Paper Submission - NPS Australia Submission System

A Low-Cost Embedded Data Acquisition System For Concrete RCP Test

The durability of reinforced concrete structures is strongly governed by chloride ion penetration, which accelerates corrosion of embedded reinforcing steel and shortens service life. The Rapid Chloride Permeability Test (RCPT) is a widely adopted method for evaluating concrete’s chloride resistance, involving a 60V DC voltage applied across a specimen with current monitored over six hours to determine total charge passed. However, commercially available RCPT systems with automated data acquisition remain costly, limiting accessibility in resource-limited contexts. This study addresses this gap through a low-cost, microcontroller-based embedded data acquisition system for RCPT applications, integrating voltage and current sensors, an LCD display, and an ESP32-S2 microcontroller for real-time monitoring, control, and data storage. The system logs voltage and current at defined intervals, computes total charge passed, and stores results for post-test analysis, with calibration against digital multimeters ensuring measurement accuracy. Results show the system delivers precise, automated performance comparable to commercial equipment at substantially lower cost, while addressing key challenges including noise reduction, calibration accuracy, and long-term reliability. The proposed system offers a practical, affordable alternative for chloride permeability testing, particularly for laboratories in resource-limited settings.

Keywords—Rapid Chloride Permeability Test, RCPT, Concrete Durability, Chloride Permeability, ESP32-S2, ADS1115, voltage sensor, current sensor, Embedded System, Data Acquisition System (DAS).

Runoff Forecasting and Flood Risk Classification in Sunamganj: A Comparative Analysis of Statistical, Machine Learning, and Deep Learning Models

This study has developed an near real-time runoff forecasting and flood risk classification framework for Sunamganj. Statistical, Machine Learning and Deep Learning models such as Persistence, Linear Regression, Random Forest, XGBoost, LSTM and GRU have been used to predict 1–24 hours of runoff and compare models. Hydrological patterns have been analyzed using time, lag and rainfall-runoff related features. In addition, three flood risk classes—Normal, Moderate and High—have been created based on runoff and the classification performance has been evaluated. This study provides an effective flood early warning approach for flood-prone areas of Bangladesh.

Beyond Point Accuracy: Regime-Conditional SHAP and Walk-Forward Validation for Australian GDP Growth Forecasting

Our contributions are:
• A walk-forward (rolling-origin) evaluation of six forecasting models spanning classical econometrics, ensemble ML, and quantile-regression deep learning, with bootstrap confidence intervals and formal Diebold-Mariano
significance testing between every model pair.

• Aregime-conditional SHAP analysis that quantifies how each model’s reliance on macroeconomic features shifts between crisis and calm periods—an interpretability angle largely absent from existing GDP forecasting litera
ture.

• A learning-curve analysis quantifying the training-data threshold at which each model class stabilizes, directly addressing the small-sample concern rather than treating it as an unexamined limitation.

SyllabusSherlock: A Curriculum-Aware and Knowledge Graph-Enhanced Generative AI Tutoring System for NCTB Education

SyllabusSherlock is a curriculum-aware AI tutoring system built around the NCTB curriculum. At its core, it links a Knowledge Graph of textbook topics and prerequisites with textbook-based retrieval and an estimate of what the learner already knows. Together, these let the system move beyond simple question-answering, offering prerequisite-based learning paths, adaptive assessment, and practice tailored to each student. We also present an initial evaluation of the generated tutoring responses, looking at how correct, textbook-grounded, relevant, clear, and personalized they are.

AgriSemNet: A Semantic-Aware Hybrid Deep Learning and Knowledge Graph Architecture for Precision Crop Recommendation

Most crop recommendation systems either act as uninterpretable black boxes or struggle to process raw sensor data. In this paper, we built AgriSemNet to bridge this gap by combining a tabular MLP with a Graph Attention Network (GAT) and an external 42-node Agronomic Knowledge Graph. Instead of only predicting a single static crop label, our system generates a practical 4-tier recommendation: the primary crop, a low-risk backup crop, a sustainable crop rotation sequence, and clear biological justifications for the farmer. We also built in two-stage soil safety guardrails to flag extreme or toxic soil conditions before planting. On a benchmark 22-crop dataset, our model achieves 98.79% accuracy and stays remarkably stable against IoT sensor noise, dropping only 1.52% under 20% Gaussian noise, where standard models like Random Forest and XGBoost degrade significantly.

Handwritten Prescription Recognition Through Transfer Learning Models Evaluation Under Multiple Dataset Partitions

1. Before the model evaluation, a duplicate-aware dataset verification and leakage-control framework was developed with the help of exact image hashing and perceptual grouping.

2. Four different train-validation-test partition strategies were systematically evaluated to investigate the effect of data partitioning on model performance.

3. The five transfer-learning architectures including VGG16, DenseNet121, ResNet50, EfficientNet-B0 and MobileNetV3-Large are trained on the same training and evaluation settings.

4. To prevent test-set selection of the final model, a validation-based model-selection strategy was used, with Macro-F1 as the validation measure.

5. To test the stability and robustness of the selected configuration, and to avoid visual-group leakage between folds, five-fold stratified cross-validation is employed.

Synthetic Data Augmentation for Bangla Hate Speech Detection Using Slot-Based Semantic Extraction

Novel Framework for Bangla Hate Speech: Developed a robust multi-class classification framework (Hate, Non-Hate, Sarcastic) tailored for implicit hate and sarcasm in low-resource Bangla.

Slot-Based Semantic Extraction: Introduced a structured slot extraction method (TARGET, Negative Expression, Neutral Context, Sarcastic Cue) to capture contextual semantic roles.

Balanced Synthetic Data Augmentation: Built a template-driven, slot-replacement synthetic augmentation technique that expanded a 5,000-comment corpus into a fully balanced 16,302-comment dataset.

State-of-the-Art Performance: Achieved 91.90% accuracy and 91.76% Macro F1-score using BanglaBERT, along with a lightweight TF-IDF+LR baseline reaching 85.00% accuracy, outperforming existing benchmarks.

Three-Layer Forensic Fingerprinting for Insider Leak Attribution: Robust Tracing, Tamper Localization, and Content Binding

To address the challenge of insider leak attribution, this paper presents a three-layer DCT-domain forensic fingerprinting architecture that jointly provides: (i) robust user tracing via QIM-embedded fingerprints with Reed-Solomon error correction; (ii) key-shared tamper detection with block-level localization; and (iii) cryptographic content binding that rejects transplant attempts. All three layers are embedded in a single pass over disjoint DCT bands to prevent interference. Evaluated on 138 benchmark images, the scheme identifies the correct leaker in every benign test, flags every tested manipulation with 0.46-0.88 IoU, completely blocks transplant attempts, and identifies colluders in averaging attacks with zero false positives. Finally, a forced trade-off analysis demonstrates that erasing the fingerprint necessarily destroys the content (dropping quality to at most 25.2 dB PSNR) or triggers a full tamper alarm, proving that no clean, untraceable copy can be laundered.

A Deep Learning Framework for Cross-Medium Deepfake Detection in Independent Images

The rapid progress in deep generative models has led to the creation of hyper-realistic synthetic media that easily evades human perception. While modern networks can generate convincing fake assets, existing forensic tools struggle to generalize because they are optimized either for isolated images or continuous video streams, but rarely both. To address this limitation, we propose a unified, cross-medium deep learning framework designed to evaluate authenticity across independent images. The proposed framework evaluates independent images by utilizing a custom Convolutional Neural Network (CNN) baseline alongside advanced Vision Transformers (ViT) and Hierarchical Mixed-Attention (MaxViT) architectures to capture micro-texture spatial anomalies. The framework was comprehensively evaluated on large-scale datasets under real-world data distributions. In the image-level analysis, the baseline CNN achieved an accuracy of 55.74%, which improved to 89.11% with the ViT backbone, and peaked at 93.40% using the MaxViT architecture. Ultimately, this work provides a highly generalizable and robust solution that significantly advances digital media forensics and multimedia verification systems.

Concept Drift Detection in Streaming Bangla Text Using an Ensemble Statistical Framework

1. We propose an ensemble that combines a BanglaBERT-plus-autoencoder reconstruction-error signal with a label-distribution signal through weighted voting over three statistical detectors.
2. We provide a controlled benchmark against six established detectors (ADWIN, DDM, EDDM, HDDM-W, Page-Hinkley, and KSWIN), reporting F1-Score and false-alarm rate on each drift type.
3. Through a per-detector analysis, we characterise which signals contribute to detection, showing that the label-distribution channel carries most of the discriminative power in our dataset.

Leakage-Aware Ordinal Prequential Forecasting of District-Level Conflict Severity in Bangladesh

Abstract—Reliable district-level conflict forecasting requires
models that distinguish ordered severity while controlling temporal
leakage, geographic heterogeneity, class imbalance, reporting
processes, and non-stationarity. We propose a leakage-aware
prequential framework for forecasting next-month Low, Medium,
and High conflict severity across all 64 districts of Bangladesh.
The study integrates three data sources: (1) ACLED, comprising
31,613 reconciled conflict-event records; (2) NASA POWER,
providing precipitation, temperature, and wind variables; and (3)
geoBoundaries ADM2, providing district-level spatial information.
Their integration produces a complete 11,904-observation districtmonth panel spanning 186 months. Severity is defined by a
study-specific analytical index combining next-month events and
reported fatalities, with the High cutoff (c = 6) frozen from an
initial 60-month history.

Explainable Federated Early-Warning for Commercial Banks: Privacy-Preserving Financial-Distress Prediction and Decision Support in Bangladesh

Bank supervisors need early-warning models that learn from multiple institutions without centralizing confidential bank records. This paper evaluates an explainable cross-silo federated learning framework for one-year-ahead bank-distress prediction in Bangladesh. The compiled panel contains 780 bankyear records for 57 banks over 1997–2025; after forward target construction, 614 observations remain for 2013-2024, with 452 training and 162 strictly out-of-time test observations. Distress
is defined as a next-year breach of either a 10% NPL or 10% CAR threshold. Seventeen CAMELS-derived predictors are winsorized and standardized, and five unbalanced silos train classweighted logistic models aggregated by sample-weighted FedAvg. Centralized Logistic Regression, Random Forest, XGBoost, and Explainable Boosting Machine provide reference baselines, while FedProx, clustered personalization, and differentially private FedAvg test robustness and privacy. FedAvg reaches ROC-AUC 0.9685 versus 0.9714 centrally, retaining 99.7% of reference discrimination without moving raw records. At an RDP-accounted operating point of ϵ ≈ 3.52, δ = 10−5
, DP-FedAvg reaches ROCAUC 0.926. SHAP identifies NPL ratio, ROE, log Z-score, ROA,
and loan-to-assets as leading drivers; independent CAMELS ablation confirms asset quality, earnings, and capital as the most consequential groups. A two-year horizon loses substantial threshold-dependent performance, and a supervisory watchlist
translates probabilities into risk bands and drivers. The results support privacy-aware collaborative bank-risk screening while clearly distinguishing retrospective federation emulation from live deployment.

ThyroNet-XAI: An Explainable Multimodal Framework for Thyroid Cancer Analysis Using Clinical and Ultrasound Data

Thyroid cancer diagnosis requires accurate assessment of clinical characteristics and ultrasound imaging, yet existing artificial intelligence (AI) approaches often rely on a single data modality and provide limited interpretability. This paper presents ThyroNet-XAI, an explainable multimodal framework that integrates clinical tabular data and thyroid ultrasound images for automated thyroid cancer diagnosis and nodule analysis. The proposed framework employs SMOTE-ENN to address severe class imbalance in clinical data, followed by ANOVA-based feature selection to identify informative clinical attributes. A hybrid Random Forest–Deep Neural Network (RF-DNN) architecture is used for clinical classification, while an FCN–TransUNet architecture combines convolutional feature extraction with transformer-based global contextual learning for ultrasound nodule segmentation. To improve transparency, SHAP and LIME are incorporated for clinical predictions, while Grad-CAM provides visual explanations for image-based decisions. The framework is evaluated using the Kaggle thyroid clinical dataset, TN3K ultrasound dataset, and DDTI dataset for external validation. Experimental results reported in the study demonstrate 98% classification accuracy for the clinical branch, with the proposed hybrid clinical model achieving an F1-score of 0.7179 and ROC-AUC of 0.9803. For ultrasound analysis, the proposed segmentation framework achieves a Dice coefficient of 0.8974 and IoU of 0.8138, while multimodal imaging evaluation reports an AUC of 0.9942. These findings demonstrate the potential of combining multimodal learning, hybrid architectures, imbalance-aware preprocessing, and explainable AI to develop accurate and clinically interpretable thyroid cancer decision-support systems.

ECHO-LOCAL: Edge Context Orchestration for Privacy-Preserving Smart-Home Automation

ECHO-LOCAL introduces a privacy-preserving edge architecture that moves the entire smart-home presence estimation and actuation control loop onto a commodity OpenWrt router. By compressing a POMDP presence-control problem into a 5-state FSM estimator paired with a 100-byte 5×5 tabular Q-learning policy, it eliminates cloud round-trip dependencies and metadata leakage. The system achieves a fast 2.2 ms control loop and a compact 697 KB resident set size on live router hardware while enforcing zero WAN telemetry at the kernel level using nftables drop rules.

A Low-Cost ESP32 TinyML System for On-Device Arrhythmia Screening with Train-Deploy Preprocessing Parity

This paper presents an ultra-low-cost, privacy-preserving edge AI system for continuous arrhythmia screening deployed on an ESP32. The primary contribution is achieving train-deploy preprocessing parity by recreating the exact MIT-BIH dataset format on-device using a custom firmware pipeline and Pan-Tompkins R-peak detection. Additionally, the system features a hardware-specific filter chain optimized for local 50 Hz power grid interference and performs full on-device classification, ensuring that only scalar metrics—and no raw ECG waveforms—are transmitted to the cloud.

Spatial Image Steganalysis Using Fixed SRM Preprocessing and SRNet

1. A fixed 30-filter SRM preprocessing stage integrated into an SRNet-style backbone, replacing the learned front-end with handcrafted spatial residual filters.

2. Both configurations (BOSSBase alone and a combined BOSSBase+BOWS2 corpus ) are evaluated against SRNet and ResNet baselines across accuracy, AUC, precision, recall, and F1-score.

3. A paired augmentation pipeline that applies identical spatial transformations to corresponding cover-stego images, preserving their structural correspondence during
training.

4. An ablation study using an SRM+SRNet configuration to quantify the contribution of fixed SRM preprocessing relative to the baseline SRNet architecture.

5. Grad-CAM visualizations identify which image regions the model weighted most heavily when making classification decisions.

Toward Trustworthy Stroke Risk Prediction: A Balanced, Explainable and Externally Validated Machine Learning Framework

This study presents an interpretable and trustworthy stroke risk prediction framework that integrates Boruta–mRMR feature selection, training-only Borderline-SMOTE, comparative evaluation of nine classifiers, and an optimized CatBoost model with an OOF F2-based threshold. The framework further combines SHAP, LIME, and counterfactual explanations with fairness, calibration, robustness, bootstrap-based reliability, and cross-dataset external validation.

A Multi-Strategy Machine Learning Framework for Real-Estate Price Prediction: Pooled, Contextual, and Segmented Approaches

The primary contributions of this research are outlined below:
Empirical Comparison: An empirical comparison of three modeling strategies—a pooled baseline (Model A), a pooled contextual model (Model B), and dedicated per-segment models (Model C)—on 30,472 Bangladeshi property listings spanning four market segments, using five regression algorithms under a shared preprocessing and five-fold cross-validation protocol.
Statistical Testing: Paired, fold-level statistical testing with Holm–Bonferroni correction, confirming that the observed strategy differences are statistically supported.
SHAP-Based Interpretation: A SHAP-based model interpretation showing that the added market-context features are informative and that price determinants differ across the four segments.
Strategy Preference Evidence: Evidence on when each strategy is preferable: a pooled contextual model is strongest overall and on the sale segments, whereas dedicated segment models perform best on the rental segments.

EAM-XAI: Auditing Malware-Identifier Overlap with a Hybrid Behavioral–Retrieval Framework

Repeated memory snapshots carrying the same
dataset-provided malware identifier can cross record-level splits.
A high score may then measure recurring-identifier triage
rather than behavioral generalization to an unseen identifier.
We audit this dependence in CIC-MalMem-2022 and present
Evidence-Aware Malware XAI (EAM-XAI), a routed behavioral–
retrieval system: unseen or unavailable identifiers use a behavioral
LightGBM model, known pure identifiers use label-count retrieval,
and mixed-label identifiers use a local Extra-Trees specialist. The
filename-derived token is not treated as a verified executable hash.
Snapshot-stratified five-fold evaluation reaches 99.85% ± 0.06%
accuracy, whereas malware-identifier-disjoint evaluation reaches
85.49% ± 1.06%. A retrospective multi-snapshot extension that
requires joint access to an unseen identifier’s available captures improves a reproduced single-snapshot baseline from 85.50%±0.98%
to 89.93% ± 1.30%.

A Source-Grounded Multimodal Edge-Agent Framework for Inclusive Digital Service Access in Bangladesh

This research proposes a source-grounded multimodal edge-agent framework for improving access to digital services in Bangladesh. Its main contribution is the integration of Bangla/Banglish interaction, multimodal input, local edge inference, domain-specific routing, source-grounded retrieval, and risk-aware response selection within a single architecture. The work also demonstrates the feasibility of local execution through 120 software trials, achieving 92.5% overall success and 95.0% offline-continuity performance, while clearly distinguishing tested components from proposed future extensions.

Real-Time Detection of Toxic and Hate Speech in Esports Live Chat: A Comparative Study of Machine Learning and Deep Learning Approaches

The significant research contribution is the creation of the first multi-game annotated esports live-chat dataset for toxicity detection, containing 4,957 consensus-labeled messages from five major games. The study also provides a systematic comparison of eight ML, DL, and Transformer models, showing that DistilBERT achieved the best F1-score (0.8151), while Logistic Regression remained a competitive, much faster option for real-time moderation.

Beyond Balanced Benchmarks: A Deployment-Aware Evaluation of Prompt-Injection Detectors under Source Shift

This study evaluates three prompt-injection detectors under a frozen-threshold, deployment-aware setting.
It measures robustness across source shift, hard-benign prompts, and successful JailbreakBench artifacts without threshold retuning.
The results show substantial variation in recall, false-positive behavior, and cross-threat transfer across detectors.
The study also highlights deployment trade-offs through prevalence-aware precision, latency, and decision-margin analysis.

LEGCID: Lightweight Edge-Generated Classifier for Intrusion Detection on OpenWrt

Proposes LEGCID, an edge-native intrusion detection architecture deployed directly on commodity OpenWrt home gateways. The system trains a sparsified 30-feature random forest on IoT botnet traffic, compiles it into dependency-free C99 source via emlearn, and executes in-process flow classification directly from conntrack data with kernel-level nftables quarantine. It operates within strict embedded budgets—achieving 2.9 µs per-inference latency and a 772 KB resident set size while eliminating cloud telemetry egress and recurring compute costs.

A Non-Destructive IoT-Based Multi-Sensor System for Fruit Preservative and Surface Coating Detection

• Designing and developing a portable fruit sensing chamber
for controlled data acquisition.
• Design and integration of a multi-sensor platform, using
optical and gas sensing technologies.
• Co-ordinated data acquisition software and real-time
visualization.
• Production of a multi-modal dataset with synchronized
data from all sensors in a structured form.
• Development of a dataset capable of being used for future
fruit quality assessment systems using AI.

An End-to-End Deep Learning Framework for Animal Classification with Ensemble CNNs and Web Deployment

Animal-image classification is often evaluated as an
isolated model task, while the software path required to turn
a trained model into a usable system receives less attention.
This paper presents an applied deep-learning and softwareintegration workflow for closed-set classification of 30 animal
categories from a custom project corpus of 3,000 images.
The workflow combines image standardization, duplicate
screening before partitioning, a nominal class-wise 70/15/15
split, training-only augmentation, transfer learning, fixed classindex alignment, heterogeneous CNN fusion, and a browseraccessible Flask inference service. The deployed ensemble combines MobileNetV3Small, EfficientNetV2B0, and DenseNet121
through equal-weight averaging of aligned softmax outputs.
On a 448-image development validation set, the observed
development result was 98.44% accuracy with macro precision,
recall, and F1 of 99%, 98%, and 98%, respectively.

ScopeGuard: Mitigating Goal Drift and Confused Deputy Attacks in MCP Agent Pipelines

This paper presents ScopeGuard, a middleware-based security architecture for Model Context Protocol (MCP) agent pipelines that jointly addresses goal drift and confused-deputy attacks. ScopeGuard mediates planner-generated tasks before execution using four complementary checks: semantic intent alignment, provenance verification, least-privilege permission validation, and prompt-injection detection. The prototype is evaluated on 831 labeled tasks, including real goal-drift trajectories and benchmark injection and confused-deputy attacks, demonstrating 80.6% detection of real goal drift and 100% detection of the evaluated injection and confused-deputy cases.

MindBridge: A Multi-Modal AI Platform for Depression and Anxiety Risk Stratification

1. Developed an integrated multi-modal AI platform combining self-report instruments, ClinicalBERT free-text triage, facial affect classification, and EHR vitals with an XGBoost risk ensemble and SHAP attribution.
2. Implemented a 30-day Temporal Fusion Transformer (TFT) to accurately forecast PHQ-9 trajectories and symptom progression.
3. Designed a retrieval-augmented CBT chatbot featuring explicit safety refusal triggers and crisis-line escalation pathways (0 harmful responses on a 200-prompt benchmark).
4. Validated on a retrospective cohort of 4,820 patients, achieving state-of-the-art performance (AUROC 0.892, AUPRC 0.741, F1 0.842) with comprehensive demographic bias analysis and HIPAA-compliant architecture.

X-CRSAA: An Explainable, Clinically-Regularized Self-Attention Autoencoder for Unsupervised Thyroid Anomaly Detection

The core contributions of this work are as follows:
1) A Feature-Tokenized Multi-Head Self-Attention (FTMHA) encoder paired with dual-decoder reconstruction
heads, optimized via a composite loss enforcing biological TSH/FTI constraints (Sections IV and V).
2) A Peaks-Over-Threshold EVT threshold calibration
pipeline that replaces heuristic percentile cut-offs with
extreme-quantile estimation (Section VI).
3) A dependency-free Kernel-SHAP explainability module
providing global and local clinical feature attributions
validated against synthetic ground truth (Section VIII).
4) An extensive empirical ablation demonstrating that
reconstruction-only attention optimization outperforms
masked-feature pretraining and latent-density fusion on
tabular medical data (Section XII).

Sequential Modeling of Corporate Bankruptcy Risk Using Hidden Markov Model Augmented Transformers.

This study contributes a hybrid framework for bankruptcy risk prediction that integrates Hidden Markov Models to capture latent temporal financial dynamics with Transformer-based models for multivariate financial time-series classification. The study also provides a systematic comparison with classical machine learning baselines, demonstrating the potential of temporal representations for identifying bankruptcy risk under severe class imbalance.

Rice Yield Prediction and Crop Health Assessment in Bangladesh Using Satellite Imagery and Machine Learning

This study proposes an integrated machine learning framework for rice yield prediction and crop health assessment in Bangladesh using satellite imagery and environmental data. The study combines NDVI derived from Sentinel-2 satellite imagery with rainfall and temperature data to characterize rice-growing conditions. Machine learning models are employed to predict rice yield and assess crop health across different rice seasons in Mymensingh and Khulna districts. The proposed approach demonstrates the potential of integrating remote sensing and machine learning for data-driven rice monitoring and yield estimation in Bangladesh.

Explainable Calibrated Hierarchical XLM-R for Bilingual PHQ-9 Depression Severity Classification

The study proposes a calibration-aware hierarchical XLM-R framework for five-level PHQ-9-guided depression severity classification in Bangla and English. Its key contributions are hierarchical modeling of three open-ended responses with ordinal-aware learning, bilingual performance analysis, improved confidence reliability through temperature scaling, and interpretable prediction analysis using hierarchical attention and LIME.

Latent-Space Counterfactual Explanations for Cervical Cytology Classification

Deep convolutional networks now classify cervical
(Pap smear) cytology with high accuracy, yet clinical adoption
is held back not by accuracy but by opacity: saliency methods
such as Grad-CAM and SHAP show where a network looked,
but never what would need to change for the decision to differ.
Treating accuracy as a solved sub-problem, this paper contributes
a counterfactual explanation module for Bethesda-system cervical
cytology classification, presented as a pilot feasibility study. We
first select a backbone in a data-driven manner from four
CNNs (ResNet-50, EfficientNet-B3, DenseNet-121, MobileNetV2)
and fuse the top two, DenseNet-121 and EfficientNet-B3, via a
softmax-gated adaptive cross-scale fusion module with a Convolutional Block Attention Module (CBAM), reaching 95.45%±1.37%
test accuracy and 0.9506±0.0135 macro-F1. This classifier is then
frozen and used purely as a differentiable oracle for our main
contribution: a convolutional variational autoencoder coupled
with a latent-space optimization procedure that searches for a
constrained, class-flipping, background-preserving perturbation
of a cytology image toward a target class. Across four diagnostically important Bethesda transitions (n = 20), the module
achieves a 100% validity rate, mean SSIM of 0.549, mean
LPIPS of 0.625, mean sparsity of 0.403, and mean target-class
confidence of 0.985, while keeping background disturbance low
(|D| = 0.082), indicating that most of the edit is confined to
the cell-bearing foreground. The resulting difference maps are
qualitatively consistent with, yet informationally complementary
to, Grad-CAM++ and SHAP saliency.

An Edge-AI Driven Multi-Layered Air Defense Framework for Real-Time Cost Asymmetric Threat Prioritization

The primary contributions of this paper include developing a novel Cost-Asymmetric Threat Prioritization (CATP) algorithm that merges kinematics, trajectory alignment, and CNN-derived visual features (payload severity and destructibility) to resolve multi-target tie-breaking. Additionally, the framework introduces a Cost-Asymmetric Ratio (CAR) that achieves a 32.0% reduction in interceptor costs via a tri-layered response strategy, all while being fully validated on resource-constrained edge hardware achieving an end-to-end latency of 145.33 ms using only 3.5% CPU and 14.3% RAM.

Wind Turbine Blade Structural Health Monitoring Using Piezoelectric Sensors and Machine Learning

This paper presents a machine learning framework
for classifying fatigue-induced damage stages in a wind turbine
blade using piezoelectric (PZT) guided-wave measurements. A
cycle-aware, group-based data-splitting strategy is employed to
prevent measurements from the same fatigue cycle from
appearing in both training and test partitions. Four classifiers –
Random Forest, XGBoost, Logistic Regression, and Support
Vector Machine are trained using 44 time- and frequencydomain features extracted from four PZT receiver channels at
six excitation frequencies (20–120 kHz). The dataset comprises
10,968 feature vectors spanning 1,828 unique fatigue-cycle
instances from a Sonkyo Windspot 3.5 kW composite blade
subjected to cyclic loading. Four structural states are
considered: Pre-crack, Crack-3mm, Crack-10mm, and Postoverload. On the held-out test set of 2,196 feature vectors,
Random Forest and XGBoost achieve accuracies of 99.64% and
99.68%, respectively, with AUROC values of 0.9996 for both
models. Logistic Regression and Support Vector Machine
achieve accuracies of 98.00% and 96.68%, respectively.
TreeSHAP analysis identifies dispersion- and energy-related
features from two PZT channels as consistently influential
across the tree-based models. A label-permutation control
reduces accuracy to 34.6% when training labels are shuffled,
supporting the conclusion that the observed classification
performance depends on the original damage-label structure
rather than the partitioning procedure alone. The results
demonstrate the feasibility of machine-learning-based fatiguestage classification on a single benchmark specimen, while crossspecimen generalization remains an open question.

Efficient Wound Segmentation Using Lightweight U-Net with Adaptive Skip Fusion and Wound-Specific Attention Mechanisms

The main contributions of this work are summarized
as follows:
• Adaptive Skip Fusion (ASF), Wound-Specific Attention
(WSA), and Progressive Refinement (PR) are all inte-
grated into a single U-Net architecture in this lightweight
wound segmentation system.
• In order to retain computing efficiency while improving
the depiction of heterogeneous wound tissues and uneven
boundaries, a wound-oriented attention mechanism is
presented.
• With a Dice coefficient of 0.7584 and an improvement
in segmentation performance of 4.61% over the baseline
lightweight U-Net, the efficacy of the suggested frame-
work is demonstrated on a public wound segmentation
dataset.

A Comprehensive Review of Charging Methods for Electric Vehicle Batteries CC, CV, CCCV, MSCC, and Advanced Strategies

The paper highlighted a comprehensive analysis of the charging methods used in both conventional and advanced EV Li-ion battery charging processes. These are judged based on five critical parameters, including charging time, efficiency, temperature increase, cycle life, and ease of implementation. The optimized five-stage MSCC is a good technique for balancing fast charging technology and implementation, in addition to the adaptive, MPC, AI, and temperature-based methods.

An Educational RTL-to-GDSII Flow for a Non-Pipelined 8-bit Accumulator-Based CPU Using Open-Source Tools

Designed and implemented a complete open-source RTL-to-GDSII flow for an 8-bit accumulator-based CPU, verifying six sub-blocks individually before top-level integration, achieving zero DRC/LVS errors and clean timing closure using Yosys, OpenLane 2, and the SKY130 PDK.

Calibration-Safe Heterogeneous-Rank Federated LoRA: A Multi-Backbone Equivalence Study with a Lossless SVD-Merge Aggregator

We have three main contributions to this paper. First, we have analyzed not only the accuracy but also the calibration in the case of heterogeneous-rank federated LoRA aggregation for the first time. Second, we have adapted the FlexLoRA materialize–average–recompress recipe into SVD-Merge, a lossless aggregator that we have used as a faithful-aggregation baseline for the calibration study. Third, we have presented a statistically robust equivalence result.

A Proposed Framework of Transfer Learning-based Explainable Hybrid Model for Early Cervical Cancer Detection in Bangladeshi Women

The study proposes a transfer learning-based explainable hybrid framework for early cervical cancer detection for women in Bangladesh. The study is ongoing and plans to combine Bangladeshi Pap-smear data, CNN-based image features, and clinical risk factors. Various transfer learning models such as VGG, ResNet, InceptionV3, DenseNet and EfficientNet will be evaluated and SVM, Random Forest and XGBoost will be considered for classification.
In addition, it is planned to explain the decision of the model using Grad-CAM, SHAP and LIME and apply focal loss, SMOTE/ADASYN and GAN-based techniques to deal with class imbalance . Overall, the study aims to combine transfer learning, hybrid CNN-ML, multimodal XAI and Bangladeshi local data to create an interpretable and effective cervical cancer screening framework.

AF-KT: An Alternate Autoregressive and Forgetting-Aware Transformer for Knowledge Tracing in Adaptive Examination Preparation

This research proposes AF-KT, a transformer-based knowledge tracing model that explicitly models human memory. Its key contributions are a forgetting-aware attention mechanism with learnable per-head time-decay rates, a Bayesian mastery gate that tracks per-concept knowledge estimates, and a calibrated alternate-sequence architecture. On the EdNet-KT1 benchmark (5,000 students), AF-KT achieves an AUC of 0.8608 and accuracy of 0.8054, outperforming recent state-of-the-art baselines and providing a practical foundation for adaptive examination preparation systems.

Evaluating Machine Learning Models for Heart Disease Classification Using a Hybrid Ensemble Framework

Our contributions are as follows.:
(i) an evidence-driven stacking framework built
from the three strongest of eight benchmarked
classifiers;
(ii) a joint evaluation of predictive performance and
predictive uncertainty; and
(iii) agreeing global and local explanations from
permutation importance, SHAP, and LIME.

Smart Auto Rickshaw Violation Detection System: A Convergence of IoT, GPS Tracking, RFID Authentication and Cloud Database

The main contributions of this work are:
• An inexpensive, energy-efficient system that uses RFID
tags to identify cars and records infractions for processing
by edge-node microcontrollers.
• Ray casting and Kalman filtering are integrated to provide
precise, lightweight geofencing that can manage several
zones and differentiate between restricted roadways and
approved regions.
• Effectively multitask data streams with separate database
nodes and a single dashboard, cutting latency by separating local edge processing from cloud enforcement.

TrueColorPro: A Handy Device for Color Blind People with IoT Devices

1. Low-cost and portable color detection system based on
the IoT. 2. More accurate when using lighting normalization
techniques. 3. Two output displays (LCD and mobile app)
to improve the usability. 4. A generalizable architecture of
assistive enhancements in the future like speech feedback.

A Multi-Factor Smart Motorcycle Ignition Interlock with Face Recognition, Helmet Verification, and Alcohol Screening

The significant contribution of this research is the development of a low-cost, integrated motorcycle ignition interlock that combines rider authentication, helmet-use verification, and alcohol screening within a single fail-safe decision framework. Unlike prior systems that typically address these functions separately, the proposed prototype permits ignition only when all three conditions are satisfied, thereby integrating theft prevention and rider safety at the point of vehicle startup. The system further demonstrates practical feasibility through embedded, cloud-independent face recognition, HC-12-based helmet-to-bike communication, controlled prototype validation, and a cost-conscious design suitable for resource-constrained settings.

Assessment of Urban Evacuation Demand Based on Population Density under Seismic Events

Rapid and unplanned urbanization in different regions of Bangladesh has increased the need for evidence-based evacuation planning during natural hazards. Evacuation demand refers to the number of people who need to evacuate from an unsafe structure. However, this is not evenly distributed across urban populations. Conventional evacuation planning in Bangladeshi cities has often treated the population as a uniform mass. This paper assesses urban evacuation demand based on population density and demonstrates its application to the Rajshahi City Corporation (RCC). Rajshahi is a rapidly growing city in northwestern Bangladesh, divided into 30 administrative wards. To assess evacuation needs, ward-level population density is calculated using census population counts and the areas of the wards. The results show a clear pattern from the center to the periphery. The findings indicate that the demand for evacuation is not evenly distributed; instead, there is a higher concentration of population pressure in the central urban area. This study identifies the hotspots that represent potential evacuation demand and offers a spatial baseline for emergency preparedness and helps prioritize resource allocation in the RCC.

Firm-Specific Determinants of Environmental Information Disclosure: Evidence from DSE

The study has implications for regulators, corporate managers, investors, and other stakeholders. Regulators may consider strengthening environmental reporting requirements and developing clearer disclosure guidance for listed firms. Firms should move beyond minimal narrative disclosures toward systematic reporting of environmental policies, waste management, environmental investment, certification, and measurable environmental performance. Investors and other stakeholders can also use environmental information as an additional dimension of corporate transparency and accountability.

Leakage-Free, Calibrated, and Explainable Machine Learning for Heart-Disease Screening from Large-Scale BRFSS Survey Data

Our contributions are as follows.
• A leakage-free benchmark. Under honest resampling,
the task is genuinely hard (test ROC–AUC ≈ 0.81), and
stacking does not beat a well-regularized linear baseline.
• A screening-oriented evaluation. We report ROC–AUC
with sensitivity, specificity, and MCC, add isotonic cal-
ibration, and make the sensitivity/specificity trade-off
explicit through threshold analysis.
• Trust and transparency. We audit fairness across sex,
race, and age, show that age-stratified thresholds reduce
the age-related inequity, quantify uncertainty, and ex-
plain the model with permutation importance, SHAP, and
LIME.

Minimum Detectable Improvement: Calibrating the Noise Floor of Microarray Cancer Benchmarks

Contributions.
1) We define MDI@95, a per-dataset, per-protocol resolu-
tion threshold with a direct operational reading, with a
bootstrap interval.
2) We report the floor for five widely used cohorts under
three fold counts and four learners, and release the table
so that future work can situate its margins.
3) We show the floor is set overwhelmingly by partition
choice rather than model stochasticity, that it falls with
both cohort size and fold count, and that the 3-fold
CuMiDa reference protocol sits at the worst end of that
range.
4) We demonstrate the consequence: 61 % of pairwise
comparisons among four standard learners have margins
below their own floor, and on one cohort the identity of
the winning method changes with the seed.
5) We identify a saturated cohort on which every learner
is perfect on every seed, and which therefore has no
capacity to discriminate methods.

GreenBangla: A Grid-Aware Framework for Measuring Compression Energy in Bangla NLP

The principal contributions are as follows:
• We have measured training and inference energy for
full fine-tuning, LoRA, QLoRA, and INT8 post-training
quantisation on a Bangla NLP task while controlling
precision, idle power, sampling duration, and accelerator
visibility.
• We have demonstrated that 4-bit QLoRA can consume
more energy than LoRA on a Tesla T4 despite its lower
memory footprint, while a smaller task-specific Bangla
encoder can achieve comparable classification accuracy
at substantially lower energy cost.
• We have quantified the sensitivity of reported carbon
to Bangladesh grid-emission assumptions and derived a
practical reporting checklist for reproducible Green-AI
studies in resource-constrained settings.

A Deep Learning Approach for Dialect Neutralization in Bengali Academic Writing

Bengali academic writing frequently reflects regional dialectal influences, particularly among students from diverse linguistic regions of Bangladesh. These influences are evident as non-standard vocabulary, spelling forms, and grammatical structures, which make it less clear, formal, and acceptable to automated assessment. This study presents an AI writing assistant based on deep learning that can identify and neutralize regional dialects in Bengali academic writing. The proposed model follows a two-stage architecture that includes dialect classification with transformer-based models and dialect neutralization with sequence-to-sequence text generation models. A custom parallel corpus of dialect-influenced and standardized academic Bengali sentences is constructed through curated sources, synthetic dialect injection, and optional crowdsourced annotation. Experimental evaluation demonstrates promising performance, where the dialect classification stage achieved an accuracy of 0.77 and an F1-Macro score of 0.77 across five regional dialect categories. In the dialect neutralization stage, the system achieved an average Word Error Rate (WER) of 0.3970, a Character Error Rate (CER) of 0.2490, and a BLEU score of 34.52. These results represent a notable reduction in WER relative to single-stage baseline metrics reported in prior dialect-translation literature; because that comparison spans a different translation direction, dataset, and dialect count, it is best read as indicative rather than a controlled ablation, and a matched same-dataset ablation is identified as a priority for follow-up validation. Ultimately, this work illustrates the potential of dialect-conscious normalization for improving the quality of academic texts in Bengali and contributes new data, methods, and empirical evidence to the growing body of Bangla Natural Language Processing research.

Multi-Scale CBAM Fusion Network: A Lightweight and Explainable Deep Learning Approach for Acute Lymphoblastic Leukemia Detection from Blood Smear Images.

This work proposes MSCBAMFN, a lightweight (0.91M-parameter) network that matches or outperforms four much larger ImageNet-pretrained backbones (ResNet50, EfficientNetB0/B3, DenseNet121) with 4–27× more parameters, achieving 99.59% accuracy on the original dataset and 96.94% on the segmented dataset. The model also shows strong clinical reliability, missing only 2 of 490 malignant cases on the original dataset, which matters most in screening where a missed diagnosis is far costlier than a false alarm. Grad-CAM applied at the fusion layer confirms the model’s explainability by consistently highlighting the nucleus and cytoplasmic boundary as the regions driving its predictions, rather than background artifacts. Since the segmented dataset is a background-removed, cell-focused version of the same original images, the model’s consistent performance across both variants demonstrates strong generalization based on genuine cell morphology rather than incidental background cues. This is further validated by a component-wise ablation study confirming the contribution of CBAM and multi-scale fusion, along with 5-fold stratified cross-validation (99.11%±0.23% and 97.02%±0.86%) showing the results are stable and reproducible rather than an artifact of a single data split.

A Deep learning approach for anemia detection using peripheral blood smear images

Anemia remains one of the most prevalent public health conditions globally, and accurately identifying its morphological subtype is crucial for appropriate clinical management. Conventional diagnosis involves the microscopic assessment of peripheral blood smears by a trained hematologist or pathologist to classify red blood cells (RBCs) based on cellular size and shape, a process that is time-consuming, subjective, and constrained by the limited availability of expert personnel, particularly in low-resource healthcare environment. In response, this paper introduces a deep learning-based framework for automated classification of RBC morphology directly from peripheral blood smear microscopy images into three clinically relevant classes: Normocyte, Macrocyte, and Microcyte. We construct and preprocess a dataset containing 2,500 RBC images and apply preprocessing class-balancing augmentation to address dataset imbalance. Three transfer-learning-based architectures. ResNet50, DenseNet121, and a Vision Transformer (ViT-Base), are individually fine-tuned and evaluated, achieving individual test accuracies of 89.87%, 88.53%, and 89.07% respectively, with mean AUC scores above 0.96 in all cases. We further propose a performance-weighted hybrid ensemble that fuses the softmax outputs of all three models, improving test accuracy to 91.20% and mean AUC to 0.9792, outperforming every individual architecture. These results illustrate that combining convolution-based and attention-based feature representations improves robustness in fine-grained RBC morphology classification, and suggest a viable pathway toward low-cost, automated anemia subtyping to support diagnosis in settings where expert microscopy review is limited.

Heart Disease Prediction using ECG Image Analysis with Grad-CAM Explainability Visualization

The following are the major contributions of this research:
1. An end-to-end ECG preprocessing and region-of-interest (ROI) based segmentation system that helps to extract diagnostically useful waveform regions before model training.
2. Comparative performance analysis of different DL architectures (VGG16, ResNet50, DenseNet121, MobileNetV2, EfficientNetB0, EfficientNetB0 with CBAM Attention mechanism, and hybrid Fusion) in the same experimental setting.
3. Use of Grad-CAM for interpretation of the waveforms responsible for each prediction and thereby making decision-making clinically reliable.
4. Real-time prediction system using a GUI interface that enables clinicians to upload ECG images and get interpretable predictions.

GAN-Driven Synthetic Financial Data Generation: Ensuring Privacy and Statistical Fidelity for Fraud Detection and AI Applications

1.A CTGAN-based framework for privacy-preserving
synthetic financial transaction data;
2. An in-depth evaluation of
statistical fidelity and privacy protection;
3. Experiments
showing that synthetic data can improve accessibility and privacy while
achieving near-real downstream performance.

Vehicle-to-Grid for Renewable Energy Integration in Bangladesh: A Degradation-Aware Critical Review and Future Research Framework

This paper presents a degradation-aware critical review of Vehicle-to-Grid (V2G) technology for renewable energy integration in Bangladesh. The study systematically synthesizes recent evidence using PRISMA 2020 and SALSA frameworks, identifying key challenges related to battery degradation, EV mobility, renewable variability, distribution-grid constraints, economics, aggregation, and cybersecurity. The paper highlights Bangladesh-specific research gaps and proposes an integrated future research framework that combines renewable generation, battery health, EV availability, grid limitations, and techno-economic considerations for sustainable V2G deployment.

A Transformer-Based Approach to Understand and Identify The Fake News In Bangla Media

The rapid growth in the digital media in Bangladesh has increased the spread of misinformation in online news portals and social networks. Detection of fake news in Bengali is a special problem because of the complexity of the language, informative manifestations and often a combination of the textual presentation of persuasion and false visual information. The proposed paper suggests a transfer learning-based multimodal model that incorporates BanglaBERT to provide text contextual representation and ResNet50 to obtain visual features. The features obtained are pooled in one representation and then subject to fully connected layers to perform binary classification. The model is trained on the AdamW optimizer and is tested on stratified data splits and cross validation. As the experimental results illustrate, the suggested system is 92.81% accurate, which is higher than the text-only and image-only baselines. The results prove that multimodal learning is effective in detecting fake news in Bangali and stress that it can be used in the real-life.

An Explainable Attention Enhanced Model for Breast Cancer Detection

Breast cancer is a major worldwide health concern, where delayed or erroneous diagnosis critically worsens patient outcomes and increases mortality. Therefore, early and accurate detection of breast cancer is necessery to improve patient outcomes and lower mortality rates. Conventional diagnostic workflows, which depend heavily on manual histopathological and ultrasound image analysis are time consuming, subjective and prone to human error, specially as patient volumes continue to rise. Existing computer aided diagnosis (CAD) systems also have notable limitations like traditional machine learning (ML) classifiers depend on manually crafted features that do not generalize well across different magnifications and staining variations, while standard convolutional architectures treat every spatial position and channel equally, offering minimal insight into their own decision making processes and limiting clinical trust and adoption. To address these challenges, we propose an explainable framework combining DenseNet201 with a Convolutional Block Attention Module (CBAM) and a Transformer encoder for the classification of breast cancer histopathology images into benign and malignant categories. The integrated attention mechanism enables the model to selectively focus on diagnostically relevant spatial and channel-wise features, while Gradient weighted Class Activation Mapping (Grad-CAM) is employed to visualize these regions, enhancing interpretability and clinical trust. Comprehensive experiments demonstrate that our model achieves a high accuracy of 97.22%, significantly outperforming eight independently benchmarked backbone networks as well as reported traditional machine learning baselines, highlighting its potential as a reliable and interpretable tool for clinical decision support.

Evaluating Provenance-Based Defenses Against Prompt Injection in LLM Agents: What Taint Tracking Can and Cannot See

Large language model agents read email, browse the web, run shell commands, and send messages on a user’s behalf, so content an agent merely reads can become an action it performs. Guardrails against this rely on one assumption: an action becomes dangerous because untrusted content preceded it. We test that assumption directly. Using the JED red-team environment, its sandboxed tool suite, 19,679-page web corpus, and breach-scoring function, we evaluate a provenance tainttracking guardrail against four attack classes with full source access to both the defense and the scorer. Three classes are structurally closed. An exact-substring block denies the reads that exfiltration and destructive-write attacks require, and the guardrail’s five-event taint window fully contains the scorer’s two-event breach window, so every scored breach is already denied. Only 221 of the 19,679 corpus pages, 1.1 percent, trigger the injection parser at all, and 194 of those, 88 percent, collapse to one outcome the guardrail already blocks. The fourth class succeeds: a confused-deputy chain built entirely from direct user requests produces no web- or email-sourced tool event, so the taint mechanism has nothing to observe and issues no denial. Taint tracking answers one question, did untrusted input precede this action, and an attack built entirely from trusted input falls outside what it was built to detect.

PhysicsMate: A Curriculum-Grounded Bengali Benchmark for Secondary Physics QA with Small-Model Adaptation

This paper presents PhysicsMate, the first curriculum-grounded Bengali physics QA benchmark and multi-relational knowledge graph (1,760 nodes, 2,600 edges across 10 ontological types, and 1,834 QA pairs) derived from the NCTB Grade 9–10 syllabus. We conduct a controlled LoRA adaptation study across three small language model scales (Qwen3 0.6B, 1.7B, and 4B), demonstrating monotonic capacity-dependent closed-book accuracy gains (+5.5, +15.0, and +23.3 percentage points) and enabling offline 4-bit edge deployment. Finally, an ontological diagnostic analysis reveals that adaptation primarily aids structured curricular categories (physical quantities: +93.9%, named laws: +56.6%) while pinpointing remaining gaps in entity-level reasoning.

A District-Level Regional Planning Framework for CCUS Deployment in Bangladesh: Spatial Decision Support Approach

This study has constructed and applied Regional CCUS Deployment Suitability Index (RCDSI), an integrated multi-criteria spatial tool to identify suitable Carbon Capture, Utilization, and Storage (CCUS) deployment zones across all 64 districts of Bangladesh. The RCDSI integrates storage feasibility, climate risk, and utilization synergy into a single multiplicative index, and is based on data-driven principles and a transparent formula that allows for easy reproduction, thereby ensuring a foundation for climate risk assessment of CCUS locations in developing countries that is both reproducible and transparent.

An Uncertainty-Aware Joint Sizing and Queue-Constrained Scheduling Framework for Solar-Assisted Electric Three-Wheeler Battery-Swapping Stations: A Field-Informed Dhaka Case Study

This paper proposes an uncertainty-aware joint sizing and queue-constrained scheduling framework for solar-assisted electric three-wheeler battery-swapping stations. The study integrates photovoltaic capacity, charger number, grid capacity, and battery inventory optimization with realistic operational constraints, including finite READY battery availability, FCFS queueing, uncertain vehicle arrivals, battery SOC variation, and solar variability. A field-informed Dhaka case study and discrete-event Monte Carlo simulation framework are developed to evaluate reliability, energy performance, and economic trade-offs. The proposed approach provides a practical planning methodology for designing resilient and cost-effective E3W battery-swapping infrastructure in emerging urban mobility systems.

Feature-Based EKF-SLAM for Differential-Drive Mobile Robots: Implementation and Parameter Sensitivity Analysis

Implementation of FEKF-SLAM on a Differential Drive Mobile Robot using a modular approach.

Evaluation of three progressive stages: motion-only localization, SLAM with predefined features, and full SLAM with dynamic feature addition.

Development and integration of observation models and their Jacobians to enable accurate state estimation and feature mapping.

Analysis of the effect of varying key parameters such as sensing range and feature update frequency on localization performance.

Degradation-Aware V2G Coordination for Grid Stress Mitigation in a Calibrated IEEE-33 Bus Distribution Network

The swift expansion of electric vehicles (EVs) and
renewable generation is basically changing distribution networks ,
from boring passive power-delivery setups into flexible , data-
driven smart grids. Still, if EV charging is not coordinated, it can
push up the evening peak demand, make voltage deviations get
worse, overload feeders , and also drive operating costs higher,
mainly in renewable-rich areas. So in this paper we offer a
degradation-aware vehicle-to-grid (V2G) coordination approach
for EV fleets inside a future smart distribution network that is
calibrated for Bangladesh. The idea is to treat an EV fleet as an
aggregated distributed energy resource, and then schedule both
grid-to-vehicle (G2V) and V2G power exchanges using an
aggregator-based control scheme. For the experiments,
simulations are carried out on an IEEE 33-bus radial feeder,
together with a Bangladesh-calibrated daily demand pattern.
Dhaka solar photovoltaic inputs come from NASA POWER, and
the stress case is built as a 2035 future scenario. That scenario
assumes 450 equivalent EVs, 2.40 MW of PV capacity, 4.05
MWh of daily EV energy demand, plus an aggregate charging cap
of 2.60 MW and an aggregate V2G cap of 1.35 MW. Altogether,
six operating cases are tested: no EV, uncontrolled charging,
price-only smart charging, rule-based V2G, degradation-blind
V2G, and the proposed degradation-aware V2G strategy. The
results are pretty clear. With uncontrolled charging the peak grid
import climbs to 7.80 MW and the smallest bus voltage drops to
0.8193 pu. Meanwhile, the proposed strategy cuts the peak grid
import by 42.09% , raises the minimum voltage to 0.8883 pu, and
reduces total daily operating cost by 4.44% , all while keeping
unnecessary battery cycling under check. Overall, these findings
suggest that battery-aware EV fleet coordination can help with
peak-load reduction, better voltage profiles, and real cost savings
in renewable-integrated distribution systems. That said, even with
this, feeder reinforcement and reactive-power support might still
be necessary once electrification gets extremely high.

The “Aakash Deep” Hub: A Sustainable and AI- Driven Li-Fi Framework for the Consumer Market in Bangladesh

The global effort to bridge the digital divide requires wireless solutions that are not only high-performing but also sustainable, secure, and resilient to local infrastructure challenges. This is particularly true in developing nations like Bangladesh, where issues such as power instability (load shedding) and shared bandwidth congestion can severely degrade the user experience. This paper proposes the “Aakash Deep” Hub, a futuristic, consumer-focused Li-Fi framework designed specifically to address these challenges. We move beyond traditional performance metrics to evaluate the system’s viability through a series of high-fidelity Python simulations grounded in the context of a typical Bangladeshi home. The study demonstrates, with quantitative reasoning, how Li-Fi’s inherent properties, when enhanced by AI, create a superior consumer product. Our simulated results show that the system’s high energy efficiency ensures connectivity survives through multi-hour power cuts where traditional Wi-Fi fails. We demonstrate how an AI-driven Quality of Service (QoS) engine prevents buffering and lag in a congested, multi-user family environment. Furthermore, we visualize Li-Fi’s inherent physical-layer security, a critical advantage for user privacy. A final techno-economic analysis reveals that while the initial hardware cost may be higher, the total cost of ownership and the vastly improved reliability and user experience present a compelling value proposition. This research concludes that an AI-enhanced, sustainable Li-Fi framework is a highly consistent and viable solution for the next generation of home networking in Bangladesh and other developing markets.

Medprice-XAI-BD: A Kind of Explainable & Uncertainty Mindful Machine Learning Framework, for Medicine Price Prediction in Bangladesh

Medicine price transparency is, like, an important problem in data-driven healthcare analytics, because medicine affordability does not only depend on whether the product is available, but also on manufacturer, dosage form, generic composition, strength, and even that package-level variation thing. In this paper we propose MedPrice-XAI-BD, a framework that’s both explainable and uncertainty-aware, aimed at predicting medicine prices using structured medicine data from Bangladesh. The dataset has 21,714 medicine records with attributes such as brand name, generic name, dosage form, strength, manufacturer, package container, and package size details. Because the dataset itself does not provide a direct numerical price column, we build a preprocessing pipeline that extracts valid prices in Bangladeshi Taka from package-text fields, and that produces 21,636 usable price records. Then we test several regression models for price prediction, including Random Forest, XGBoost, LightGBM, and CatBoost. Overall results suggest that CatBoost performs best across the board, giving MAE = 172.75 BDT, RMSE = 2689.38 BDT, and R² = 0.368, while still coping well with heterogeneous categorical medicine features. Besides that, we also train a low-medium-high price category classifier, and it reaches 97.30% accuracy and 0.973 macro-F1. This indicates a good separability between the different price ranges. For interpretability, we use SHAP-based explainability to pinpoint what really drives the price, dosage form, package size, package container, generic name, strength, and manufacturer they come out as the major influencers. Finally, we apply split conformal prediction to quantify uncertainty, and we get 90.25% empirical coverage for a nominal 90% prediction interval. Overall, the proposed framework shows how explainable AI plus uncertainty-aware prediction can help medicine market analysis, affordability research, and more transparent pharmaceutical data analytics in Bangladesh.

Ensemble Machine Learning Models for Multi-Target Nanoparticle Cytotoxicity Prediction

Nanoparticle cytotoxicity prediction can help with early safety screening, though a reliable evaluation is kinda hard when datasets have repeated nanoparticle identities, plus the toxicity labels are very skewed. In this work we put forward a leakage-audited machine-learning pipeline that aims to classify nanoparticle cytotoxicity by using physicochemical, exposure, and biological-response descriptors. The curated dataset has 308 records and 22 engineered features across three toxicity categories, yet the rarest category only has two records, and it does not show up in the frozen, held out test split. We then combine Random Forest and LightGBM models using logistic regression stacking. All preprocessing and resampling are done inside the training partitions, so it stays leakage-checked. For a fair like for like comparison, every model is tested with nanoparticle-grouped cross validation, using macro F1, balanced accuracy, and per class metrics. We also include a supplementary analysis based on the frozen split, but we treat the overall numbers as partial, two-class style evidence, not as a complete three-class validation. Feature importance and case level explanations point to how much biological-response, exposure, and physicochemical variables contribute. Overall the results suggest the pipeline can be made auditable for screening, but they also underline that we still need bigger and more balanced datasets, and independent external validation, before anyone should claim broad generalization or regulatory readiness.

A Framework for Generative AI-Designed Synthetic Gene Circuits for Targeted In Vivo Senolytic Therapy

Cellular senescence kind a contributes to age
related tissue dysfunction , but the senolytics right now they don’t really have that context specificity, and also safety can be kind of shaky. Here we show a computational framework for designing and screening senescence responsive synthetic gene circuits, using generative AI plus a risk aware evaluation layer. The whole workflow mixes together several bits, like: (i) a tissue aware senescence expression atlas built from harmonized single cell RNA‑seq datasets ; (ii) a senescence state classifier trained on atlas derived features, with explicit train/validation/test splits, (iii) a constraint aware conditional sequence generator aimed at regulatory DNA, (iv) a logic compiler that turns Boolean specs (AND/NOT) into genetic parts, (v) biophysical simulation of circuit dynamics through chemical reaction network ODEs, and (vi) a safety scoring module that tries to penalize OFF state leakiness, off target activation, and cellular burden. For candidate selection, we also lean on explainability tools—motif enrichment and SHAP—plus distribution free uncertainty quantification using conformal prediction, so risk sensitive picks are not just based on point estimates. We report dataset identifiers, model architectures, training hyperparameters, baseline comparisons, and ablation studies, and we include confidence intervals. All figures are numbered, referenced in text, and paired with statistical summaries. This study is explicitly computational, so no wet lab work or in vivo validation is claimed. Instead, we provide an auditable reproducible design protocol and a ranked set of candidate circuits meant for future pre-clinical evaluation. Overall, the work clarifies what methodological requirements and evaluation standards should look like for generative AI assisted senolytic circuit design.

CleftNet v2: A Dual-Pipeline Deep Learning and Handcrafted-Feature Machine Learning Framework for Pre- and Post-Surgical Classification of Cleft Lip Images

The primary contribution of this paper is CleftNet v2, a dual-pipeline framework that resolves the clinical trade-off between high predictive accuracy and model transparency for pre- and post-surgical cleft lip image classification. The architecture pairs an end-to-end deep learning CNN (combining residual learning, multi-scale inception blocks, and squeeze-and-excitation channel attention) that achieves 93% test accuracy and 0.9776 ROC-AUC with an independent, highly interpretable machine learning ensemble. This second pipeline extracts a 305-dimensional vector of clinically motivated features—such as facial symmetry, quadrant-based tissue ratios, color, and texture—reaching up to 91.79% cross-validated accuracy while providing auxiliary outputs for severity, symmetry, age, and gender. Together, the dual approach delivers both top-tier automated diagnostic performance and transparent, human-readable metrics to support objective surgical auditing.

Machine Unlearning Based Pre and Post Surgery Classification of Cleft Lip and Nasal Images

The core research contribution of this work is the development of a privacy-preserving medical image classification framework that integrates a custom CleftNet v2 deep learning architecture with SISA-based Machine Unlearning and Grad-CAM explainability. Designed for classifying pre-operative and post-operative cleft lip facial images, the framework combines Inception modules, residual connections, and squeeze-and-excitation attention blocks to extract multi-scale spatial features, achieving an overall classification accuracy of 93.36% and an ROC-AUC of 97.76%. Crucially, by incorporating the SISA (Sharded, Isolated, Sliced, and Aggregated) paradigm alongside a complementary 305-dimensional handcrafted feature extraction pipeline, the system enables the selective, rapid removal of specific patient data upon request through isolated shard retraining. This eliminates the need for full model retraining while supporting clinical interpretability and compliance with stringent data privacy regulations like GDPR’s “Right to Be Forgotten”.

ESG and Investment Decisions: The Moderating Role of Risk Tolerance among Individual Investors in Bangladesh

The study contributes to the ESG investment literature by considering risk tolerance as a potential moderator within an emerging-market context. The results from the moderation analysis indicate that willingness to absorb risk cannot be the sole explanation for the influence of ESG on investment decisions.Governance and environmental factors directly support SDG 16 and SDG 13, while social factors indirectly support SDG 8. Ultimately, this research clarifies sustainable investment dynamics in developing markets and provides practical guidance for Bangladesh’s financial sector, policymakers, and corporate leaders seeking to advance responsible investing.

An Integrated Machine Learning and Web Dashboard Framework for Post-Type 2 Diabetes Complications and Recommendations Using Chittagong-Based Survey Data

This study develops an integrated machine learning framework using survey data from 1,002 Type 2 Diabetes patients in Chittagong, Bangladesh, to predict individual insulin-use risk across 30 clinical and lifestyle features. Benchmarking eight classification models showed that a Soft Voting ensemble achieved the highest overall accuracy (96.02%), while a Support Vector Machine (SVM with RBF kernel) was selected for deployment due to its superior precision (95.67%) and 95.45% accuracy. Beyond raw prediction, the framework advances clinical utility by combining SHAP and LIME to generate patient-level explanations—identifying self-monitoring frequency, age, and eye check-up rates as key risk factors—and incorporating a customizable decision engine that lets clinicians tune model weights and risk thresholds live. The entire pipeline is deployed as an interactive web dashboard providing population analytics alongside plain-language, actionable clinical recommendations.

Predictive Ceilings in Generative-AI Learning Analytics: A Six-Model Benchmark of Burnout Risk, Grade Change, and Skill Retention Across 50000 Student Records

This paper demonstrates that standard supervised machine learning models hit a strict predictive ceiling when using tabular self-reported AI usage data to forecast student outcomes across 50,000 records. Across six distinct model families—spanning linear, kernel, ensemble, and neural network architectures—performance converged to a narrow band without statistically significant separation, indicating that predictive bottlenecks stem from limitations in the feature space rather than algorithm capacity. Furthermore, a feature group ablation study revealed that AI usage data alone is highly informative for psychological targets, capturing 94% of full-model performance for burnout risk, but is virtually uninformative on its own for academic outcomes like grade change, where predictive power depends almost entirely on complex interactions between study habits, prior GPA, and AI usage patterns.

Myth, Model and Reality NLI Framework for Hallucination Mitigation in SLM

Mitigating Hallucination in Small Language Models using NLI framework by focusing on cost management (by using SLM)

Time-Aware Explainable Machine Learning for CO2 Emission Analysis in Bangladesh: Decoupling, Decomposition, and 2030 Scenarios

In Bangladesh, the continuing economic and energy-demand growth has been paralleled by the continuing increase in emission intensity across the country’s territory, which raises the question of what territorial CO2 emissions and economic growth mean together, and how these can be understood and anticipated. This study proposes an analytical framework to explain and understand the CO2 and Greenhouse Gas indicators from the Our World in Data database in a timeaware way, covering the description of 1990-2024and completecase modelling over 1990–2022 (n = 33). Gross domestic product, population, and primary energy consumption are used to predict territorial CO2 emissions with Linear Regression, Ridge Regression, Random Forest, and XGBoost, evaluated under both repeated random cross-validation and an expandingwindow temporal-validation protocol that mimics future-year forecasting. Model behavior is interpreted with SHAP and permutation importance and ablation (leave-one-feature-out) methods; historical structural dynamics are explored with Tapio decoupling and Kaya–LMDI decomposition; and 2030 scenarios are projected with Monte-Carlo. The best model is validated by random cross-validation with R2 ≈ 0.996, and by expanding window temporal validation with R2 ≈ 0.974, suggesting that indistribution fit overestimates the out-of-distribution fit for this short annual series. Primary energy consumption turns out to be the most explanatory driver in the SHAP analysis, while the decomposition of the emissions increase by Kaya–LMDI reveals an affluence effect of +71.93 MtCO2. With various growth roadmaps, emissions are estimated to be around 158.8–187.1 MtCO2 in the illustrative 2030 scenarios. The finding suggests that temporal (not only random) validation of the credibility of forecasting is essential, and that the emissions pathway in Bangladesh is still on a trajectory more about economic scale and less about structural decoupling.

When Fine Tuning Does Not Help: A Two Stage EfficientNet B0 Evaluation for Four Class Lung CT Classification on a Small Cohort

This study evaluates standard transfer learning practices on small medical datasets using an EfficientNet-B0 model trained on a 990-image chest CT collection. The authors’ primary contributions demonstrate that partial fine-tuning (unfreezing the top backbone layers) yielded no improvement over frozen feature extraction, reaching its peak validation accuracy (88.89%) prior to fine-tuning. Additionally, the paper highlights a significant divergence between metrics—where a high macro ROC-AUC (0.9466) masked lower test accuracy (76.83%) and poor recall (0.58) for adenocarcinoma—showing that ROC-AUC alone misrepresents practical model quality. Finally, the study illustrates how balanced class weighting merely redistributed errors by over-predicting minority classes, and it documents a critical label space mismatch in the source dataset requiring explicit harmonization.

Design and Implementation of an Intelligent Solar Panel Monitoring and Single-Axis Auto-Tracking System

1)A complete, reproducible, single-axis LDR-based solar tracker built on an Arduino UNO, with a published wiring diagram and pin allocation suitable as a teaching
reference.

2)A discrete-step closed-loop control law with an explicit dead-band threshold and a step-limited update rule,which removes the limit-cycle oscillation around the
equilibrium angle seen in earlier designs.

3)A mechanically mirrored twin-servo actuation scheme,where the second servo is driven by the complementary angle, giving the panel bilateral support and reducing
torque on either individual shaft.

4)An on-device measurement pipeline a 16×2 I2C LCD plus a pipe-delimited serial log that exposes irradiance, terminal voltage, and servo angle as user-readable sig-
nals at 1 Hz.

5)A quantitative bench comparison between the proposed tracker and a co-located fixed-orientation panel, together with cost and feature comparisons against published
alternatives.

A Compact Multi Scale Attention CNN for Brain MRI Tumour Detection Under a Duplicate Aware Evaluation Protocol

The paper’s main contribution is a reproducible, leakage-aware evaluation protocol (content-hash deduplication of a public brain MRI dataset reducing 506 → 228 unique images, followed by stratified grouped 5-fold CV with inner validation for early stopping and single-use test folds) applied to a compact multi-scale channel-spatial attention network (MSCA-Net, 0.34 M parameters).
Under this rigorous protocol it reports realistic performance (~0.82 accuracy / ~0.89 AUC) that is statistically indistinguishable from a stronger ResNet-18 baseline, while Grad-CAM evidence shows some correct predictions rely on non-brain cues. The work therefore supplies a careful reference point and a caution against overly optimistic published numbers on this (and similar) small public collections rather than claiming architectural superiority.

Conservation-Residual Augmented Learning for Telemetry Integrity Monitoring of Rooftop Solar-Storage Systems

Novel monitorning strategy for rooftop solar pv system

ARMeet: A Dual-Mode Mobile System for AR-Anchored and Virtual-Room Multi-User Meetings with Spatial Presentation

This research presents ARMeet, a dual-mode mobile meeting system that integrates AR-anchored and virtual-room environments to support multi-user meetings with spatial presentation. The proposed system enables users to transition between physical AR-based collaboration and virtual meeting spaces within a unified platform, providing an interactive and spatially oriented approach to remote communication and presentation.

NORA: A Multimodal Edge-AI Framework for Offline Context-Aware Intelligent Assistance

The main research contribution is the development of NORA, a local-first multimodal edge-AI assistant that combines speech, visual information, contextual reasoning, local LLM inference, and risk-aware decision making within a single framework. The work also introduces a reliability-weighted multimodal risk-fusion model and evaluates the software prototype for task success, latency, resource usage, and offline operation, showing the practical feasibility of intelligent assistance without continuous cloud dependence.

SOROLIPI: Automatic Bangla Image Captioning Using Deep Learning Based

Natural Language processing

Automatic Detection of Total Sleep Time using Nasal Airflow and Pulse Oximetry Signals

The severity of obstructive sleep apnea (OSA) is based on apnea hypopnea index, which is dependent on total sleep time (TST). TST is measured with the electrode-dependent, costly, and labor-intensive electroencephalography (EEE) recording derived from polysomnography. The TST estimation problem has been largely overcome by simplified screening alternatives, where total recording time (TRT) or a fixed percentage of TRT is used as TST. This study proposed an automatic algorithm to estimate TST from airflow (AF) and oximetry (SpO2) by applying a subject-adaptive sleep-wake classification strategy. The algorithm took AF and SpO2 signals as input, pre-processed AF (smoothing, filtering, peak-excursion) and SpO2 (round-off and correction). After segmenting both signals into 30 second epochs, different features were extracted. A data-driven adaptive threshold was applied to distinguish epochs of sleep from wake, from which TST was estimated. On a held-out testing set (N = 68 subjects), the estimated parameters were compared against annotation. The method resulted in negligible systematic underestimation at the population level (Pearson’s correlation coefficient, r = 0.463). A minimal underestimation, with a mean bias of -0.004 hours and a limits of agreement of −2.147 to +2.139 hours was resulted in between the estimated and annotated TST. The performance of the designed algorithm can’t be compared with EEG-based studies, and there is no such study of estimating TST from AF and SpO2. Hence, the method can be applied to OSA screening device, since the apnea and hypopnea events screening is based on these two signals.

IoT-Enabled Smart Aquarium Monitoring and Automated Cleaning System

IoT based cost effective smart aquarium

DECL: Disease Evidence Consistency Learning for Robust Plant Disease Classification

Plant disease classification models can achieve high
predictive accuracy while remaining sensitive to appearance variations and relying on inconsistent visual cues. We propose Disease
Evidence Consistency Learning (DECL), a transformation-aware
framework that encourages stable disease representations and
spatial evidence across disease-preserving views of the same
image. DECL employs a pretrained ResNet18 backbone with
a learnable disease-evidence module that generates spatial evidence maps and constructs evidence-weighted feature representations. Training jointly optimizes classification loss, feature level consistency, and transformation-aware evidence consistency. Specifically, cosine consistency encourages invariant disease
representations, while spatial evidence maps are geometrically
realigned before consistency regularization, promoting invariance
to photometric changes and equivariance to geometric transformations. This enables the model to learn stable task-relevant
evidence without requiring lesion-level annotations. Experiments
on bean leaf disease classification show that DECL achieves an
accuracy of 99.04%, providing an absolute improvement of 3.85
percentage points over standard training. DECL also produces
well-calibrated predictions, achieving an Expected Calibration
Error (ECE) of 0.75% and a Brier score of 0.0185. These results
demonstrate that jointly enforcing representation and disease evidence consistency can improve both predictive performance
and reliability in plant disease recognition.

SOROLIPI: Automatic Bangla Image Captioning Using Deep Learning Models

Bangla Image Captioning

Machine Learning Based Real Estate Price Prediction

Based on machine learning, the cost of any home or apartment can be estimated. We can accomplish this utilizing test and training data derived from real-time data. We categorize and identify distinguishing characteristics appropriate for algorithms. We tested many algorithms, Decision Tree, Linear Regression, Random Forest Regression, etc. The work is appropriate for the general public, businesses buying real estate, and businesses selling real estate. It may be useful for administration, organizations, universities, to defend themselves against earthquakes and other natural disasters etc., for instance, in making judgments on effective allocation or establishing new facilities. Usually, there are a lot of people in line to purchase a new home or apartment. They can obtain a better estimate of how much they should prepare for. Along with helping the consumer prepare, it might lessen the likelihood that businesses would trick clients by offering expensive lodging. similar to how any seller may profit from a rough price point estimate. We were able to determine the top and lowest prices for every region in Bangladesh, including the District, Upazila, Thana, and Village. Our suggested housing price predicting algorithm has a 94% to 95% accuracy rate. It was a challenge for us.

Multilingual Fake News Classification in Bangla and English with ML and Transformer Architectures.

Natural Language Processing

Explainable Federated Learning for Factory-Level Disease Burden Prediction in Bangladesh’s Ready-Made Garment Industry

This paper provides an empirical integration of federated learning (FedAvg and FedProx) with SHAP-based explainable AI for factory-level disease-burden-share regression using real multi-factory occupational health records from Bangladesh’s RMG sector.Key elements include client-partitioned (privacy-preserving) modeling across three factories that outperforms a matched centralized MLP baseline (≈14% lower macro TEST RMSE), held-out TEST SHAP analysis identifying Lag1_Burden_Share_01 as the dominant feature with relatively stable rankings across seeds, and demonstration that FedAvg and FedProx perform nearly identically in this setting. Results are specific to the three observed factories.

Cross-Corpus Transfer and Label-Calibrated Learning for Imbalanced Multi-Label Bangla Hate Speech Detection

The research introduces a cross-corpus encoder transfer framework for mismatched Bangla hate-speech taxonomies, using BANHATE as the source corpus and BOISHOMMO as the imbalanced multi-label target task. It further contributes a leakage-aware evaluation protocol, imbalance-aware weighted learning, validation-only threshold calibration, and three-seed probability ensembling without using the test set for model selection. The final system is evaluated once on a frozen test set with bootstrap uncertainty and post-hoc explainability analysis using Integrated Gradients.

Fish Disease Classification Through Deep Fusion of Visual and Environmental Data

The study’s significant contribution is the development of a multimodal deep-learning framework that integrates fish images with 72-hour environmental time-series data for seven-class fish-health classification. It introduces a dual-branch CNN–BiGRU architecture and demonstrates that multimodal fusion can outperform image-only and sensor-only approaches, achieving 98.36% accuracy and greater robustness when visual symptoms are ambiguous.

Conditional GANs for Joint Facial Denoising and Multi-Angle Synthesis

This paper presents a two-module conditional GAN pipeline. It first denoises a facial image, then synthesizes the same face at four additional viewing angles. A noise-conditional GAN handles the first stage, taking an explicit noise-level map as input so it can adapt its denoising strength across severities rather than applying one fixed strategy. The second stage is an attention-gated GAN conditioned on the target angle, generating four non-frontal views from a single frontal face while preserving expression. A face-alignment bridge connects the two modules, and experiments evaluate the full pipeline end-to-end across CelebA, KDEF, and real-world images spanning increasing domain distance. Module 1 reaches 38.41 dB PSNR and 0.9776 SSIM at the lowest tested noise level, dropping to 28.91 dB and 0.8575 SSIM at the highest. Module 2 reaches 23.90 dB PSNR and 0.7502 SSIM. The paper also analyzes how the pipeline behaves across domains and reports this in detail.

A Privacy-Preserving Depth-Based Human Identification Framework Using Consumer TrueDepth Sensing

This research presents a privacy-preserving depth-based human identification framework that integrates consumer TrueDepth sensing with hardware-isolated biometric authentication through Apple’s Secure Enclave and LocalAuthentication API. The key contribution is the development of a secure and modular biometric authentication architecture supporting template enrollment, real-time verification, and structured event logging. The framework was experimentally evaluated with 100 participants across normal lighting, low lighting, masking, and facial occlusion conditions. These results demonstrate the feasibility of using consumer depth sensing for secure, repeatable, and privacy-conscious biometric authentication and access control.

TCN+LightGBM: A Hybrid Deep Learning–Gradient Boosting Architecture for Interpretable Short-Term Electricity Demand Forecasting in Bangladesh

Main contribution:
The paper proposes a novel two-stage hybrid architecture for short-term electricity demand forecasting on Bangladesh’s PGCB grid data that is structurally different from prior hybrid approaches. Instead of fusing two neural branches end-to-end (as in [1] and [2]) or combining two independently-trained model outputs via linear regression (as in Gong et al. [4]), this study uses an already-trained TCN purely as a feature extractor: it takes the 32-dimensional learned embedding from the TCN’s penultimate layer and concatenates it with 49 hand-engineered tabular features (weather, calendar, economic, lag/rolling statistics) to form an 81-dimensional input, on which a LightGBM regressor is trained to produce the final prediction.

Why this matters (the novelty claim):

The final predictor is a tree ensemble, not a black-box neural network — so exact, native feature-importance analysis is possible instead of an approximate/surrogate explainer, which is a genuine interpretability gain over prior Bangladesh-grid hybrids.
It achieves 97.26% accuracy (2.74% MAPE) on the full hold-out set, and 1.87% MAPE on a literature-matched Jan–Feb 2024 window — beating Islam et al.’s 2.30% MAPE on the same window, and substantially better than Bhadra & Rois’s 5.45% MAPE (though that comparison is caveated by daily vs. hourly granularity).
The study is methodologically thorough: it adds a five-part evaluation protocol (point accuracy, statistical significance via Diebold-Mariano test, exact feature-importance/explainability, quantile-based uncertainty coverage, multi-step 24-hour forecasting, and 5-fold walk-forward robustness validation) that neither reference paper applies in full.
It also honestly reports a nuance/limitation rather than overselling: the Diebold-Mariano test shows standalone TCN actually has significantly lower squared error than the hybrid — meaning the hybrid’s advantage is specifically in relative-percentage error (MAPE), not absolute-MW error, a trade-off the paper explicitly flags for grid operators (peak-hour reserve planning vs. day-ahead trading use cases).

Addressing Extensive Missingness in Newspaper-Reported Suicide Data Through Imputation and Clustering Techniques

The study’s significant contribution is the development of an expanded, comprehensive dataset of newspaper-reported suicide cases in Bangladesh and a systematic evaluation of imputation–clustering approaches. It advances existing work by comparing these approaches against no-imputation baselines using a balance-aware cluster-validity metric, while also assessing their robustness across different metric weightings to identify the most effective combinations.

Not correct

A Vision-Based Gesture and Eye-Blink Controlled Exoskeleton Robotic Hand for Rehabilitation of Paralyzed Patients

This work proposes a novel, sensorless dual-modality robotic hand exoskeleton controlled entirely through computer vision via MediaPipe healthy-hand gesture and eye-blink tracking, eliminating the need for complex, calibration-heavy wearable biosignal sensors (EEG/EMG). The system introduces an end-to-end hardware control pipeline using the Web Serial API and an ESP32 microcontroller to drive an assistive glove. Empirical validation shows that while gesture control offers faster response times, a 2.0-second temporal gating mechanism for eye-blink control yields 100% gate-adherence accuracy, providing a reliable, calibration-free, and flexible assistive solution for patients with varying degrees of residual motor function.

A Blockchain-Based Revocation and Expiry Aware Framework Using AHIBE for Secure Document Verification

The problem is that while most document verification systems on the blockchain ignore revocation and expiry, there is an important gap in managing both of these simultaneously while also maintaining confidentiality and preventing on-chain storage overload. We use a Counting Bloom Filter on IPFS to support an O(1) revocation check and use AHIBE based encryption for documents that tie document accesses to its expiry, enforcing expiry cryptographically not just checking it against a timestamp. We also describe an actual one-time key-delegation method that verifiers have to implement in order to use the system, while previous formulations of this type of design simply left it unaddressed, and test the approach with a working prototype, not just theory.

A Hybrid Rule-Augmented Stacked Ensemble Framework for Market Basket Completion with Leak-Free Customer History

Predicting the next item a customer is likely to add to a partially filled shopping basket is a highly actionable sub-problem of recommendation, but existing approaches either rely on association rules that overstate confidence for small itemsets and ignore customer segment differences, or on machine-learning classifiers that discard the interpretable statistical structure of transactional data. This paper presents a hybrid rule-augmented framework for market basket completion that mines association rules separately per country-and-season segment and corrects them using Wilson-Score Adjusted Confidence (WSAC), Segment-Normalized Lift (SN-Lift), and a cross-fold rule stability score, all computed under a strictly out-of-fold protocol to prevent leakage. These rule features are combined with a leak-free customer purchase history feature and fed into six architecturally diverse base classifiers, fused through a stacked logistic regression meta-model. On the UCI Online Retail dataset (541,909 transaction rows, 15,619 baskets), the full pipeline lifts accuracy from an 80% majority-class floor to 87.3%, with the stacked model achieving the best overall AUC of 0.855. Customer purchase history alone is the single strongest signal, present in 42.3%
of correct predictions against 6.9% of incorrect ones. Evaluation uses Lift-over-Floor and Hit@K ranking metrics with 10-fold time-series cross-validation and paired bootstrap testing over held-out baskets, and confirms every reported gain is statistically significant (p < 0.05). Results show that segment-aware, leakage-free feature engineering, not classifier complexity, drives the performance gain, with all six base classifiers converging to a tight 0.869–0.873 accuracy band, and residual error concentrates in the
sparsest data segments rather than being distributed uniformly.

TumorMultiNet: A Statistically Validated Hybrid Deep Learning Framework for Brain Tumor MRI Classification and Segmentation on BRISC2025 with Explainable AI

Our contributions are:
(i) A unified two-branch benchmark on BRISC2025.
Thirteen models, seven ImageNet-pretrained classification backbones and six segmentation architectures, are
trained and evaluated under one preprocessing, optimization and evaluation protocol on identical splits, so the
numbers are comparable across models and across both
tasks.
(ii) TumorMultiNet, a hybrid for each task.
TumorMultiNet-Cls fuses EfficientNet-B0 and ResNet50
by weighted soft voting with test-time augmentation;
TumorMultiNet-Seg averages the probability maps of
SegFormer and DeepLabV3+. Members and fusion
weights come from validation data only, with the test
split untouched.
(iii) Evidence on whether hybridizing is worth it. Every
metric carries a bootstrap 95% confidence interval and
a paired test against each constituent (McNemar’s exact
test; Wilcoxon signed-rank), which separates a real gain
from a cosmetic one that point accuracy would hide.
(iv) A deployment-oriented efficiency profile. Parameters,
GFLOPs, latency, checkpoint size and peak memory are
reported for all thirteen models, quantifying what each
hybrid costs to run.
(v) Explainability and failure characterization. GradCAM confirms the classifier attends to tumor tissue, and
an error analysis ties segmentation failures to small and
multifocal lesions and classification errors to anatomically adjacent classes.
(vi) A working web interface. Both branches are wrapped
in a browser tool returning class, confidence and segmentation overlay for an uploaded slice.

Deep Learning-Based Hybrid Framework For Pneumonia Detection From Chest X-Ray Images with XAI And Real-time Web-interface

The main contributions of this paper
to automated pneumonia screening are:
•A pipeline to preprocess the raw chest X-ray images by
resizing, denoising using Gaussian filter, pixel normalization, and enhancing contrast cleaning up images before
they are fed to any model.
•Compare different transfer learning models (custom
CNN, VGG, ResNet, DenseNet) to observe the performance of each model in pneumonia detection, not just
assuming one architecture is sufficient.
•A hybrid ensemble of the two best-performing pretrained
models with soft voting probability aggregation, designed
to achieve higher accuracy, stability and recall than any
single model can achieve on its own.
•Web-based deployment of the final ensemble pipeline, so
that it can be used for real-time diagnostic support
particularly in areas with limited medical resources.

Archives: Paper Submission