Current machine learning algorithms have achieved remarkable success in quantitatively interpreting spectral gamma-ray logging for uranium exploration. However, several challenges remain: (1) Existing models typically neglect intervariable dependencies among multiple output variables and (2) they generally lack interpretability. To address these issues, we propose a multitarget regression Takagi–Sugeno–Kang fuzzy model (MTR-TSK) in this paper. The MTR-TSK employs fuzzy rules to model feature-target relationships and jointly optimizes multitarget high-order correlations through fuzzy inference and regression loss during training. The proposed model not only models the complex relationship between the input and output spaces but also captures and exploits the inter-variable correlation structure to enhance predictive performance. In addition, the model features a transparent reasoning process, allowing for the identification of feature importance. The model demonstrates strong generalization performance on public benchmark datasets and achieves notable results on our custom uranium mine logging dataset.
Long-term corporate solvency forecasting is fundamentally constrained by two intertwined obstacles: severe class imbalance in default events and the nonstationary dynamics of financial markets. Deep learning architectures optimized on short-horizon objectives characteristically overfit to transient market noise, yielding predictive performance that deteriorates sharply as the forecasting horizon extends. This instability is addressed through the Gated Spectral–Temporal Transformer (GST-Net) a hybrid architecture that introduces a learnable gating mechanism to suppress transient, high-frequency signal components in favor of persistent structural trends. A dual-branch design further decouples the modeling of localized anomalies from global temporal dependencies, enabling the network to preserve discriminative signal across multiyear horizons where conventional models degrade. Given the extreme class imbalance inherent to solvency prediction, the evaluation protocol centers on the Matthews correlation coefficient and macro F1 score, metrics that penalize minority-class misclassification far more rigorously than accuracy alone. At a three-year forecasting horizon, GST-Net achieves an MCC of 0.5704, a macro F1 score of 78.38%, and an accuracy of 95.95%, nearly doubling the MCC of the strongest tuned baseline, XGBoost, which peaked at 0.3294, while state-of-the-art tabular deep learning models, TabNet and TabPFN, collapse entirely under identical conditions. Beyond predictive robustness, GST-Net sustains subsecond inference latency of approximately 0.37 s per instance, rendering it practical for quarterly operational auditing at scale. Coupling the model with SHAP and LIME yields the interpretability required for regulatory scrutiny, exposing an economically coherent shift from short-term liquidity stress to long-term structural decay as the dominant predictive signal, an insight with direct implications for early-warning credit risk systems.
Clinical data often exhibit heterogenous feature distributions, nonstationary physiological dynamics, and class imbalance, which collectively increase the complexity of accurate prediction. Despite the popularity of support vector machines (SVMs) in clinical data classification, the traditional kernels have globally fixed hyperparameters and may fail to reproduce the local properties of data. We propose an adaptive density–variance (ADV) kernel for the SVM classifier for addressing these shortcomings. The ADV kernel is based on sample-conditioned bandwidths that adapt to the local neighborhood density and the feature-level variance within each sample’s k-nearest neighbors, directly encoding clinical heterogeneity into the similarity space. It is constructed as a combination of valid kernel Hadamard products (elementwise), an adaptive RBF component, and an interaction term, which is a polynomial. This design explicitly captures localized similarity structures alongside global feature correlations. Furthermore, the resulting kernel matrices behave as an empirically verified positive semidefinite (PSD) via eigenspectrum analysis. This kernel is fully sklearn-compatible via the BaseEstimator interface (available: https://pypi.org/project/adv-kernel/). The proposed approach is tested according to 6 benchmark clinical datasets, including breast cancer, heart disease, PIMA diabetes, fetal health, Parkinson’s disease, and chest X-ray pneumonia. Clinically relevant metrics are compared between the performance and conventional SVM kernels and standard machine learning models. Statistical significance is determined through Friedman and Wilcoxon signed-rank tests, and an ablation study is performed to determine the contribution of each kernel component. To overcome the black-box nature, SHAP-based explainability and permutation feature importance are also implemented. Nested cross-validation does offer the reliability of the generalization estimates and unbiased model selection. Experimental findings reveal that the ADV kernel has very competitive performance, especially in F1-score, balanced accuracy, and agreement measures, which means that it is appropriate in heterogeneous and clinically sensitive healthcare contexts.
This study compares five supervised machine learning models in their ability to classify students as Regent Admission Index (RAI)–eligible or not RAI-eligible—that is, whether they meet the RAI automatic admission threshold. Because this binary outcome is computed deterministically from the same three predictors used by the models, the task is framed as a methodological comparison of how accurately and stably different algorithm families recover an existing threshold rule, rather than as a prediction of an independent admission, college-readiness, or college-success outcome. The study compares linear discriminant analysis (LDA) and k-nearest neighbor (KNN) as distance-based models, conditional inference tree (CIT) and random forest (RF) as tree-based models, and Naïve Bayes as a Bayes’ theorem-based model, using ACT composite scores, high school GPA, and number of core courses completed as predictors. Results show that all five models achieve high classification accuracy, with RF achieving the highest accuracy (99%). Bootstrapping with 10 resamples indicates stable performance, as reflected by low standard deviations across resamples. Small adjustments to the cutoff score cause slight changes in accuracy rankings, but the top and bottom performers remain consistent. Across models, ACT scores are strongly associated with the eligibility classification; however, the relative importance of predictors differs across models and from the policy weights assigned by the operational RAI formula, reflecting how each algorithm separates students near the threshold rather than evidence about college readiness.
This paper proposes a more efficient resource-constrained decoding algorithm for Adaptive Offset Min-Sum (A-OMS) low-density parity check (LDPC) optimized decoder parameters in fifth-generation New Radio (5G-NR) Base Graph 2 codes (BG2). To curb the challenge of observing the all-time finest offset to use the irregular BG2 structure, a genetic algorithm (GA) is utilized to enhance the respective offset aspect in the updating of check nodes to achieve the required objective of creating a near-optimal solution offset aspect. Depending on the GA, the bit error rate (BER) can achieve optimal performance over a wide range of signal-to-noise ratios (SNRs), which is a metaheuristic optimizer. These results of the simulation indicate that this proposed optimized decoder outperforms the conventional fixed-offset OMS by 0.2–0.5 dB in terms of belief propagation (BP) decoding performance, coding gains, and a BER that are 0.3–0.6 dB and a BER of 4.6, respectively. The proposed method achieves BP at about 35 percent decoding computation. In the case of 5G-NR next-generation receivers, the suggested approach will provide a fair trade-off between the implementation efficiency and performance.
Metaheuristic algorithms reduce the size of combinatorial test suites, lowering testing costs. Researchers have developed several combinatorial test suite generators in the literature that adopt metaheuristic algorithms, such as particle swarm optimization (PSO), harmony search strategy (HSS), and whale optimization algorithm (WOA). Although many studies report near-optimal results, some do not address the representation of combinatorial optimization. This paper models combinatorial optimization in a two-dimensional (2D) search space using WOA to address the limitations of 1D representations, which may limit search capability, especially in bio-inspired algorithms. The proposed model represents the combinatorial optimization problem in a 2D search space. To evaluate effectiveness, a bio-inspired WOA is adopted, as it mimics living organisms’ behavior beyond 1D space, enabling exploration of multidimensional combinatorial configurations. Metaheuristic algorithms update candidate solutions as position vectors in a multidimensional space. In WOA, each agent updates its position relative to the best solutions using vector-based distance calculations. Representing solutions using spatial coordinates improves neighborhood-based search by enabling spatial proximity between candidate solutions. In contrast, linear index representations in combinatorial testing may limit this spatial interpretation, motivating the proposed 2D model. Results show that 53% of experiments achieve smaller test suite sizes with the 2D representation. Statistical analysis using the Wilcoxon signed-rank test (p<0.05) indicates a significant difference between 1D and 2D representations. These findings demonstrate that the 2D model improves the ability of WOA to reduce test suite size. This study presents a 2D combinatorial model using WOA that outperforms most existing strategies in producing smaller test suites. The implications of applying the model extend beyond the current algorithm, offering significant advantages over other algorithms that generate near-optimal numbers of test suites.
Accurate and interpretable brain tumor classification from MRI scans remains a critical challenge in medical image analysis, owing to limited annotated data, high intraclass variability, and the need for trustworthy models. We present an F1-weighted ensemble that integrates handcrafted texture descriptors (LBP) with deep features extracted from lightweight convolutional neural networks ([CNN] MobileNetV2, SqueezeNet 1.0, and EfficientNet-B0). The CNN backbones are used as frozen ImageNet feature extractors without any fine-tuning, ensuring a modular and leakage-free pipeline. Grad-CAM visualizations are generated using a separate, task-specifically fine-tuned CNN instance (5 epochs, intrafold), ensuring that saliency maps reflect brain tumor-relevant activation patterns rather than generic ImageNet features. Each component model is weighted according to its macro-F1 score estimated on an internal 20% validation subset of the training fold, preventing weight collapse due to classifier overfitting on training data. The optimal number of neighbors for the KNN classifier was determined by a cross-validated grid search over k∈1,3,5,7,9, yielding k=1 as the best configuration (macro-F1 score = 97.38 ± 0.39%). The framework is evaluated on two heterogeneous datasets, including a rebalanced 15-class variant, under a strict stratified 5-fold cross-validation protocol in which normalization and augmentation are applied exclusively within the training partition of each fold, preventing any statistical leakage between training and validation sets. It achieves 97.93%±0.18% accuracy (97.91%±0.19% macro-F1 score) and 99.86%±0.06% AUC on the 4-class dataset, and 97.75%±0.65% accuracy on the challenging 15-class setting. Beyond predictive performance, we provide technical post hoc model explanation through Grad-CAM, SHAP (stability σ¯=4.4×10−5, three independent runs, extended to both 4-class and 15-class settings), and error-correlation analysis, which highlight the complementary nature of CNN and LBP features. A representative state-of-the-art (SOTA) method was reproduced under identical 5-fold cross-validation conditions Acc=99.38%±0.23%, confirming that the residual performance gap reflects deliberate design trade-offs—computational efficiency, interpretability, and robustness under class imbalance—rather than evaluation protocol heterogeneity. This work introduces a rigorous and computationally efficient framework that combines performance and transparency, paving the way for reliable computer-aided systems to support neuroradiologists in brain tumor diagnosis.
Explainable artificial intelligence (XAI) is being actively pursued to promote the transparency and fairness of AI-based financial decision-making systems, such as credit scoring systems, loan approval systems, fraud detection systems, and risk assessment systems. However, machine learning algorithms developed based on past financial information can be biased and, therefore, yield discriminatory results toward specific demographic groups. To address this problem, the paper presents a framework called Correcting Algorithmic Discrimination with Scaled Priority Feature Vector (CAD-SPFV), which involves XAI techniques to identify and address the bias in financial decision models. The proposed approach combines the feature importance derived by SHAP with the bias evaluation based on correlation and counterfactual explanations to pinpoint the features linked to discriminatory results and mitigate them by using fairness-conscious optimization. The experiments based on a real-life financial lending dataset (UCI Adult Income dataset, 48,842 instances) show that the proposed CAD-SPFV framework is effective in enhancing predictive performance and equity. The framework is 91.63% accurate and has an F1-score of 0.914 with fairness scores such as a demographic parity difference of 0.061 and an equal opportunity difference of 0.052, significantly exceeding current methods. Such results show that the suggested framework can significantly increase the interpretability, equity, and transparency of AI-based financial decision-making systems and provide a feasible way to make AI-based finance responsible and regulation-conscious.
Brain tumors can be diagnosed and prognosed using magnetic resonance imaging (MRI). The intricate nature of brain tumors makes them difficult to diagnose. In this study, we evaluate the performance of two vision transformer (ViT) models (MobileViT-V2_175 and MobileViT-V2_200) and four advanced pretrained convolutional neural networks (CNNs), ResNet18, DenseNet201, MobileNetV2, and InceptionV3, on the publicly available Masoud dataset (7023 MRI images compiled from SARTAJ, Br35H, and Figshare). Images are categorized into four groups: glioma, meningioma, pituitary tumors, and healthy brains. The data were preprocessed using standard methods (train-test-validation split, imgaug-based augmentation to balance classes to 1850 images each, and normalization to [0,1]). The models were implemented with PyTorch with input size 224 × 224 × 3, batch size 4, Adam optimizer (learning rate 0.0001), and early stopping. The proposed approach achieves state-of-the-art overall accuracy of 99.61% (MobileNetV2 and MobileViT-V2_175), followed by MobileViT-V2_200 (99.47%), DenseNet201 (99.05%), InceptionV3 (99.00%), and ResNet18 (98.42%). Through 5-fold cross-validation, confusion matrices, and ROC-AUC curves (mean AUC = 1.00), these results demonstrate the efficacy of lightweight ViTs alongside CNNs for classifying brain tumors accurately. It further discusses the integration of these models into the Internet-of-Medical-Things (IoMT) framework for real-time remote monitoring, extending traditional classification to practical clinical applications. In the early detection of brain tumors, the new generation of hybrid CNN-transformer techniques shows great promise. The work applies MobileViT architectures to the Masoud dataset while providing actionable insights for IoMT-enabled assistive healthcare.
Bearings-only tracking (BOT) in sonar’s passive mode presents significant nonlinearity challenges, addressed here through a measure of nonlinearity (MoN) based on the filter’s covariance matrix. This study computes MoN to evaluate filter performance and enhance target motion estimation precision. To advance beyond traditional methods, we introduce a novel LSTM-based MoN framework that captures temporal nonlinear patterns, an adaptive MoN using reinforcement learning for dynamic adjustments, and attention-based multisensor fusion for improved accuracy. MoN values are estimated across various filtering techniques, including extended Kalman filter (EKF), unscented Kalman filter (UKF), particle filter (PF), and the proposed enhancements. Monte Carlo simulations (1000 runs) demonstrate the following quantified improvements: the LSTM–MoN achieves a 10%–15% RMSE reduction during maneuvering phases over static covariance-based MoN methods and a 24.8% overall RMSE improvement over the Gaussian sum shifted Rayleigh filter (GSRF) in Scenario 1; the adaptive Q-learning MoN provides a 19.8% improvement over the constrained multiple model particle filter (CMM-PF) in high-noise maneuvering conditions; and the attention-based fusion accelerates convergence by 20% and delivers an 18.2% accuracy gain over fixed-weight fusion. Collectively, these innovations achieve an overall RMSE reduction of 14.5%–25% against contemporary benchmarks, with detailed trade-offs in computational complexity. These advances enhance robustness in multisensor underwater environments, validated under varying noise levels (1°–10°) and target maneuvers.
Magnetic resonance imaging (MRI) is widely used for brain-tumor assessment, but conventional deep learning workflows often require patient images to be centralized, creating privacy, governance, and regulatory barriers. This study presents a lightweight federated learning (FL) framework for brain MRI classification using a multilayer perceptron (MLP) trained with the Flower framework and Federated Averaging (FedAvg). In the revised formulation, the task is treated strictly as classification: each MRI-derived sample is represented by a 239-dimensional feature vector and assigned to one of the encoded diagnostic classes in the label column. Segmentation terminology has been removed or clarified, and the Dice coefficient is reported only as a class-level F1/Dice-style classification metric rather than as a pixel-level tumor-overlap score. The prepared BraTS-derived subset used for simulation contains 3000 .npy samples organized for training, normal/abnormal representation, and mask-associated metadata; the full BraTS source and the preprocessed subset are now distinguished to avoid confusion. The FL setting is explicitly described as a simulation in which a single prepared dataset is partitioned into artificial clients, not as a true multihospital deployment. The proposed MLP contains 39,814 trainable parameters, requiring approximately 0.152 MB per model update in a 32-bit floating-point format, which makes it substantially cheaper to communicate than typical CNN-based alternatives. The best reported FL–MLP configuration achieved 95% accuracy, 0.92 class-level Dice/F1, 0.91 sensitivity, and 0.97 specificity. The revised manuscript also clarifies that FL reduces the need for raw data sharing but does not provide formal privacy guarantees unless combined with secure aggregation, differential privacy, encryption, or attack-resistance mechanisms.
Breast cancer continues to be one of the most common cancers in women throughout the entire world. Medical professionals analyse mammogram images to identify suspicious patterns, which most frequently appear as masses. The development of these masses stems from groups of cells that have become more dense than their surroundings. Medical diagnosis requires computer systems that use machine learning to deliver precise evaluation results. The process of extracting features and selecting important ones from large datasets remains difficult because these methods aim to achieve their highest possible classification performance. Nature-inspired metaheuristic algorithms have shown promise in tackling these optimisation problems. The research applies AlexNet for transfer learning to extract deep features, while GBMPA operates as an improved version of the Gompertz binary marine predators algorithm and ELM functions as the fitness function to boost classification results and reduce breast mass classification feature requirements. This research tests the proposed algorithm by comparing with seven different standard datasets, which serve as benchmarks. The performance metrics of the proposed algorithm is better compared to the results from EO, GWO, PSO, WOA, AOA and HHO algorithms.
The non-independent and identically distributed (non-IID) nature of data across different sites and the dynamic participation of client nodes represent two critical challenges faced by distributed industrial visual defect detection systems. While asynchronous federated learning (AFL) reduces training inefficiencies, its asynchronous update process makes the global model more vulnerable to data heterogeneity. This research proposes a contrastive clustering–based asynchronous federated learning (CC-AFL) framework. The method integrates contrastive learning into the client clustering process, working simultaneously across instance-level and cluster-level spaces to identify hidden similarity patterns among participating clients. The server maintains update queues for each cluster and then applies staleness-weighted aggregation based on the combined timestamp. Meanwhile, the framework uses client scheduling based on cluster queue lengths, giving priority to clusters with fewer updates to balance update frequencies across different clusters. Without requiring raw data sharing, the framework achieves improvements in both convergence speed and stability of AFL under changing scenarios. On the furniture board dataset, CC-AFL achieves better performance with an average area under the receiver operating characteristic curve (I-AUROC) of 74.38%, showing a 2.22% improvement over AFL-DCS under different non-IID settings. Experimental results show that CC-AFL maintains stable convergence under different conditions, confirming the effectiveness and practical value of contrastive clustering and scheduling in AFL.
The invention of infrastructure facilities in intelligent transportation systems is skyrocketing every day. The improvements in such things have many pros and cons. For instance, the accident rate increases due to a rise in the number of vehicles and changes in weather conditions. Therefore, it became very essential to find a method to remove the effect of the weather conditions in the images of the moving vehicle. Detecting vehicles and their number plates in foggy weather can be challenging due to reduced visibility and contrast. Numerous studies and developments have already taken place to solve these problems, but the concrete solution is still not available. The transformer with convolutional networks is one of the most recent, effective, and widely applicable methods for image processing. In this paper, a Light Weight Vision Transformer for License Plate image dehaze named LPDehaze–Transformer is proposed. This paper focuses on license plate image reconstruction using a lightweight Vision Transformer that combines perceptual loss with Charbonnier loss, specifically targeting improved performance in haze conditions.
A novel metaheuristic algorithm named worm optimization algorithm (WormOA) is introduced. WormOA is embedded in an evolutionary search framework that models the behavior of internet worms when they attack computer systems. The algorithm employs a frame-shifting technique to enable worm replication. Additionally, the transferring phase in worm attacks is implemented using the SIRS propagation model. The paper evaluates the proposed approach on 20 conventional benchmark optimization functions, along with five CEC 2019 functions and five CEC 2011 functions, to demonstrate its performance. The algorithm is assessed using metrics such as best cost, average cost, standard deviation, and convergence rate and is compared with eight established and novel metaheuristic algorithms from the current state of the art. The results indicate that WormOA outperforms the other metaheuristic algorithms and exhibits a high convergence speed.
Breast cancer is a widespread cancer among women worldwide. Automated segmentation of breast cancer from imaging modalities is essential in conducting an early diagnosis and optimizing treatment. This paper proposes a new deep learning model called DeiT–ASPP–UNet, which has never been employed before for an accurate breast cancer segmentation from mammography images focusing on regions of interests (ROIs). The proposed model includes a lightweight DeiT-based transformer encoder for contextual representations, an atrous spatial pyramid pooling (ASPP) module for enabling an accurate multiscale feature extraction, and an effective UNet decoder to ensure precise lesion boundary delineation. The experimental study was performed on ROI extracted from public datasets, encompassing INbreast, CBIS-DDSM, and DMID. Furthermore, an ablation study was conducted to evaluate the contribution of each component of the proposal. The experimental results indicate that DeiT–ASPP–UNet outperforms concurrent methods, SwinUNet, ResUNet, UNet, UNet++, TransUNet, and PSPNet, in terms of Dice coefficient and IoU.
The exponential growth of high-dimensional medical data has led to critical challenges such as increased computational complexity and reduced predictive accuracy in machine learning models. Feature selection has become a key solution to mitigate these issues by reducing data dimensionality and enhancing computational performance. However, conventional filter methods often ignore interactions among features, while wrapper methods, despite achieving higher accuracy, suffer from heavy computational costs. To overcome these limitations, this study introduces a hybrid filter–wrapper framework that integrates the advantages of both approaches. In the filter stage, an ensemble filtering method combined with graph-based ranking is employed to assess feature relevance and interdependencies. Additionally, a novel trigonometric optimization algorithm, termed SCOA, is proposed and shown to outperform existing techniques. In the wrapper stage, a binary version of SCOA (BSCOA), utilizing an inverse tangent (arctan) transfer function, is developed to enable multiobjective optimization of feature subsets. Evaluation conducted on 27 high-dimensional medical datasets and COVID-19 case study demonstrate that the proposed approach consistently achieves superior accuracy, stability, and computational efficiency, while in many cases selecting less than 1% of the original features.
This study applies machine learning to predict customer churn in commercial banking, using a publicly available dataset of 10,000 customers from Kaggle, comprising 14 demographic, behavioural and account-level features with an approximate 4:1 class imbalance ratio between nonchurners and churners. The 7 classifiers (logistic regression, decision tree, random forest, support vector machine, K-nearest neighbours, LightGBM and XGBoost) were tested on five performance metrics: precision, recall, F1-score, AUC-ROC and MCC. To address class imbalance, we applied SMOTE + ENN to the training set following data partitioning, while the test set retained its original distribution for unbiased evaluation. The optimisation of hyperparameters was done using Grid Search with stratified cross-validation. In addition to predictive benchmarking, this study proposes a three-level business analytics framework: Layer 1 filters significant churn predictors using permutation feature importance; Layer 2 utilizes K-means to categorize users into interpretable risk tiers; and Layer 3 leverages a bank-specific cost matrix to identify the optimal threshold that maximizes net financial value. Random forest outperformed other classifiers, achieving the highest AUC-ROC, F1-score and MCC, potentially the most effective model for customer churn prediction. The proposed framework provides a pipeline for commercial banks to move from a reactive to a proactive customer retention. Moreover, it provides an interpretable model and is both computationally efficient and practical, given the operational constraints of commercial banking. This framework can be extended in future work by leveraging longitudinal data at the transaction level with sequential techniques for better prediction regarding “when” customers actually churn.
Fraudulent SMS messages are a significant threat in Pakistan, impacting financial security, identity theft, and misinformation and enabling exploitation by adversaries. This study proposes a localized AI-based framework for SMS fraud detection, utilizing machine learning (logistic regression [LR], random forest [RF], support vector machine [SVM], extreme gradient boosting [XGB]) and deep learning models (multilayer perceptron [MLP], long short-term memory [LSTM]) alongside an ensemble voting classifier. The framework is built on a curated dataset of 5617 SMS messages, including fraud, promotional, and normal categories, which were preprocessed with tokenization, normalization, and TF–IDF feature extraction. Experimental results show that ensemble and tree-based models achieved accuracy exceeding 0.95. Adversarial robustness testing, including synonym substitution, typo injection, noise addition, data poisoning, and label flipping, was conducted to evaluate model resilience. Although the models maintained high baseline accuracy, adversarial attacks underscore the need for robust evaluation in security-critical applications. This research contributes to (i) the creation of a representative SMS dataset tailored to Pakistan, (ii) a comparative analysis of classical, deep, and ensemble learning models, and (iii) an extensive evaluation of adversarial robustness. Limitations include the monolingual dataset and restricted adversarial testing, which present opportunities for future research on multilingual datasets and real-time deployment. To further strengthen the evaluation, this study incorporates transformer-based multilingual models, such as mBERT and XLM-R, that leverage contextual embeddings to improve semantic understanding. Experimental results show that these models outperform traditional and deep learning approaches, achieving up to 99.4% accuracy and a 99.2% F1 score. Additionally, an extended adversarial analysis demonstrates that all models experience performance degradation under attack; transformer-based models maintain comparatively higher robustness to synonym substitution, MLP to noise injection and data poisoning, ensemble to typo injection, and finally RF to label flipping.
Bangla Sign Language (BdSL) serves as a critical communication bridge for children with hearing and speech impairments, enabling interpersonal communication, social inclusion, and cognitive development. Existing sign language recognition systems predominantly focus on alphabet and numeric classification, neglecting complete words essential for meaningful daily interactions. This study addresses this gap by developing a machine learning–based system specifically designed to recognize static Bangla Sign Language words frequently used by children with communication disabilities. We present a deep learning approach utilizing the VGG19 architecture to classify hand gestures corresponding to 16 commonly used Bangla words. Our custom-built dataset comprises 1703 training images and 720 testing images, carefully curated to represent vocabulary relevant to children’s daily communication needs. Through comprehensive preprocessing, data augmentation, and class balancing techniques, we achieved training accuracy of 96.59% and testing accuracy of 88.47%, demonstrating promising generalization on the collected test set. Comparative analysis with VGG16, traditional CNN, and support vector machine models confirms VGG19’s superior performance. This word-level recognition approach represents a meaningful step beyond character-based systems, enabling more natural communication for children with hearing and speech impairments.
