Article 1
A Study of Machine Learning Models with Population and Swarm Based Optimization for Cardiovascular Disease Prediction
Ruponti Muin Nova, Jawad Anzum Fahim, Sharmeen Jahan Seema
Pages: 1-24
Cardiovascular Diseases (CVD) remain a leading cause of global mortality, necessitating accurate and scalable diagnostic tools. This study proposes a machine learning (ML)-based framework for early CVD prediction by integrating advanced feature selection, model tuning, and hybrid optimization techniques. Five ML models were evaluated, and their performance was enhanced through the application of Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) for hyperparameter optimization. Feature selection techniques were employed to identify clinically relevant predictors, including binning strategies for age, BMI, and MAP. PSO was applied within a cross validation (CV) loop, yielding a notable improvement in Logistic Regression (LR) performance raising CV accuracy from 81.5% to 82.51%. Similarly, GA optimization improved cross validation accuracy for Random Forest (RF) and XGBoost from baseline test accuracies of 82.03% and 81.1% to 82.51% and 82.04%, respectively. Evaluation metrics such as precision, recall, F1-score, and AUC were also utilized to ensure robust model assessment, especially considering imbalance within the dataset. A comparative analysis against traditional hyperparameter tuning methods (e.g., GridSearchCV) demonstrated the superiority of PSO and GA in enhancing predictive performance. However, the study is limited using of a single public dataset without external clinical validation. These findings highlight the necessity for more extensive data validation in subsequent studies, but they also highlight the promise of hybrid optimization and feature engineering in improving ML-based CVD diagnosis.
Article 2
Data Center Standards: A Review of Metrics and Standardization Gaps
S M Saiful Islam, Sayma Alam Suha, Tasnim Binte Shiraj
Pages: 25-52
Certifying data centers (DCs) necessitates a rigorous evaluation of diverse metrics to assess performance, efficiency, sustainability, and operational effectiveness. Key technological and infrastructural parameters are analyzed to ensure compliance with industry standards. However, inconsistencies in measurement methodologies often lead to discrepancies in evaluation, highlighting the need for standardized metrics. Current industry standards, including Uptime Institute Tier Classification, ANSI/TIA-942, ASHRAE, The Green Grid (TGG), NIST, and EPA ENERGY STAR—exhibit significant variations in scope, methodology, and applicability. A comparative analysis reveals that while these frameworks emphasize infrastructure redundancy and energy efficiency (e.g., PUE, MTBF). While energy efficiency remains a critical challenge, a more comprehensive certification approach must integrate additional dimensions such as sustainability, operational reliability, availability, and security. This paper presents a systematic review of existing data center metrics, focusing on their role in standardization and certification. These metrics are categorized into six key sub-dimensions: efficiency, sustainability, operations, reliability, availability, and security. The findings underscore the need for a unified certification framework that balances physical infrastructure metrics with emerging requirements such as software resilience, real-time adaptability, and cross-platform interoperability. Such an approach would enable more comprehensive benchmarking, particularly for hybrid and hyperscale data centers. This approach will not only improve the certification evaluation process but also enable a more holistic assessment of data center performance.
Article 3
Bangla Sign Language Recognition: A Comprehensive Review of Machine Learning Approaches and Data Sources
Jasiya Fairiz Raisa, Rumana Yasmin
Pages: 53-86
Sign language is the primary medium of communication for deaf and dumb individuals, but it is difficult to interpret for every demographic, which makes communication extremely difficult. Bangla is among the most widely spoken languages worldwide, and substantial research on Bangla Sign Language (BdSL) has emerged to address this issue. In recent years, researchers have been working to automate BdSL recognition using different techniques. This review paper evaluates research trends in BdSL by comparing the features and evaluation outcomes of various systems and approaches applied to both existing and novel datasets. We have gathered and integrated metadata from datasets encompassing all BdSL alphabets and numbers implemented to date. The analysis of this paper shows that most suggested models work well on images with static and single-handed signs, but performance drops in complicated backgrounds. Additionally, we concentrated on identifying insights and parallels within the existing systems, identifying research gaps, and suggesting potential future directions.
Article 4
HDP: Simulated IoT Architecture Design Applying GridSearchCV
Nusrat N. Suzana, K. Habibul Kabir
Pages: 87-110
Globally, cardiovascular disease remains a major burden of morbidity and death with more than half a billion affected population. Heart Disease Prediction (HDP) is strenuous for medical staff because of the risk factors. Classifying people into various risk groups according to their traits and medical history can be done effectively and powerfully leveraging machine learning techniques. This research study leverages the Framingham Heart Study dataset to propose a machine learning enabled system regarding heart disease prediction (HDP). Utilizing GridSearchCV, six supervised machine learning algorithms were trained and optimized: gaussian naive bayes, k nearest neighbour, decision tree, support vector machine, random forests, and logistic regression. With 97.9% accuracy, 96.4% precision, 99.5% recall, and 97.9% f1-score, random forest outperformed the six algorithms with a 10-fold validation technique. To illustrate the prediction model’s possible real-time integration into patient monitoring systems, a notional Internet of Things (IoT) architecture was created. Even though the architecture is now just in simulation, it offers a framework for future use with real IoT infrastructure. The results demonstrate that combining ML with IoT for early cardiovascular risk prediction and decision assistance is both feasible and beneficial.
Article 5
Analysis of Safety Margin Aspects of the BAEC TRIGA Research Reactor using the Deterministic Code System SRAC2006
Md Masud Rana, M J H Khan, Md. Asaduzzaman
Pages: 111-124
The goal of this study is to test the ability of the BAEC 3 MW TRIGA Mark-II research reactor, intended to run at its steady state full power 3 MWth with suitable safety limits. The neutronics design safety margins of the core were calculated based on the deterministic method. Two methods have been applied, and they are (i) the collision probability lattice transport code SRAC-PIJ for the region averaged group coefficients of fuel and non-fuel units separately, and (ii) the neutron diffusion theory code SRAC-CITATION for global core computations. The realistic calculated safety margin features of the BAEC TRIGA core are effective multiplication factor (Keff), surplus reactivity (ρex), efficacy of control rod, shutdown margin (SDM), safety reactivity factor (SRF) and fuel temperature reactivity coefficient (FTC). The calculated results were evaluated with the experimental data, as well as the values provided by the International Atomic Energy Agency (IAEA) and the results obtained from the Monte Carlo N-Particle (MCNP) and WIMS-CITATION simulations. The IAEA safety principle says that the safety reactivity factor of a research reactor should be greater than 1.5 and the calculated values of SRF are 1.588 and 1.589, which exceeds 1.5 and also indicates that the reactor exhibits improved operational and safety characteristics, which echoes that the SRAC2006 code system yields the physical TRIGA core model truthfully. Therefore, the analysis gratifies each safety margin feature of existing core arrangement of 3 MW TRIGA Mark-II research reactor following IAEA safety viewpoints for its effective safe operation.
Article 6
An Explainable Ensemble Learning Framework for Brain Tumour Using Pretrained Models and XAI Techniques
Muhammad Mahbub Sarwar Shafi, Jannatul Ferdous Mirza
Pages: 125-146
Brain tumours pose serious health challenges worldwide, demanding early and accurate diagnosis for effective treatment. Traditional diagnostic methods are time-consuming and subject to human error. While deep learning models have advanced automated brain tumour classification, their black-box nature limits clinical trust. This study introduces an ensemble learning framework that integrates EfficientNetB0, InceptionV3, and DenseNet for improved classification accuracy. High-pass filtering (HPF) is employed in the preprocessing phase to enhance tumour-specific features. Three machine learning classifiers—Random Forest, Support Vector Machine, and Logistic Regression—are combined using a soft voting strategy to deliver final predictions. To ensure interpretability, Local Interpretable Model-Agnostic Explanations (LIME) is used to visualise feature contributions. The proposed method achieved a validation accuracy of 97.3% and an AUC-ROC score of 0.9986 on a publicly available brain MRI dataset. These results outperform individual models and enhance clinical reliability. This study demonstrates that combining deep learning, ensemble learning, and XAI techniques can bridge the gap between model performance and interpretability, promoting real-world adoption of AI-based medical diagnosis.
Article 7
Comparison of Machine Learning for Sentiment Analysis in Movie Reviews
Jannatul Ferdous Mirza, Muhammad Mahbub Sarwar Shafi, Saeed Hossain Moheb
Pages: 147-170
This study presents a comparative analysis of machine learning and deep learning algorithms for sentiment classification in movie reviews. Three benchmark datasets—IMDb (50K and 20K reviews) and Rotten Tomatoes—were used to evaluate six classifiers: Na¨ıve Bayes, Logistic Regression, Support Vector Machine, Decision Tree, Random Forest, and BERT. Preprocessing included tokenization, stop-word removal, stemming, and feature extraction using Count Vectorizer and TF-IDF. Evaluation metrics such as accuracy, precision, recall, sensitivity, specificity, and F1-score were used to assess model performance. Logistic Regression achieved 88% accuracy on the IMDb dataset, while Random Forest exhibited the highest specificity. BERT outperformed traditional models in both accuracy and F1-score across all datasets, particularly in handling informal and context-heavy language. The results highlight the impact of dataset characteristics on classification performance and provide insights for deploying sentiment analysis in real-world applications like recommendation systems and audience profiling.
Article 8
A Systematic Analysis of Machine Learning and Deep Learning Techniques to Detect Fake News
Nazneen Akhter, Sahara Alam Tuba, Afrina Khatun
Pages: 171-184
Expanding social media platforms produce a large amount of news daily that is not always accurate, true, or reliable. Some dishonest news vendors intentionally generate misinformation to misguide public views, manipulate people’s opinions, and further create a political agenda or confusion for their economic benefit. Automated techniques such as Machine Learning (ML) and Deep Learning (DL), are frequently used to prevent the spread of fake news. This work conducts a systematic assessment of the existing ML and DL strategies with a focus on data pre-processing, feature extraction, feature selection, and contextual analysis that enhance the performance of the fake news detection model. A comparison-based description of the fundamental models of ML and DL is provided, where it is found that DL techniques are more explainable and perform well in detecting fake news than ML classifiers.
Article 9
Towards Early Intervention for Panic Disorder Detection and Dominant Feature Selection through Machine Learning Techniques
Sayma Alam Suha, Mosa. Sumiya Akter
Pages: 185-205
Panic disorder, marked by recurrent and unexpected panic attacks, significantly impairs daily functioning and overall well-being. Early detection is crucial to improving patient outcomes, yet traditional diagnostic methods often delay timely identification. This study investigates the application of machine learning (ML) techniques for the early detection of panic disorder and the identification of key features that contribute to its development. Utilising a comprehensive dataset of clinical and physiological data including demographics, symptoms, and vital signs from individuals with and without panic disorder, multiple ML classification algorithms were trained, tested, and evaluated. The ensemble voting feature selection method was employed to pinpoint the most relevant predictors of panic disorder. Among the models tested, the Extra Tree Bagging Ensemble ML model demonstrated exceptional performance, achieving 99.8% accuracy, along with high sensitivity and precision. Feature significance analysis revealed critical physiological and psychological factors associated with panic vulnerability, offering valuable insights into the disorder’s underlying mechanisms. This research underscores the potential of ML-based approaches in enabling early detection of panic disorder, paving the way for personalised prevention and intervention strategies. The findings highlight the importance of integrating advanced computational techniques in mental health diagnostics to enhance accuracy and timeliness in identifying panic disorder.
