IV-D. The increase in peak-hour car speeds was 136 percent, moving from a baseline of 5.2 kilometers per hour to an empirically measured 12.31 kilometers per hour. Averaging 17.43 kilometers per hour, two-wheelers outperformed other modes of transportation. We also examined the new variances observed at key intersections in signal cycles, ranging from 180-second cycles to as long as 260 seconds during heavy traffic. Data indicates that manual ITS overrides are an essential component of boundary metering, which restricts traffic to some extent and prevents the internal network from completely locking up. In the end, our findings indicate that traffic volume is not solely determined by the physical lane capacity. The combination of overhead transit and access controls backed by technology can be instrumental in eliminating latent delay and maintaining the speed of urban highways.

Novel Attacks Arrive Diffuse: Rethinking Few-Shot Open-Set Enrollment for ROS Intrusion Detection

The paper’s central contribution is showing that single-prototype enrollment in few-shot open-set recognition has a geometric ceiling that more labels cannot lift — novel-class F1 actually declines with support size, and an oracle prototype built from all 4,326 novel flows matches the 100-shot result exactly, ruling out sampling error. The diagnosis generalizes beyond ROS: because the episodic loss never shapes the unseen class, it arrives 2–4× more diffuse than trained classes and its mean falls inside a rival class basin. The fix is cheap and follows directly — enrolling three k-means sub-prototypes instead of one raises mean novel-class F1 from 0.694 to 0.862 with no retraining.

Interface-Dependent Current and Voltage Losses in a Lead-Free ZnSe/FASnI3/Cu2O Perovskite Solar Cell: A SCAPS-1D Study

This work examines a Glass/ITO/ZnSe/FASnI3/Cu2O/Au lead-free perovskite solar cell in SCAPS-1D, degrading the two absorber interfaces independently rather than together. The Cu2O/FASnI3 interface is shown to be the costlier one to degrade, taking 1.981 mA/cm2 of photocurrent and 5.02 percentage points of efficiency against 0.483 mA/cm2 and 3.41 percentage points at FASnI3/ZnSe, a photocurrent ratio of 4.10 at the highest defect density modeled. This ordering is the reverse of the ETL-side dominance reported for other lead-free absorbers, and the 0.80 eV electron barrier at Cu2O offers a plausible explanation, since electrons generated near that face must traverse the remaining absorber thickness before extraction at ZnSe. The device is also run under separate front and rear one-sun illumination, which leaves the photovoltage and fill factor nearly unchanged and concentrates the loss in photocurrent; because the modeled rear electrode is opaque Au, that result measures the electrical symmetry of the layer sequence rather than bifacial performance.

Calibrated Federated Intrusion Detection for Drone Swarms under Non-IID Feature Skew

For the protection of critical infrastructure with
drone swarms, an intrusion detection system must be able
to detect intrusion without concentrating sensitive telemetry.
Evaluation leakage, which has been commonly used in prior
federated learning (FL) based intrusion detection approaches, has
been shown to lead to high discrimination scores, but these prior
works have overlooked the probability calibration aspect and
have resulted in overconfident alerts under non-IID feature skew.
We propose an uncertainty-aware intrusion detection system
that is trained using federated learning and tested under a
leakage-safe protocol, comprised of chronological holdout splits,
cross-split duplicate hashing, and feature hardening. With the
ISOT Drone Intrusion dataset, we create a setting of ten clients
with a feature-skew distribution, and compare four aggregation
strategies (FedAvg, FedProx, FedNova, FedDyn) using operating
points found through Youden’s J statistic. Accuracy and Macro
F1 are used to measure the predictive quality, while Expected
Calibration Error (ECE) after post-hoc temperature scaling is
used to measure the reliability, and SHAP is employed as a
behavioral inspection tool. The upper bound is given by a
centralized Random Forest with Macro F1 = 0.9896. When the
skew is severe, FedAvg achieves 98.60% accuracy and Macro
F1 = 0.9844 and temperature scaling makes the ECE 0.0027
thus allowing selective alerting to reduce false alarms. The
federated model is also better at withstanding skew, as shown in
explainability analysis, in that it is more agile than centralized
models with respect to spatial identifiers. These results apply to
the particular dataset and construction of the non-IID system
and may not represent all shifts in deployments; however, they
suggest that an FL system can provide near-centralized UAV
intrusion detection with reliable human-in-the-loop deferral.

HemaStack: Harnessing Heterogeneous Stacking with Neural Meta-Learning and Dual Explainability for Cost-Effective Dengue Diagnosis

Dengue fever in Bangladesh is a critical public
health issue, particularly in the Dhaka area, with over 30,000
cases reported in 2025, highlighting the need for quick and
reliable diagnostic tools. While routine Complete Blood Count
(CBC) parameters are easily available, traditional diagnosis relies
on NS1 antigen testing, which is expensive and susceptible to false
negatives in later stages of infection. To overcome this limitation,
we propose HemaStack, a stacking ensemble framework for
binary dengue classification using routine hematological blood
parameters. A dataset of 1,037 patients from various tertiary
hospitals in Dhaka, Bangladesh, from June to August 2025, was
preprocessed using label encoding, IQR-based outlier removal,
StandardScaler normalization, and SMOTE balancing. HemaS
tack integrates five base learners—Random Forest, Extra Trees,
XGBoost, LightGBM, and SVM—with an MLP meta-learner
using passthrough feature augmentation, benchmarked against
eight baseline classifiers under 10-fold cross-validation and an
independent test set. HemaStack attained an accuracy of 98.59
± 0.98%, an F1-score of 98.59 ± 0.97%, and a ROC-AUC of
0.9966 ± 0.0037; on the held-out test set it achieved an accuracy
of 97.57%, a recall of 99.38%, an F1-score of 98.45%, a ROC
AUC of0.9852, and an MCCof0.9292, outperforming all baseline
models. SHAP and LIME analyses identified Lymphocytes (%),
Platelet Count, and Hematocrit as the most important features.
The results indicate that HemaStack enables detection of dengue
in a timely, cost-effective, and interpretable manner without
specialized diagnostic tests, which could be useful in the resource
limited healthcare system in Bangladesh and other endemic areas.

A Noble Technique of Patient Data Handling for Bangladesh.

The significant research contribution is the development of a secure, interoperable, and scalable patient information management system that integrates heterogeneous healthcare databases through a middleware layer. The system reduces patient-data fragmentation, enables efficient retrieval of comprehensive records, protects sensitive information, and supports improved healthcare and telemedicine services, particularly in remote areas.

RSANet: An Advance in Oral Cancer Prediction using Residual Network with Soft-Attention Mechanism

Oral cancer is still a major worldwide health problem, especially in South and Southeast Asia, with late detection and variability in diagnosis leading to poor survival. Histopathological investigation is the gold standard of diagnosis but is time-consuming, subjective and pathologist dependant. We propose RSANet in this work, a deep learning framework integrating a ResNet50 backbone with a soft-attention mechanism for discriminating normal tissue from oral squamous cell cancer (OSCC) in histological pictures. The residual backbone encodes hierarchical morphological patterns, while the attention mechanism distributes higher weights to selected spatial regions and provides visual maps to check model focus. RSANet obtained an accuracy of 97.56%, a sensitivity of 98.33% and an F1-score of 97.75% on the held-out test set. The ResNet50 baseline acquired an accuracy of 91.40%, for comparison, equal to an absolute gain of 6.16 percentage points. Our results show that using soft attention improves classification performance in the experimental setup we analyzed. RSANet is a promising computer-assisted OSCC classification system that might support pathologists in making diagnostic decisions, however independent validation in multicenter setting is needed prior to its possible clinical use.

BCViT++: Locality-Aware Large-Kernel Vision Transformer for Breast Cancer Histopathology Classification

Automated categorization of breast histopathology pictures can benefit computational pathology, however tissue appearance changes with optical magnification, and annotated datasets remain tiny for Vision Transformers (ViTs). In this study, we propose BCViT++, a magnification-specific framework for binary categorization of benign and malignant breast tumor pictures. This pipeline incorporates channel-wise principle component analysis (PCA) reconstruction with a variance threshold of at least 99%, training-set class balancing, geometric augmentation, Locality Self-Attention (LSA), and Large Kernel Attention (LKA). LSA masks token self-attention with a diagonal matrix and learns the attention temperature, while LKA refines value representations using depth-wise convolutions with a larger effective receptive field. All BreakHis experiments were done separately at 40×, 100×, 200× and 400×. We obtained an accuracy of 86.57%, 86.87%, 90.37% and 90.44% using a PCA-ViT baseline, respectively. The whole BCViT++ framework improved these accuracies to 93.53%, 91.17%, 96.79%, and 96.45% with absolute improvements of 6.96, 4.30, 6.42 and 6.01 percentage points, respectively. The best performance was achieved at 200× with accuracy of 98.55%, sensitivity of 96.79%, specificity of 96.80%, F1-score of 97.66% and MCC of 0.926. These results show consistent overall improvements over the PCA-ViT baseline for the studied image-level technique.

MVPR-PF: A Multi-View Peptide Representation and Probability Fusion Framework for Anticancer Peptide Prediction

Anticancer peptides (ACPs) are promising therapeutic candidates because of their selective activity against cancer cells and relatively low toxicity toward normal cells. However, experimental ACP identification is time-consuming and costly, motivating efficient computational screening methods. This study proposes Multi-View Peptide Representation and Probability Fusion (MVPR-PF), a hybrid machine and deep learning framework that integrates Binary Profile Features (BPF), Composition of $k$-Spaced Side-Chain Pairs (CKSSCP), a 20-dimensional pI-derived physicochemical property vector, and grouped 3-mer descriptors. XGBoost learns nonlinear relationships from the structured multi-view features, while a Bidirectional Long Short-Term Memory (BiLSTM) network captures bidirectional residue dependencies from the position-aware BPF representation. Their prediction probabilities are combined through weighted probability-level soft voting. Using stratified 10-fold cross-validation, MVPR-PF achieved 86.89% accuracy and 93.66% AUC on ACP740, and 85.83% accuracy and 91.89% AUC on ACP240. These headline numbers are consistent across all reported tables, the Discussion, and the Conclusion. These results demonstrate that combining complementary handcrafted and sequential representations provides effective computational ACP prediction.

Design of an Intelligent Solar-IoT Irrigation System for Sustainable Agriculture

The main contribution of this research is the integration of real-time automatic irrigation based on soil-moisture, ESP32 control, LoRa long-range communication, IoT monitoring into a single smart irrigation platform and dual axis solar power. The study also explores the water saving and operational cost benefits, suggesting its applicability for sustainable and efficient agricultural irrigation.

An Explainable Machine Learning Framework for Cyberbullying Detection in Social Media Comments: Hybrid Lexical–Sublexical Modeling with Global and Local Explanations

This research provides an explainable and high-performing cyberbullying detection framework by combining (1) hybrid lexical–sublexical text representation (TF-IDF with word-level and character-level) (2) optimized LinearSVC classification (3) explanation for decisions based on SHAP and LIME. The limitations of existing black-box detection models are overcome as the proposed approach reasoning about the predictions at the feature level in a transparent manner. The framework is evaluated using a custom dataset of 22,630 social media comments, and achieves a score of 97.77% accuracy, 98.04% F1-score, and 0.9976 ROC-AUC, with explanation faithfulness reported using token-deletion analysis. This research is making strides towards more trustworthy AI-assisted content moderation while also providing high predictive performance alongside interpretable decision explanations.

AI-Powered Smart Clothing Management System

This research introduces an AI-powered smart wardrobe system that leverages deep learning and reinforcement learning to optimize clothing classification and outfit recommendations. It reduces decision time, enhances wardrobe utilization, and promotes circular fashion practices, thereby contributing to sustainability and inclusivity in the textile industry.

Hierarchical MLP for EEG- Based Decoding of Cognitive Load in Multi-Level n-Back Tasks

This paper contributes by performing following steps: (i) extracts a rich multi-domain EEG feature set spanning time, frequency, and time-frequency domains, (ii) applies rigorous non-parametric statistical selection to retain 241 informative channel–feature pairs, (iii) achieves high-accuracy binary and multi-class classification of n-back load levels, and (iv) complements classification with exploratory PLV and Time-Frequency Representation (TFR) analyses that illuminate the underlying neural mechanisms.

Leveraging AI-Driven Cloud Computing for Advanced Health Prediction Models

Modern healthcare systems are undergoing a paradigm shift from reactive clinical interventions to continuous, proactive risk prediction. However, deploying advanced predictive models in real-world clinical settings is severely constrained by heterogeneous data silos, computational latency, and the opaque black-box nature of complex algorithms. To overcome these challenges, this paper presents PERFORM (Patient Edge Resilient-Fast Optimized-Reliable Monitoring), a novel, multi-layered architectural framework that synergizes Artificial Intelligence (AI) with scalable hybrid Edge-Cloud computing infrastructure. The proposed framework operates across five interconnected layers: harmonizing disparate data streams (EHRs, imaging, IoMT telemetry, and multi-omics); dynamically orchestrating computational workloads between edge nodes and cloud environments to minimize processing latency; executing adaptive model selection; embedding Explainable AI (XAI) for transparent decision support; and maintaining continuous governed evolution via drift tracking and clinical feedback. By synthesizing data integration, low-latency execution, and audit-ready interpretability into an end-to-end system, this study provides a unified blueprint for deployable, safe, and clinically trustworthy risk prediction.

Interpretable High-Dimensional Advertisement Classification with Regularized Logistic Regression: A Numerically Stable Framework

High dimensional advertisement classification is hard due to the distribution of predictive information across disparate geometric and sparse lexical features, imbalanced class distributions, and the need for systems that explain their decisions in simple and reproducible ways. This work proposes an interpretable and numerically stable binary classification framework based on UCI Internet Advertisements dataset that has 3,279 observations and 1,558 predictive features. The model is a logistic regression trained with a self-written L2 regularization that fits within a leakage safe pipeline that contains stratified partitioning, training-only median imputation, and training-derived feature standardization. To avoid computational instability, numerical safeguards are used, such as sigmoid clipping and stabilized binary cross-entropy computation. The model achieves 97.56% test accuracy, 95.24% precision, 86.96% recall, 99.29% specificity, and a 90.91% F1-score. Training and validation losses are close and go to each other steadily, learning-curve analysis exhibits better generalization as sample size grows. A stable high-performance region is also identified through learning-rate sensitivity analysis. The framework is intrinsically interpretable, since the predictions are directly related to the coefficients that can be inspected. Overall, the results illustrate that a well-regularized linear model can provide good predictive performance, while maintaining transparency, numerical stability, reproducibility, and computational simplicity.

A Multi-Dataset Deep Learning and Sequential Pattern Mining Framework for Interpretable Disease Progression Tracking in Brain and Liver MRI

This study makes a significant contribution by proposing a novel hybrid framework that integrates deep learning (CNN/ResNet-18) with Sequential Pattern Mining (PrefixSpan) to bridge the gap between predictive accuracy and clinical interpretability in medical image analysis. The framework introduces a robust discretization technique that transforms continuous, high-dimensional deep imaging features into clinically meaningful symbolic states (Low, Medium, High), enabling the application of pattern mining without losing critical progression information. It further demonstrates a dual application of pattern mining by capturing temporal disease evolution in longitudinal data (LUMIERE) while also uncovering recurring spatial phenotypes in cross-sectional data (BraTS2021 and CirrMRI600plus). The framework’s generalizability is validated through comprehensive testing across three public MRI datasets covering both brain and liver diseases, achieving strong accuracies of 91.2%, 87.5%, and 85.1%, respectively. Ultimately, this research delivers clinically interpretable insights that allow healthcare professionals to understand and trust model outputs without specialized machine learning expertise, offering a practical, scalable, and explainable AI-based clinical decision support solution, particularly suited for resource-limited healthcare environments.

Secure Medical Image Communication using ECC-based Hybrid ROI-Aware Chaotic Cryptograpy

The principal contributions are as follows:
1. A multimodal database is constructed with the combination of CT, X-ray, ultrasound, and MRI images to examine encryption consistency across different images.
2. The prepared dataset is utilized to evaluate the statistical measures in encryption-decryption by the five widely used methods.
3. The communication perspective is also assessed in terms of their resistance to brute-force, differential, and known-plaintext attacks.
4. A weighted decision-oriented ranking approach is proposed to identify the most suitable method according to security, efficiency, communication suitability, and image-recovery performance.
5. A hybrid method is also proposed availing the best performances of different frameworks to check whether the method can show optimal performances in aforementioned criterion. The proposed method is a ROI-Aware ECC-Assisted Chaotic AES hybrid encryption framework. The method uses ROI-based selective encryption with ECC-based lightweight key establishment along with the logistic chaotic map (dynamic chaotic S-box) expansion technique.

Enhanced CBAM-LMSPNet: A Lightweight CNN with Learnable Multi-Scale Pooling for Apple Leaf Disease Classification

We propose a lightweight CBAM-enhanced CNN integrated with a novel Learnable Multi-Scale Pooling (LMSP) strategy for apple leaf disease classification. Unlike conventional transfer-learning approaches, the proposed model is trained from scratch and dynamically fuses fine-, medium-, and coarse-scale spatial features to capture complex lesion patterns. Evaluated using stratified 5-fold cross-validation, the proposed network outperforms six benchmark pretrained models, achieving 99.87% test accuracy and a 99.88% macro-F1 score. Furthermore, Grad-CAM++ visualizations demonstrate that the model consistently focuses on disease-relevant lesion regions rather than background areas, enhancing the interpretability of its predictions

Transformer Guardian 360: An IoT-Based Multi Parameter Condition Monitoring System for Distribution Transformers

This research presents an IoT-based multi-parameter transformer condition monitoring system that integrates voltage, current, oil temperature, oil level, humidity, and vibration sensing using an ESP32 microcontroller. A Transformer Health Index (HI) is developed by combining multiple operating parameters to provide an overall condition assessment and early fault warning. The system enables real-time monitoring through Blynk IoT, helping improve transformer reliability and support predictive maintenance.

A Systematic Review and Cross-Domain Taxonomy of Emerging Cybersecurity Threats and Zero-Trust-Based Defence Architectures

1. We develop a structured taxonomy of emerging cybersecurity attack vectors covering network and infrastructure attacks, IoT and next-generation communication threats, application and physical-layer attacks, and AI-enabled attack vectors.
2. We analyze existing studies using common criteria, including attack domain, methodology, evaluation environment, security mechanism, major findings, and reported limitations, enabling a consistent comparison across heterogeneous research areas.
3. We investigate how Zero Trust principles, including explicit authentication, least-privilege authorization, continuous monitoring, and dynamic policy enforcement, can complement modern attack detection and mitigation mechanisms.
4. We identify major research gaps and propose a conceptual adaptive cross-domain Zero Trust defence framework that connects attack detection, risk assessment, policy enforcement, decentralized trust, privacy-aware learning, and continuous security feedback.

IOT based Over and Under Voltage Protection for Automated Silo Storage System Using LOGO PLC

Automated silo storage systems need predictible protection in the event of abnormal supply-voltage conditions, as well as continuous supervision of environmental parameters. Typically, conventional implementations require another protection and monitoring device, which adds complexity and expense to systems. This paper introduces a common PLC based architecture which includes over and under voltage detection, multi-stage protection of motors, temperature and humidity measurement and alarm handling in a single Siemens LOGO! controller. Transportation of measurements to the protection layer is done locally so that the protective actions are not dependent on the communication layer. The Modbus-based interface, Node-RED and Blynk platform is used for remote supervision, visualisation and event monitoring. The control logic developed worked well in identifying the high and low voltage tested conditions and then starting the programmed protection sequence and the appropriate environmental alarms were generated. The communication pathway also passed on representative operating information to the remote dashboard. The results show that a low-cost and integrated protection and monitoring system for automated silo storage systems is viable; however, hardware validation is needed for full field-level evaluation.

A Lightweight Hybrid CNN–DBN Framework for Lung Cancer Classification Using CT Images

The current research work proposes an automated lightweight hybrid CNN-DBN framework for the classification of lung cancer from the CT images. The framework consists of CLAHE for enhancing images, CNN for spatial feature extraction, DBN for hierarchical representational learning and Logistic Regression for classification that ensures efficient and accurate diagnosis. A low-confidence sample refinement strategy is presented which filters out the samples that the model does not predict accurately, and therefore improves the accuracy of the prediction without re-training the model. The proposed framework yield 98.80% classification accuracy, which shows the efficacy of using both supervised and unsupervised feature learning in medical image analysis. The use of lightweight AI-based CAD systems for accurate and efficient lung cancer screening were emphasized.

LsEyeNet: A Hybrid Deep Feature Fusion Framework for Multi-Class Ophthalmic Disease Recognition from Retinal Fundus Images

LsEyeNet, a hybrid deep feature fusion model is proposed for multi-class ophthalmic disease recognition from fundus images. The study evaluates supervised and self-supervised deep learning models in a comprehensive fashion in a common framework. The proposed framework using combined complementary feature representations of ConvNeXtBase and EfficientNetB3 attains improved classification results, such as classification accuracy of 97.80% and F1-score of 0.96. It also investigates the potential of self-supervised learning using SimCLR and shows the advantage of feature fusion in building a reliable, efficient and accurate AI supported ophthalmic screening framework.

A Decentralized and Privacy-Adaptive Ride-Sharing Framework Using Blockchain and Differential Privacy

This work proposes a decentralized and privacy-preserving ride-sharing system that allows for secure multi-provider mobility data aggregation with the support of a trust management system based on blockchain, a differential privacy framework, an OAuth 2.0 authorization system and a hybrid storage system. The proposed approach addresses existing systems’ shortcomings by offering privacy-preserving ride discovery, easy consent management, and secure sharing of location information while maintaining the service’s utility. Prototype evaluation shows the feasibility of the framework by demonstrating location perturbation of ~80 meters at ε = 0.025 and acceptable performance for moderate workloads.

Deep Learning-Based Ordinal Severity Grading of Mustard Flea Beetle Infestation from Plant Images

Damage to mustard caused by feeding by flea beetles is an important constraint to mustard production, especially during early stand establishment when feeding may cause a reduction in plant vigor and development. However, in contrast to the detection of pests and the determination of overall damage, the gradation of severity of flea beetle infestation has not been thoroughly studied. This research proposes an automated grading system for the severity of a mustard flea beetle infestation using a deep learning approach and four ordered classes: healthy, mild, moderate and severe. The framework includes source aware data partitioning, redundancy control, imbalance-aware learning, ordinal severity modelling, and comparison of five CNN architectures. The accuracy, balanced accuracy, macro-F1, Quadratic Weighted Kappa (QWK) and ordinal mean absolute error (MAE) are used for performance evaluation. The highest individual accuracy is obtained by ConvNeXt-Tiny with 94.93%, the best balanced accuracy is 0.7341, the best macro-F1 is 0.6866, the best QWK is 0.9295 and the best ordinal MAE obtained is 0.0510. The performance is further enhanced to 95.26% accuracy, 0.6930 macro-F1 and 0.9351 QWK by using prediction-level ensemble averaging. The mean of the three-fold cross validation is 0.6296 ± 0.0028 (source grouped). The activation energy using Grad-CAM analysis is focused in the estimated leaf areas only 72.1%, whereas the mean leaf area index is 38.8%. The results indicate that deep learning has the potential to be used for automated multi-level flea beetle infestation grading and can be used as a starting point for further validation in different field conditions.

An Explainable Transfer Learning Framework Based on EfficientNetB3 for Multi-Class Brain Tumor Classification from MRI Images

Accurate and interpretable brain tumor classification from magnetic
resonance imaging (MRI) remains challenging due to substantial
variations in tumor appearance and the limited interpretability of
deep learning models. This paper presents an explainable transfer
learning framework based on EfficientNetB3 for four-class brain tumor
classification, covering Glioma, Meningioma, Pituitary Tumor, and No
Tumor. The framework combines ImageNet-based transfer learning,
training-time data augmentation, selective backbone fine-tuning with
frozen Batch Normalization layers, and a customized classification
head with regularization. The model is evaluated on 7,153 MRI images
using a stratified 70%/15%/15% train-validation-test split. On the
held-out test set, the proposed framework achieves 99.16% accuracy,
with weighted precision, recall, and F1-score of 99.17%, 99.16%, and
99.16%, respectively. Only nine misclassifications are observed among
1,073 test images. Furthermore, Grad-CAM visualizations provide
qualitative insights into the image regions contributing to the model’s
predictions. The results demonstrate that the proposed framework
combines high multi-class classification performance with
model-level interpretability, providing a promising foundation for
explainable AI-assisted brain tumor analysis.

Towards Edge Artificial Intelligence for Predicting Antimicrobial Resistance in Mycobacterium tuberculosis

The main research contribution of this paper is the creation of a lightweight machine learning method. This can run on edge devices to predict resistance in Mycobacterium tuberculosis using whole‑genome sequences. Existing methods usually need features that use a lot of computing power or need prior knowledge of resistance variants. The new model instead uses a 256‑dimensional 4‑mer frequency vector fed into a Random Forest classifier. This change cuts the feature space by orders of magnitude while retaining strong predictive ability. The result is a tool that works well in hospitals in low‑ and middle‑income countries where computers are slow and internet is weak. An average ROC‑AUC of 0.820 is achieved across 20 antibiotics.

ARIS: An IoT–Digital Twin Architecture for AI-Assisted Adaptive Management of Fragile Island Tourism Ecosystems

The ARIS paper makes four significant research contributions:
1. Closed-Loop Governance Architecture It proposes the first IoT–Digital Twin–AI pipeline that converts real-time ecological sensing into adaptive tourism permit decisions (Sensor → Data → Twin → AI → AGCI → Governance), moving beyond static dashboards to actionable governance.
2. Socio-Ecological Digital Twin Unlike conventional environmental twins, ARIS deploys three coupled sub-twins—Ecological, Tourism (agent-based), and Governance—creating a bidirectional feedback loop where tourism pressure is modelled as an endogenous system variable rather than external forcing.
3. AGCI as a Computational Governance Index The Adaptive Governance Capacity Index mathematically operationalises ecological state (E, R) against tourism pressure (I) and seasonal stress (S) to compute daily adaptive permit ceilings, replacing fixed carrying capacity with dynamic, data-driven thresholds.
4. Human-in-the-Loop AI Decision Support The AI layer (prediction, scenario, optimisation, recommendation agents) is explicitly designed as decision-support, not decision-authority —ensuring algorithmic transparency and institutional human oversight, which is critical for governance ethics in developing-country contexts.
The proof-of-concept demonstrates that under high tourism pressure (1,200/day), AGCI declines from 0.35 to 0.18 within 45 simulated days, triggering progressive restrictions—validating the framework’s capacity to prevent ecological overload without requiring permanent bans.

An Integrated Edge AI and IoT System for Quality Control, Waste Reduction and Predictive Maintenance in Small-Scale Food Processing

Small-scale food processing industries in
Rwanda and Malawi face persistent challenges in quality
control, energy efficiency, and supply chain management.
Traditional quality assessment methods are manual,
labour-intensive, and error-prone, leading to high
rejection rates, increased waste, and operational
inefficiencies, while the absence of real-time monitoring
results in excessive energy consumption and production
delays. This paper presents the design and development of
an integrated system that combines Artificial Intelligence
(AI) and the Internet of Things (IoT) to enhance
manufacturing efficiency, reduce waste, and improve
product quality. An AI-powered machine vision pipeline
built on Convolutional Neural Networks (CNNs) and
YOLOv8 performs real-time defect detection and
classification of maize grains, while IoT sensors monitor
machine health, environmental conditions, and energy
usage on the production line. Long Short-Term Memory
(LSTM) networks applied to sensor time-series data
forecast equipment failures, enabling predictive rather
than reactive maintenance. The system is deployed on edge
AI hardware, allowing on-site inference without
dependence on continuous cloud connectivity. Field
context motivates the work: post-harvest maize losses in
sub-Saharan Africa range between 15 and 30 percent,
roughly 25 percent of maize grains in Malawi are affected
by aflatoxin contamination, and 18 percent of processed
maize flour samples in Rwanda failed national quality
standards. Manual inspection processes approximately 100
kg of maize per hour against more than 1,000 kg per hour
for the automated system, a tenfold efficiency gain. The
solution is designed for affordability and scalability across
small and medium-sized enterprises, contributing to food
safety, waste reduction, and industrial competitiveness in
Rwanda and Malawi

Unsupervised EEG-Based Subgroup Discovery in Alzheimer’s Disease Using Spectral and Temporal Features

This study contributes an unsupervised, subject-level EEG framework for identifying reproducible electrophysiological subgroups within Alzheimer’s Disease. Unlike conventional AD classification approaches, it integrates multi-domain EEG features, PCA-based dimensionality reduction, K-Means clustering, bootstrap consensus stability analysis, and split-half validation to ensure that the discovered subgroups are both meaningful and reproducible. This provides a systematic approach for characterizing intra-AD heterogeneity and may support future research on disease progression and personalized clinical strategies.

Quantum-Resilient Self-Healing Cybersecurity Architecture for IoT-Enabled Smart Grids Using Federated Intelligence and Digital Twins

In this paper, the authors present a novel quantum-resilient self-healing cybersecurity system for smart grids with IoT technology, combining post-quantum security, federated intelligence, and digital twin technologies. The proposed framework facilitates distributed and private threat detection and ongoing monitoring of the grid’s behavior to detect any anomalous activities. Self-healing dynamically isolates adversely affected components and re-establishes secure operations with minimal service disruption. The architecture offers a scalable and adaptive security solution that helps establish the resilience of the smart grid to traditional and new quantum-related cyber attacks.

UGPDA and BDAI: Robust Medical Semi-Supervised Learning Under Varied Labeled Ratios and Class-Imbalance Profiles

Semi-supervised learning (SSL) holds huge potential for medical image classification by using vast pools of unlabeled data to reduce the burden of expert annotation. Standard SSL frameworks like FixMatch experience performance degradation when deployed under real-world clinical environments because of data scarcity and class imbalance. In these environments, unfiltered pseudo-labels create confirmation bias, resulting in the model learning the majority classes. To overcome these limitations, we present a unified framework integrating two novel modules: Uncertainty Guided Pseudo-Label Acceptance (UG- PDA) and Belief Distribution Augmentation Invariance (BDAI). UGPDA acts as a targeted entropy gate that checks the validation status of weak-view predictions using Shannon entropy to filter out high-uncertainty samples. On the other hand, BDAI preserves the full soft probabilistic target distributions (belief distributions) of accepted samples. It utilizes an asymmetric Kullback-Leibler divergence constraint to align strongly augmented views without forcing minority class nuances into rigid one-hot targets. By testing on chest X-ray dataset under varied labeled ratios (5%, 10%, 15%, and 20%) shows that our approach significantly outperforms state-of-the-art baselines. At an extreme 5% labeling within a native 3 : 1 imbalanced distribution, our framework achieves an F1-Score of 0.7521, an AUC-ROC of 0.9523, and a clinical Recall sensitivity of 0.9974. This establishes robust diag- nostic stability where traditional techniques undergo catastrophic class collapse.

A Comprehensive Benchmark of Recurrent, Convolutional, and Attention-Based Models for Radar Emitter Classification Under Photon-Starved Conditions

The identification of radar scanning patterns from
pulse amplitude sequences is a critical function in electronic in
telligence and electronic warfare. While deep learning has proven
highly effective for this task under additive white Gaussian noise,
its performance in photon-starved sensing environments remains
largely unexplored. This work presents a systematic benchmark
of eight deep learning architectures—LSTM, BiLSTM, GRU,
xLSTM, CNN, TCN, InceptionTime, and Transformer—under
identical quantum-noise-limited conditions. A physically realistic
Poisson noise model, incorporating pulse amplitude distortions,
missing pulse events, spurious outliers, and frequency-dependent
gain variations, is used to generate nine distinct noise scenarios.
Each architecture is trained and evaluated over ten indepen
dent runs with consistent preprocessing and training protocols.
Experimental results show that all models achieve 86–89%
accuracy under classical conditions but suffer a significant
degradation to 47–60% accuracy under quantum noise. Among
the evaluated architectures, InceptionTime achieves the highest
mean quantum accuracy (56.48%), while Transformer exhibits
the lowest (45.82%). Statistical analysis reveals no significant
performance differences across architectures, indicating that the
observed degradation is architecture-independent and fundamen
tally driven by photon-counting statistics. These findings establish
a clear performance bound for amplitude-only classification
under quantum-limited sensing and underscore the urgent need
for phase-aware or coherent detection strategies in future radar
emitter classification systems.

Provable Moving-Target Defense for Cyber-Physical Networks: From the Security–Performance Frontier to No-Regret Defense

The paper makes CPS moving-target defense quantitative and provable. It formalizes randomization as a game with an explicit physical-cost term and proves a security–performance frontier J⋆(p) attained by interpretable capped water-filling — a full convex curve rather than the single equilibrium point of prior grid-MTD work. Generalizing the idealized exact-match model to a graded confusion kernel yields the paper’s sharpest result: feasibility only above val(K), a matrix-game security floor that no physical budget can buy past (≈ 0.21 on the grid). A no-regret mirror-descent defender then reaches that frontier against an adaptive attacker with unknown payoffs, the novelty being the reduction rather than the standard √T rate. Because cost and security both derive from the same IEEE 14-bus physics, the price of security is stated in megawatts.

Forecasting Independent Evolution Possibilities of Code Clones

This paper is the first to forecast whether a code clone fragment’s next change will propagate to its sibling copies (dependent evolution) or happen alone (independent evolution), predicted before the change is made using only prior revision history. Existing approaches predict consistency only at copy-and-paste time, cover only copy-paste clones, and rely on shuffled cross-validation that leaks future data into training. We instead consider every clone fragment in a system, predict a recurring modification-level event across the clone’s entire lifetime, and enforce a strict walk-forward protocol that retrains after every revision with no look-ahead. Each fragment is described by 52 features capturing its edit history, pairwise coupling with siblings, and how that coupling decays over time. Across 33,003 modification events in four open-source C and Java systems, our best classifier reaches MCC 0.705 and balanced accuracy 0.850. All datasets and tools are publicly released.

Reputation over Geometry: A Validation-Guided Robust Federated Defense for Imbalanced Drone Intrusion Detection

The paper’s main contribution is a reframing of the aggregation criterion: it replaces the geometric question (“is this update an outlier?”) with a behavioral one (“does this update improve rare-class recall on a balanced probe?”). This dissolves the imbalance–robustness tension, since an honest rare-class client and a poisoner are indistinguishable by position but separable by behavior. RAVEN operationalizes this as probe scoring → tempered-softmax reputation → blended aggregation, paired with a federated-native Balanced-Softmax objective. Empirically it holds minority recall at 0.770 under strong ALIE collusion where Krum collapses to zero, with the margin widening as the attack strengthens.

The second, arguably more durable contribution is negative: the probe-optimal white-box adversary drives every evaluated defense to zero minority recall, and the natural geometric fix is shown to be structurally unavailable, since honest non-IID clients reach 0.99 cosine similarity against colluders’ 1.00. This converts what could have been a method-specific weakness into a boundary result on the entire aggregation-defense class, and points future work toward signals outside the update geometry — secret rotating probes, system-level attestation.

BAGC-FL: Benign-Anchor Gradient Correction for Imbalance-Aware Federated Intrusion Detection in IoMT Networks

The significant research contribution of this study is the BAGC-FL framework, a server-side Benign-Anchor Gradient Correction mechanism that mitigates the adverse effects of severe class imbalance and label-skewed non-IID data in federated IoMT intrusion detection. Unlike conventional FedAvg and class-weighted training, BAGC-FL selectively removes the gradient component that increases benign-anchor loss while preserving the original model architecture, local BCE objective, and communication protocol. Its effectiveness is demonstrated through a controlled E0–E2 ablation, IID/non-IID evaluation, paired statistical testing, and post-hoc SHAP/LIME analysis.

Lightweight and Explainable AI-Driven Intrusion Detection for IoT Networks: A Resource-Efficient Machine Learning Framework with SHAP-Based Interpretation

This paper presents a lightweight intrusion detection framework for IoT networks that combines machine learning-based detection with explainable AI analysis. The study evaluates several low-complexity classifiers, including Logistic Regression, Decision Tree, Random Forest, Gaussian Naive Bayes, and LightGBM, to identify a suitable model for resource-limited IoT environments. The proposed approach incorporates feature optimization, leakage-aware preprocessing, class balancing, and memory-efficient data handling to reduce computational requirements while maintaining detection capability. In addition, SHAP-based analysis is applied to explain the factors influencing intrusion decisions, improving the transparency of the detection process. The framework is evaluated using the CICIoT2023 DDoS-ICMP Fragmentation dataset, demonstrating how lightweight machine learning models can be used for accurate and interpretable IoT security monitoring.

AI-Enhanced Digital Twin with Hardware-in-the-Loop Deployment for Real-Time Lithium-Ion Battery State-of-Health Estimation

This work presents an AI-enhanced battery Digital Twin combining an aging-adaptive equivalent circuit model, an SOH-coupled Unscented Kalman Filter, and a systematic seven-model comparison for SOH prediction, validated not only through leave-one-battery-out cross-validation on the NASA dataset but also through live deployment on a physical hardware-in-the-loop sensor rig with closed-loop fan actuation. Unlike most reported battery Digital Twins, which are validated purely in simulation, this system demonstrates that a physics-informed, ML-benchmarked architecture can be deployed and cross-checked on real sensor hardware in real time.

ExplainYOLOv11: A Robust and Interpretable YOLOv11 Framework for Real-Time Traffic Sign Detection in Autonomous Vehicles

A new YOLOv11 traffic sign detection framework for autonomous driving and intelligent transportation systems (ITS) with an explanation is introduced. The proposed approach not only enhances the accuracy of the detection system but also incorporates the possibility of evaluating the robustness of the system and performing a computational efficiency analysis, along with the ability to explain why the system made the decision it did. The framework undergoes systematic testing and benchmarking against the existing YOLOv5 – YOLOv11 with the same experimental conditions and under challenging real-world conditions such as low-light, noise and occlusion. To increase transparency in safety-critical applications, the proposed model will include Grad-CAM and attention-based explanations. The YOLOv11 model that has been developed achieves excellent detection results with an mAP@0.5 of 0.989 and has been adeptly deployed through a real-time web-based deployment prototype, showcasing the model’s practicality.

A Hybrid Modified DES–Keccak Cryptographic Framework with Dynamic S-Box Design for Secure Data Confidentiality and Integrity Verification

This research presents a hybrid Modified DES–Keccak cryptographic framework that combines enhanced symmetric encryption with hash-based integrity verification in a unified architecture. The proposed method improves the conventional DES structure by introducing a 128-bit block and key mechanism, a round-dependent dynamic S-box generation approach using Keccak-256 hashing, and enhanced Feistel-based encryption operations. Unlike traditional DES variants that rely on fixed substitution tables and separate integrity mechanisms, the proposed framework integrates dynamic substitution and ciphertext authentication within a single design. The contribution of this work lies in strengthening confidentiality through improved cipher randomness and confusion while providing an additional layer of tamper detection using Keccak-256-based hash verification. The framework is experimentally evaluated in terms of computational performance, security characteristics, and integrity detection capability, demonstrating its potential as a secure alternative for modern data protection applications.

PULSE-ELF: Temporal Channel Learning and Low-Overhead Feedback for Integrated Sensing and Communication in Commodity Wi-Fi

In this paper, we propose PULSE-ELF, a lightweight Integrated Sensing and Communication (ISAC) system whereby commodity Wi-Fi hardware enhances both the sensing capability and communication efficiency. To tackle the challenges in them, the work proposes to use temporal CFR learning to capture the dynamics of channel variations to perform robust sensing, and introduce a select input low-overhead channel feedback method to reduce the channel feedback cost by sending informative parts of channels only. To achieve this, the proposed framework combines temporal feature extraction, few-shot adaptation, and an efficient channel reconstruction, all of which are required for the practically deployable ISAC framework, relying on the ubiquitous Wi-Fi infrastructure. Through the experimental results, we show that the proposed approach can achieve more than 99% sensing accuracy and up to 96% reduction in feedback overhead, which is effective for scalable and resource-efficient Wi-Fi sensing and communication.

Analysis of Various Machine Learning Model Performances for Epilepsy Prediction

The paper presents a comprehensive comparative framework demonstrating the critical impact of feature engineering (specifically chi-square and p-value statistical testing) on classical machine learning classifiers for clinical epilepsy type prediction. By bridging raw clinical patient attributes with proper data normalization, the study successfully elevates Support Vector Machine (SVM) accuracy from a baseline of 47% up to 80%, matching the overall top-performing Random Forest and Logistic Regression models at 81% accuracy

Paddy Seeds Classification using Machine Learning Models for Public Health

The paper establishes an automated computer vision classification baseline for public health and quality control across ten distinct rice cultivars utilizing transfer learning. By training and testing deep state-of-the-art models (GoogleNet, ResNet-50, and VGG-16) on automated photographic seed datasets provided directly by the Bangladesh Rice Research Institute (BRRI), the framework achieves a stable classification accuracy up to 78% without relying on time-consuming manual feature extraction.

Machine Learning Applications in the Classification of Dates

his work presents an experimental deployment of deep learning architectures (Custom CNN, GoogleNet, and ResNet-50) to mitigate multi-class classification challenges caused by data imbalances in highly similar qualitative food items. The study maps real-world nutritional variances across seven types of commercial dates, proving that GoogleNet delivers the highest robust validation accuracy of 98.57%, making it an effective baseline tool for precision agriculture and food tracking.

Elevating Skin Cancer Diagnosis: A Study of Various Machine Learning Approaches

The paper introduces a highly competitive multi-branch hybrid deep learning network integrating pre-trained EfficientNetB3, DenseNet121, and a customized Multi-Head Attention module to capture complex spatial linkages in dermoscopic skin lesions. By merging multi-scale textural features and global contextual correlation maps, the system successfully circumvents subjective diagnostic biases, capturing complex visual details to deliver a balanced 92.6% overall accuracy on the HAM10000 dataset for binary screening.

A Comparative Deep Learning Approach: CNN And Hybrid Algorithms For Lung Cancer Classification In Histopathology Images

The study designs a novel hybrid deep learning model (CNN-VGG16-ViT) that simultaneously resolves local tissue morphology extraction challenges and global multi-scale visual dependencies in histopathology samples. Evaluated against a highly balanced multi-class dataset of 15,000 images, this integrated local-global feature representation outperforms standard custom CNN baselines by boosting classification accuracy from 96.03% to 98.47% while yielding a robust micro-average AUC of 0.9913.

A Closed-Loop Supply Chain Model for End-of-Life Electronics under Carbon Tax Policy

Significant research contributions of this manuscript are as follows:
•Developed an integrated closed-loop supply chain (CLSC) model for end-of-life electronic products, combining forward and reverse logistics within a unified framework.
•Introduced distributor-level reintegration of repaired and recycled products, enabling recovered products to re-enter the forward supply chain earlier than in conventional models.
•Formulated a multi-period, multi-echelon Linear Programming (LP) model incorporating suppliers, plants, distributors, retailers, return-processing, repair, recycling, and secondary-market facilities.
•Integrated a comprehensive carbon-tax mechanism by applying CO₂-related costs to transportation across all forward and reverse logistics links.
•Combined economic and environmental considerations by minimizing total supply-chain cost while accounting for transportation emissions and revenue from scrap recovery.
•Demonstrated model applicability through a city-based case study, involving multiple suppliers, plants, distributors, retailers, customer zones, and reverse-logistics facilities.
•Established model robustness through sensitivity analysis of customer demand, carbon-tax rate, and emission factor, showing stable and predictable cost responses.
•Provided quantitative evidence on carbon-tax impacts, showing only a small increase in total supply-chain cost under the specified case-study conditions.

Interpretable Short-Term Electricity Demand Forecasting for the PGCB Grid Using Feature Fusion

The main contribution of this study is a feature-engineered short-term electricity demand forecasting framework that integrates historical demand, temporal/calendar, and weather information, combines rigorous leakage-aware chronological evaluation with machine learning optimization, and provides interpretable predictions using SHAP analysis for the PGCB grid.

30-Day-Ahead Load Forecasting for the Rajshahi Zone of the Bangladesh Power System: A Linear-Base Boosted BiLSTM Approach

This paper develops a 30-day-ahead daily peak load forecasting framework for the Rajshahi zone of the Bangladesh Power System, a horizon that governs maintenance scheduling, fuel procurement and outage planning but has received far less attention than day-ahead forecasting, and almost none for supply-constrained tropical grids.

Three contributions are made. First, a regularised linear stage is fitted before gradient boosting, so that the 4.2% per year growth trend is carried by a learner that can extrapolate it; this removes an extrapolation bias worth 0.69 percentage points against plain XGBoost on identical features. Second, supply-side outages are screened by a local V-drop criterion so that load shedding is not learned as demand, and no reported score is computed against an imputed target. Third, the worth of the target-date weather forecast is measured rather than assumed, at 2.80 percentage points, which is larger than the spread across the twenty-five models compared and identifies forecast-weather quality rather than model choice as the dominant lever at this horizon.

Evaluation is on a strictly chronological 2025 hold-out year against twenty-four benchmarks, with Diebold-Mariano tests, four-fold rolling-origin validation and seed-averaged deep models.

EndoStack-XAI: A Hybrid and Explainable Machine Learning Framework for Endometriosis Prediction

This study proposes EndoStack-XAI, a leakage-safe two-level stacking ensemble for interpretable endometriosis prediction using demographic, reproductive, symptom-related, laboratory, and clinical features. The framework combines RBF-SVM, Random Forest, and Extra Trees as base learners with Logistic Regression as a meta-learner trained on leakage-controlled out-of-fold predictions. SHAP and LIME are further integrated to provide global and local model interpretability while maintaining strict separation of the held-out test set from model development.

Toward Privacy-Preserving Federated Brain Tumor MRI Classification: A Hybrid CNN – Transformer Framework and Preliminary Evaluation

The main contributions of this study are summarized as
follows:
• Hybrid CNN–Transformer Framework: We propose
HybridNeuro, integrating ResNet50, CBAM, and Trans-
former encoder blocks for four-class brain tumor MRI
classification, with a design intended for federated de-
ployment.
• Privacy-Preserving FL Formulation: We formulate an
FL framework incorporating client-side DP-SGD, keep-
ing MRI data local and providing a basis for subsequent
privacy–utility analysis.
• Preliminary Evaluation and Failure Analysis: We
establish a centralized baseline and analyse class-wise
performance, prediction confidence, and model decision
behaviour using confusion matrices and XAI, with the
findings informing class weighting and stratified parti-
tioning for the planned FL+DP evaluation.

Physics and Communication-Aware Remote Terminal Unit Cyberattacks Detection and Classification for Smart Energy Grids

Paper contributions
(i) A causal feature set combining robust temporal statistics, three-phase power consistency, spatial RTU disagreement, and communication evidence;
(ii) Explicit fault-versus-cyber rules whose triggers have engineering meaning;
(iii) Stateful persistence that prevents rolling-baseline adaptation from prematurely clearing sustained events; and
(iv) An attack-wise evaluation in which every waveform and timeline explicitly identifies normal operation, the normal electrical fault, and each cyberattack.

Explainable AI-Based Thyroid Diagnosis Classification Using SHAP-Guided Feature Reduction

In this study, we developed an explainable machine learning framework for benign–malignant thyroid diagnosis classification using structured tabular data. We evaluated seven classifiers through a consistent preprocessing and class-imbalance pipeline.
CatBoost achieved 82.93% test accuracy, 69.92% balanced accuracy, and 55.41% F1-score. Our three-fold cross-validation produced 82.69% mean accuracy and confirmed the close performance of the leading boosting models.We used SHAP to identify eight influential predictors, and the corresponding Top-8 representation preserved the principal classification metrics of the original 14-feature model. Our central contribution is therefore a compact and interpretable feature representation rather than a new learning algorithm.

AI-Enhanced Smart Grid Integration for Solar Energy: A Comprehensive Bibliometric Analysis

Solar energy is now a large part of the smart grid. Its variability makes forecasting and stability harder. Artificial intelligence (AI) and machine learning (ML) are widely used to manage these problems. This paper maps that literature. We analyzed 273 Scopus-indexed documents published between 2010 and 2024. VOSviewer and Biblioshiny were used for keyword co-occurrence, citation, bibliographic coupling, cocitation, co-authorship and thematic analysis. The results reveal an exponential growth in publications (28.64% annual growth rate), with China, India, and the United States as the leading contributors. Three clusters were found: smart grid and solar fundamentals, AI-based energy management, and deep learning for solar forecasting. Smart power grids and solar energy appear as basic themes. AI and energy management remain niche but well developed. Keyword trends show a shift from shallow neural networks to deep and reinforcement learning. The review outlines influential works, collaboration patterns and open gaps for future research in AI-enabled solar smart grid integration.

BreastSegX: A Pipeline for High-Fidelity Breast Ultrasound Lesion Segmentation

The main contributions of this work
are summarized as follows:
• We pre-trained EfficientNet-B4 [6] on 3,369 external
breast ultrasound images to improve feature representa-
tion before fine-tuning on the BUSI dataset.
• The model was trained using 5-fold cross-validation with
class balancing and data augmentation, then evaluated on
a held-out 10% test set (78 images).
• For post-processing, we applied morphological closing
and connected-component analysis to reduce noise and
smooth tumor boundaries.
• Finally, we proposed specific evaluation metrics for “Nor-
mal” scans to avoid division-by-zero errors and to report
all classes consistently.

EXPLORING POSSIBLE ADOPTION FACTORS OF mHealth APPLICATIONS AMONG WOMEN IN BANGLADESH

The study extends the base UTAUT model by incorporating seven additional independent variables — Technology Anxiety, Habit, Trust, Personal Innovativeness in IT, Privacy, Price Value, and Resistance to Change , that are not part of the original framework, tailoring it specifically to the mHealth context. It also modifies the moderating structure by removing Gender (since the sample is women-only) and introducing Education as a new moderator alongside Age, Experience, and Voluntariness of Use.Using PLS-based structural equation modeling on survey data from 491 women, the study empirically identifies which factors significantly drive adoption and which do not, offering a more nuanced picture than the base UTAUT model alone.

Evaluating Synthetic Oversampling Strategies for Imbalanced Thyroid Disease Classification with SHAP-Based Explainability: A Comparative Study Using Ensemble Classifiers.

Here are the main points in simple terms:

The Problem: The thyroid dataset had very few sick patients (only 3.7%), making it hard for normal AI models to spot the disease.

The Solution: The authors tested different ways to create smart, artificial data for rare sick cases and combined them with 5 popular machine learning models.

Best Result: Combining Random Forest with ADASYN or SMOTE worked best—it caught 98% of sick patients (up from 90%) without making many false alarms.

Key Finding: While creating artificial data helps the model catch more sick cases, statistical tests showed it doesn’t fundamentally make the model smarter at ranking risk.

Medical Check (SHAP): Using an explainability tool (SHAP), they proved the AI is making decisions based on real medical blood markers (like TSH, T3, and T4), not random noise from the fake data.

Optimization of operational parameters of incline plate planter via Response surface methodology

Optimisation of operational parameters of the incline plate planter for increasing single seeding, reducing double seeding and missing

An Empirical Study on the Impact of Authorship on the Bug-proneness of Code Clones

A common technique in software development is code cloning, which allows for the reuse of pre-existing code. However, frequently it comes at the cost of more difficulties and unnoticed flaws (bugs). The effect of code clones on software quality has been thoroughly examined in the existing studies, but the relationship between clone authorship and bug proneness has received less attention. Code fragments cloned by developers other than the original author are intuitively more likely to contain bugs than those cloned by the original author. However, there is no empirical evaluation of this intuition. In order to bridge this research gap, we have conducted an empirical study to analyze the impact of clone authorship on the bug-proneness of code clones. For the purpose of our study, we have mined a particular clone evolutionary pattern from thousands of revisions of our subject systems. The pattern helps us discover those clone fragments that were created by one programmer through copy/pasting another programmer’s code. We have analyzed the bug-proneness of such clones. According to our analysis on five diverse subject systems, if a programmer copy/pastes and reuses another programmer’s code, the copy/pasted code fragment is more likely to contain bugs compared to a code fragment which was created through copy/pasting by the original programmer. Our research provides actionable suggestions regarding the cloning activities of programmers towards minimizing bugs in code clones.

Comparative Analysis and Shunt-Reactor Mitigation of the Ferranti Effect in 132-kV Overhead and Underground Transmission Lines

This work develops a unified MATLAB/Simulink framework to evaluate and directly compare the Ferranti effect in 132-kV overhead lines (OHL) and underground cables (UGC) under identical operating conditions. The study quantifies that underground cables exhibit nearly eight times higher voltage rise and require approximately twenty times greater shunt-reactor compensation than equivalent overhead lines. These findings offer planning and protection engineers a quantitative, technology-specific framework for allocating reactive-power compensation in hybrid OHL-UGC sub-transmission corridors.

OmniCortex: A Hybrid Framework for Explainable Multimodal Brain Tumor Classification

The significant research contribution of this study is the development of OmniCortex, a hybrid ResNet50–CNN framework for automated brain tumor classification using multimodal CT and MRI images. The proposed framework introduces a spatial tensor reconstruction layer to preserve spatial information and address cross-modal feature alignment limitations, while a custom CNN refinement head further enhances the high-level features extracted by ResNet50. The model was evaluated on 9,618 CT and MRI images and achieved 98.23% testing accuracy, outperforming the evaluated baseline architectures without requiring artificial data balancing or manual radiomic feature engineering. Furthermore, the integration of Grad-CAM provides visual interpretability by demonstrating that predictions are focused on clinically relevant tumor regions rather than background artifacts, thereby improving the transparency and potential clinical utility of the proposed system.

Multi-Class MRI Brain Tumor Classification Using a Hybrid EfficientNet-Attention Architecture

The significant contribution of this research is the development of EMA-BrainNet, a lightweight hybrid deep learning framework that integrates an EfficientNet-B0 convolutional backbone with a Multi-Head Attention (MHA) mechanism for accurate multi-class brain tumor classification. Unlike conventional CNN-based approaches that primarily focus on local features, the proposed framework simultaneously captures detailed anatomical patterns and long-range spatial dependencies, enabling more effective discrimination among glioma, meningioma, pituitary tumor, and healthy brain MRI images. The study further introduces a systematic preprocessing pipeline involving CLAHE, standardized image resizing, Z-score normalization, and class-preserving data augmentation to improve feature quality and model generalization. Experimental evaluation on a balanced dataset of 8,000 MRI images demonstrates that EMA-BrainNet achieves 99.97% training accuracy, 98.88% validation accuracy, and 99.38% test accuracy, outperforming the evaluated baseline models and recent competing approaches. In addition, the integration of Grad-CAM provides visual explanations of the model’s decisions by highlighting clinically relevant tumor regions, improving the interpretability of the automated diagnostic system. Overall, the research contributes an efficient, accurate, and interpretable framework with strong potential for computer-aided brain tumor diagnosis.

Traffic Demand Regulation by an Intelligent Management System: A Case Study in Dhaka

Traffic congestion in urban cities is acute, resulting in extensive delays, elevated fuel usage, and pollution due to mismanaged signals and unregulated traffic systems. These inhibitions primarily stem from ungoverned signal timing and unpredictable traffic patterns. An Intelligent Traffic Control System (ITCS) is developed here as a pilot study by utilizing field-observed traffic data from Gulshan-2 Circle, Dhaka. Machine learning (ML) techniques were used to recreate a daylong traffic demand profile from data obtained in specific time periods, while preserving temporal variability and realistic demand features. Three ways of synchronized traffic control are offered. First, a Traffic Demand Regulation Model (TDRM) based on incentives is proposed to transfer part of the trips in the peak periods to the nearby off-peak periods. Second, staggered scheduling is proposed for academic institutions, offices, and commercial organizations by altering opening and closing times to lessen simultaneous travel demand. Third, an Adaptive Traffic Signal Control (ATSC) technique is proposed to reduce the queue length and delay by dynamically adjusting the signal timing according to the traffic conditions. Mathematical models for TDRM, staggered scheduling, and ATSC have been built and assessed inside a MATLAB simulation environment. Comparative simulations of baseline and intervention scenarios show that transferring only 1-10% of peak period demand to off-peak periods may significantly reduce congestion and coordinated control methods can cut fuel consumption and CO2 emissions by about 10-20%. Integrated demand management, scheduling, and adaptive control can improve urban traffic efficiency under realistic heterogeneous urban traffic conditions and support sustainable transportation in Dhaka, contributing to SDGs 11 and 13.

Benchmarking Deep Learning Models for IoMT Intrusion Detection with McNemar Testing and Explainable Artificial Intelligence

The main contribution of this study is not a new architecture but a clearer picture of what existing ones actually achieve on IoMT traffic. By training seven deep-learning models under a single disclosed protocol on CICIoMT2024 and comparing them with McNemar testing under Holm correction, we show that the differences between architectures, while statistically significant across all twelve pairwise comparisons, are often practically small once the size of the test partition is taken into account. More importantly, the results expose a gap that weighted reporting conceals: the best multiclass model reaches 98.01% weighted F1 but only 58.81% macro F1, and fails to detect the rare Recon Ping Sweep class at all. Since much of the surrounding literature reports weighted or overall accuracy alone, this gap suggests that current performance figures for IoMT intrusion detection may be considerably more optimistic than per-class behaviour warrants. LIME explanations of selected predictions, together with an explicit note that the feature-sequence reshaping used by six of the models carries no genuine temporal meaning, are included to keep the interpretation of these results appropriately bounded.

LeukemiaCellNet: A Patient-Aware Multi-Scale CNN–Transformer Framework with Attention-Based Aggregation for Leakage-Free AML Cytomorphology Classification

In this work, LeukemiaCellNet is a patient-aware multi-scale CNN–Transformer framework for leakage-free AML cytomorphology classification. The proposed method fuses a multi-scale CNN with a global Transformer to extract the local morphological features and global contextual features, respectively, to represent the complementary attributes of cells. To integrate cell-level features, an attention aggregation mechanism at the patient level is introduced, and consistency regularization is used to enhance the stability of prediction in different patient samples. Patient-disjoint evaluation addresses the limitation of information leakage and allows obtaining a more accurate evaluation of the use of AI for AML classification in clinical decision support.

An IoT-Enabled Smart Cold Storage System for Real-Time Environmental Monitoring and Automated Control of Perishable Commodities

The study develops and experimentally validates a low-cost IoT cold storage system that integrates multi-parameter sensing, automated environmental control, spoilage-gas detection, and local/cloud-based alerts. The prototype maintains temperature within ±1°C of the 4°C setpoint, achieves 93% humidity-range compliance, and provides an average 2.8-second alert response, demonstrating its practical potential for resource-constrained cold storage facilities.

Assessment of Near-Field and Far-Field Electromagnetic Exposure in Daily-Use Wireless Devices

This research provides a unified comparative analysis of near-field and far-field electromagnetic exposure from common wireless devices. It evaluates both SAR and power density, compares the calculated exposure levels with ICNIRP and IEEE safety standards, and highlights the influence of distance, frequency, and device power. The study also provides practical safe-use recommendations to improve public awareness of electromagnetic exposure.

A Dual-Band Rectenna with High RF-to-DC Conversion Efficiency for Ambient Wireless Power Harvesting

1) A compact 50 mm × 45 mm dual-band antenna is
presented, with a simulated reflection coefficient below
−23 dB at 2.44 GHz and 4.71 GHz on low-cost FR-4,
and radiation efficiencies of 89% and 78–80% at the two
bands, respectively.
2) A dual-band T-junction matching network is designed
to simultaneously transform the 50 Ω antenna-port
impedance into the complex, frequency- and power-
dependent impedance of a voltage-doubler rectifier.
3) A complete rectenna is designed and verified through
harmonic-balance circuit simulation, achieving peak RF-
to-DC conversion efficiencies of 90% at 2.44 GHz and
88% at 4.71 GHz, on par with or exceeding recently
reported rectenna designs.

Impedance-Based Stability Boundary for Two-Stage Solid-State-Transformer EV Fast Chargers

Two-stage SST fast charger is appealing for high power EV charger since it provides regulated front-end converter followed by another DAB isolated stage, however, the common dc link can cause instability even if both front-end converter and DAB are individually stable. This paper will delve into the interaction through impedance-based analysis and present the closed-form expressions for the DAB input and PFC output impedances by cascading their PI-controlled dynamics. The minor-loop gain obtained is utilized to determine a realistic stability limit of a 10~kW 800/400~V charger. The nominal design has been validated to have a 10.4~dB interaction margin and $|T|_{\max}=0.301$. The minimum capacitance of the dc-link depends on the converter: $C_{dc,pfc}>12.55~\mu\mathrm{F/kW}\times P_0$; for the nominal $C=414.5~\mu$F, the converter is operated at $3.32\times$ above the minimum capacitance. A sweep of the PFC equivalent controller reveals that the critical bandwidth of the PFC voltage-loop is around 6~Hz, which indicates that the cascaded bandwidth must not be reduced too much or the cascade will be unstable. The accuracy of the predicted boundary is verified by an independent 13-state nonlinear model , whose deviations in critical capacitance and boundary frequency are 0.01\,\% and 2.9\,\% respectively . Furthermore, the boundary is validated through nonlinear free-response and switching-level impedance verifications.As shown, the results indicate that after crossing the instability boundary, the value of $\min_{\omega}|1+T|$ may be misleading whereas the value of $|T|_{\max}$ and the Nyquist plot encirclement remain reliable indicators of interaction.

A Hybrid Statistical and Machine Learning Approach for Optimizing Critical Quality A Hybrid Statistical and Machine Learning Approach for Optimizing Critical Quality

This paper advances pharmaceutical Quality by Design (QbD) by empirically demonstrating the structural failure of traditional linear OLS regression ($R_{adj}^2 = 0.00\%$) in multi-stage oral solid dosage (OSD) manufacturing due to severe multicollinearity (VIF up to 29.40), while establishing a robust machine learning benchmarking framework tailored for small industrial datasets ($N=30$). Utilizing a Leave-One-Out Cross-Validation (LOOCV) protocol to prevent data leakage, the study identifies Random Forest as the optimal predictive architecture, achieving the lowest Mean Absolute Error ($\text{MAE} = 0.8023\%$) and an RMSE of $1.1646\%$ for predicting finished tablet Assay (%). Crucially, the authors quantify cross-stage parameter propagation through feature importance metrics, isolating film coating Weight Gain ($25.6\%$), Water Content ($18.1\%$), and Moisture Content ($12.2\%$) as the primary drivers of active ingredient potency—delivering a regulatory-compliant tool to proactively mitigate Out-Of-Specification (OOS) risks in commercial production.

X-STREAM-IDS: An Explainable Streaming-Aware Deep Learning based Intrusion Detection System for IoT

The proposed X-STREAM-IDS makes the following key contributions:
1. Develops a dataset-aware multi-file attack taxonomy for the BoT-IoT dataset to support controlled multi-class intrusion labeling.
2. Introduces a robust feature encoding mechanism for handling heterogeneous and noisy IoT traffic.
3. Implements mini-batch online learning with noise-based regularization to simulate streaming traffic and reduce overfitting.
4. Integrates SHAP and LIME to provide interpretable and transparent intrusion detection decisions.

Brain Tumor MRI Classification Using Hybrid Model

• Implementation and evaluation of NeuroVision, a ResNet-
18 + Transformer-encoder hybrid, for four-class brain tu-
mor MRI classification, using genuine ImageNet transfer
learning for the CNN backbone.
• A controlled, within-notebook comparison against a fine-
tuned ResNet-18-only baseline trained with identical data
splits, optimizer, and hyperparameters.
• Identification and correction of a shared-dataset
transform-aliasing defect in the data pipeline, with a
description of the corrected implementation.
• A positioning of the results against recent (2026) open-
access literature on CNN, wavelet-enhanced CNN, Vision
Transformer, and 3D volumetric methods for brain tumor
MRI analysis.

An Explainable ML Framework for Five Customs Fraud Type Predictions Using XGBoost and SMOTE-Based Class Balancing

Customs fraud causes significant revenue losses and undermines fair international trade. Common fraud types include undervaluation, misclassification, false declaration of origin, and smuggling, yet most existing machine learning approaches only predict whether a declaration is fraudulent, without indicating the fraud type or explaining the prediction. This paper presents a multiclass and explainable machine learning framework for customs fraud detection. A dataset of 14,873 customs import declarations was extended from binary fraud labels to five classes: no fraud, false origin, undervaluation, misclassification, and smuggling. The framework applies missing-value handling, domainbased feature engineering, SMOTE-based class-imbalance treatment, and an ensemble feature-selection method combining ChiSquare, ANOVA, Recursive Feature Elimination, and SHAP. Several models were benchmarked under different preprocessing configurations and tuned via hyperparameter search. XGBoost achieved the best performance, with 96.0% accuracy, macroF1 of 0.840, and macro-AUCROC of 0.947. SHAP, LIME, and counterfactual explanations were integrated to make the predictions interpretable.

ASD-Predict: A Multi-class Explainable Autism Disorder Severity Prediction Using LightGBM and SMOTE

Timely assessment of autism disorder (ASD) severity
is significant for proper disease diagnosis and healthcare of a
person or child. While a sizable amount of machine learning
research has focused on autism screening, most of this work uses
a limited number of classifiers and seldom offers a methodical
comparison of learning (ML and DL) techniques for multiple
type of autism disorder severity assessment. In order to evaluate
autism severity across four levels, this study introduces ASDPredict, an adaptive learning framework that uses machine
and deep learning-based methods. To improve data quality and
model dependability, the framework employs comprehensive data
preparation techniques with encoding and feature scaling. Seven
predictive machine learning models (RF, XGBoost, LightGBM,
MLP, RNN, ANN and LSTM) are developed and assessed.
Stratified cross-validation, proper feature selection, and parameter tuning are conducted along with explainability results. The
experimental results highlighted that the suggested LGBM-based
framework offers an accuracy of 98.09 percent and outperforms
the traditional background work’s performance.

An AI-Enhanced Four-Class Stacking Ensemble Framework for Explainable Income Tax Fraud Detection Using Calibrated NBR IT-10B Structure

Revenue collection is severely hampered by income
tax fraud, especially in developing nations like Bangladesh. The
National Board of Revenue (NBR) has to deal with issues including severely skewed fraud statistics and a dearth of instruments
that can recognize various forms of fraud from the nation’s own
tax return system. The majority of current machine learning
research is mostly on binary fraud detection and offers scant
justification for their forecasts. Based on Bangladesh’s IT-10B
individual tax return structure and the 2025–2026 tax slabs, this
paper suggests a four-class income tax fraud classification scheme
that covers No Fraud, Underreporting, Inflated Deductions, and
False Credits. With the assistance of a practicing tax attorney,
a dataset of 2,498 records was created from anonymized tax
files and verified. Experts in the field were consulted in order to
further verify the fraud typology guidelines. A weighted Borda
aggregation of RFECV, LightGBM feature significance, Pearson
correlation, and Mutual Information was used to decrease the
initial 62 features produced by the proposed framework to
25. To solve class imbalance, SVMSMOTE was solely used on
training data. A pruned stacking ensemble was used to merge
five models (SVM, Random Forest, ANN, CatBoost, and Logistic
Regression), with Logistic Regression serving as the meta-learner.
The suggested model obtained a macro F1-score of 0.9502 and
an accuracy of 96.53% on the locked test set.It fared better
than the soft-voting ensemble (0.9272), the best individual model,
SVM (macro F1: 0.9343), and four re-implemented literature
baselines tested on the same dataset and test split. The model’s
general behavior and class-specific predictions were explained
using SHAP and LIME.

ASD a new era begins

research has focused on autism screening, most of this work uses
a limited number of classifiers and seldom offers a methodical
s performance.

URL-Only Phishing Detection Based on Structural and Character-Level Lexical Fusion Using TF-IDF and LightGBM

Phishing URLs are commonly used for credential
theft and online fraud. This study presents a URL-only phishing
detector using the raw URL and class label from the PhiUSIIL
dataset. The method combines LightGBM trained on 38 fixed
structural lexical features with Logistic Regression trained on
character-level TF-IDF URL representations. Component configurations, fusion weights, and classification thresholds are selected
through inner validation. Five registrable-domain/IP-disjoint
outer folds ensure that URLs from the same registrable domain or
IP-address group do not cross training–test boundaries. Across
234,604 cleaned URLs, the method achieved 99.81% accuracy,
99.96% precision, 99.59% recall, and a 99.78% F1-score from
pooled held-out predictions. All comparison models also received
inner-validation configuration and threshold selection. Holmadjusted exact paired McNemar tests showed significant paired
correctness advantages over six of seven comparison models; the
small numerical advantage over Extra Trees was not statistically
significant. The complete artifact averaged 7.60 MB, with batch
scoring of 0.191 ms per URL. In a contextual comparison with
selected PhiUSIIL studies, the proposed method achieved 0.02
percentage points higher accuracy than the highest reported
result, although evaluation protocols differ

Complexity-Aware Explainable Deep Learning for Skin Cancer Classification: A Dynamic Explanation Selector for Adaptive XAI on Dermoscopic Images

Skin cancer is among the most rapidly increasing malignancies worldwide, and early, accurate diagnosis is critical to improving patient survival. Conventional diagnosis relies on visual examination by dermatologists, a process that is time-consuming, subjective, and highly dependent on clinical expertise. Automated detection is further complicated by the inherent variability of dermoscopic images, including ambiguous lesion boundaries, heterogeneous colors and shapes, low contrast, and artifacts such as hair and shadows. This paper presents an automated, explainable deep learning framework for skin cancer classification from dermoscopic images. Three convolutional neural network architectures, including ResNet50, DenseNet121, and EfficientNetB3 are evaluated using transfer learning and data augmentation on the benchmark HAM10000 dataset, with EfficientNetB3 achieving the strongest performance among the three, reaching an overall test accuracy of 92% (weighted F1-score 0.92). To address the opacity of deep learning predictions, the framework integrates three Explainable AI (XAI) techniques: Score-CAM, Saliency Maps, and LIME and also introduces a Dynamic Explanation Selector (DES) that automatically identifies the most suitable explanation method for a given image based on a computed visual complexity score derived from edge density, texture contrast, color variation, and entropy. Feature separability across lesion classes is further examined using t-SNE visualization. Experimental results demonstrate that the proposed framework achieves high classification accuracy alongside interpretable, clinically meaningful visual explanations, positioning it as a promising decision-support tool for dermatological practice.

Deep Learning-Driven Detection of Temporal Lobe Seizures Using EEG Signals

Epileptic seizure detection using electroencephalogram (EEG) signals has received considerable attention because of the potential of automated systems to support continuous neurological monitoring and clinical decision-making. Among focal epilepsies, temporal lobe epilepsy is particularly important because the temporal lobe is a common site of focal seizure activity. However, many existing EEG-based approaches process all available channels simultaneously, potentially introducing redundant information and reducing the interpretability of the resulting models. This paper investigates a region-specific deep learning approach for automated seizure detection using EEG signals obtained from the temporal lobe. The study is derived from a broader brain-region-based seizure detection framework in which EEG channels are grouped according to anatomical regions and informative channels are selected using the Lion-Mapped Chaotic Whale Optimization (LMC-WWO) algorithm. The selected EEG signals are normalized, filtered, artifact-reduced, segmented, and transformed using Gramian Angular Summation Field (GASF) and Continuous Wavelet Transform (CWT) representations. Four deep learning architectures, namely Long Short-Term Memory (LSTM), One-Dimensional Convolutional Neural Network (CNN1D), Temporal Convolutional Network-1 (TCN1), and Temporal Convolutional Network-2 (TCN2), are evaluated for temporal-lobe seizure classification. A weighted ensemble of CNN1D and TCN2 is also investigated. Experimental results show that the LSTM model achieves the highest temporal-lobe classification accuracy of 97.62%, outperforming CNN1D (93.45%), TCN1 (91.67%), TCN2 (72.02%), and the weighted ensemble (78.57%). The findings indicate that temporal-lobe EEG signals contain sequential characteristics that can be effectively captured by recurrent temporal modeling. The study demonstrates the potential of region-specific EEG analysis for developing accurate and clinically interpretable seizure detection systems

A Lightweight Knowledge-Distilled CNN for Real-Time Waste Classification on Edge Devices

The exponential rise in solid waste generation, coupled with inefficient manual sorting practices, has intensified the need for automated waste classification systems that can operate reliably within resource-constrained Internet of Things (IoT) infrastructures. While deep convolutional neural networks (CNNs) have demonstrated strong performance in waste classification tasks, their substantial computational and memory requirements render them unsuitable for deployment on low-power edge devices commonly used in smart bin applications. To address this gap, we propose a lightweight CNN architecture trained via knowledge distillation to enable accurate, real-time waste classification on edge hardware. We first benchmarked thirteen pretrained CNN architectures, including VGG, ResNet, DenseNet, MobileNetV2, InceptionV3, Xception, EfficientNet, and NASNet variants, to identify the most effective teacher model. ResNet152 achieved the highest classification accuracy of 98.02\% and was subsequently employed to guide the training of a compact student network through a distillation-based learning framework. The proposed student model, trained under teacher supervision, achieved a classification accuracy of 93.35%, a substantial improvement over the 86.44% obtained when trained from scratch without distillation, while maintaining a significantly smaller parameter footprint suitable for embedded deployment. Experimental results confirm that the distilled model preserves much of the teacher’s discriminative capability while drastically reducing inference latency and memory overhead. The classifier categorizes waste into biodegradable, non-biodegradable, and background classes, making it directly applicable to intelligent smart bin systems for automated, real-time waste sorting in smart city and environmental monitoring applications, contributing toward sustainable and scalable waste management infrastructure.

A Dual-Branch Framework for Livestock Disease Screening Using Multi-Repository Visual Datasets and Symptom-Based Intelligence

• We design a seven-stage harmonization pipeline that consolidates four heterogeneous bovine disease repositories
into a single validated six-class dataset, addressing taxonomy standardization, annotation validation, duplicate
removal, class balancing, and cross-split data leakage
verification.
• We train a YOLOv8m-based multi-class bovine disease
detector on 11,804 harmonized images, achieving an
mAP@50 of 0.7136, and analyze its behavior across
confidence thresholds and disease classes.
• We develop a rule-based symptom assessment module,
built on a veterinary-derived disease–symptom pathology
matrix, that generates ranked differential diagnoses from
farmer-reported clinical observations without requiring
any image input, and we characterize its reliability and
failure modes under simulated ambiguous reporting

A FEATURE-INTEGRATED MACHINE LEARNING FRAMEWORK FOR LYSINE PTM SITE CLASSIFICATION

Post-translational modification (PTM) increases the functional diversity of proteins by
introducing new functional groups to the side chain of amino acid of a protein. The structural and
functional diversities of proteins as well as plasticity and dynamics of living cells are significantly
dominated by the post-translational modifications (PTMs). It plays an important role in diversity,
structure, plasticity, active cells even in human diseases and so on. PTMs are also responsible for
expanding the genetic code and for regulating cellular physiology. Like other PTMs, PTM of
lysine residues have proven to be major regulators of gene expression, protein-protein interactions,
and protein processing and degradation. So far so many computational methods have been
developed to identify PTM of lysine but most of them are binary classifier. It can predict various
single-label PTM sites, and a very few have been developed to solve multi-label PTM of lysine.
Our expectation is to design a simple and efficient predictor for predicting multiple lysine PTM
sites. Our goal is to get higher success rates in comparison with the existing predictors in this area.

Predicting Solar PV Power Under Real-World Conditions: A Comparative Evaluation of Interpretable Machine Learning Models

Solar power generation prediction plays a vital role in improving the stability of the power grid, energy management, and the integration of renewable energy with the existing power grid to achieve the required objectives. Due to its inherent variability and climate dependent nature, solar photovoltaic output poses a huge challenge for grid operators and energy planners such as the capacity of the energy producers to predict output with high accuracy and the capacity of the grid to adjust to the nonlinear relationship between environmental factors and PV output. In this study, the authors compare four popular machine learning models: Linear Regression, Decision Tree, Random Forest and Extreme Gradient Boosting (XGBoost) to predict solar power generation from a publicly available data set from the solar plant in South Asia. The data set consists of 68,778 samples with attributes such as irradiation, ambient temperature, module temperature and the target is the AC power. Experiments conducted in the same way for all models: Train/Test split 80/20, no shuffling of time series. The performance of the models was evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²). The experimental results prove that the model of Random Forest has the best results with an MAE of 34.700, RMSE of 53.261, and an R² of 0.976, which is quite close to the results of XGBoost with an R² of 0.974. As expected, with the basic physics of solar power generation, feature importance analysis showed that the irradiation was the most important predictor accounting for about 97% of the model’s predictive power. The irradiation and AC output power were correlated and found to be highly positive (0.98). The results indicate that the ensemble methods are not only accurate in predicting the results but also computational efficient and easy to interpret, thus they are suitable for use in resource-limited settings, like developing countries.

FedPdM: Privacy-Preserving Federated Learning for Remaining Useful Life Prediction Across Heterogeneous Industrial Systems

• We have introduced a unified DP-FL framework for
multi-domain RUL prediction that combines FedProx
with formal (ε, δ)-differential privacy;
• We have conducted a cross-architecture benchmark of five
deep learning models under federated non-IID conditions
on three industrial datasets;
• We have proposed three new DP-FL evaluation metrics,
plus evidence that personalized federation consistently
closes the gap to the centralized upper bound.

Web-based Explainable Deep Learning Framework for Skin Disease Detection

We have developed a controlled comparative framework
for evaluating six CNN and Transformer architectures on
the HAM10000 dataset under a consistent experimental
protocol.
• We have employed a two-phase transfer learning strategy
consisting of initial feature extraction with frozen backbones
followed by selective fine-tuning using a reduced learning
rate.
• We have incorporated random oversampling to address the
substantial class imbalance in the HAM10000 dataset and
improve the representation of minority disease categories
during training.
• We have integrated GradCAM++ into the classification
pipeline to provide class-specific visual explanations of
model predictions.
• We have evaluated the models using accuracy, macroprecision, macro-recall, macro-F1 score, macro-AUC, and
inference time to provide a comprehensive performance
comparison.
• We have integrated the classification and explainability
components into a responsive web-based application to
demonstrate an accessible prototype for dermoscopic image
analysis.

Archives: Paper Submission