Deep learning-based credit scoring systems face three interrelated challenges typically addressed in isolation: unavoidable class imbalance, model opacity, and demographic inequalities. This paper proposes the Explainable Fairness-Aware Deep Learning (EFADL) framework, which is a unified end-to-end pipeline designed to mitigate these risks simultaneously. The EFADL framework integrates three novel components: FC-SMOTE, a fairness-constrained oversampling module that preserves intra-group demographic balance; a multi-objective joint training loss combining focal imbalance correction with differentiable multi-attribute fairness penalties; and a dual-level SHAP module providing both instance-level adverse action explanations and group-level fairness attribution. Extensive experiments on the German Credit, Taiwan Credit, and LendingClub datasets demonstrate that EFADL achieves a better accuracy-fairness trade-off surface. Results indicate a 79% reduction in statistical parity and equal opportunity differences and a 70.5% decrease in maximum intersectional disparity, with a negligible AUC-ROC cost of only 1.2 percentage points. Furthermore, the framework reduces the fairness attribution gap by 71%, which provides evidence that it achieves fairness by suppressing reliance on demographic proxies rather than post-hoc calibration. By delivering stable, economically interpretable explanations, the EFADL framework aligns with the transparency requirements of the EU AI Act and US CFPB guidance and offers a deployable solution for regulatorily-compliant algorithmic lending.

Internet of Things (IoT) and Internet of Medical Things (IoMT) networks provide a massive amount of data. These types of data need a protection level, such as an intrusion detection framework. Deep learning models become a powerful tool for this purpose. Therefore, this work proposes an intrusion detection framework based on a deep learning technique which employs TabTransformer and self-attention mechanisms to imprison intricate dependencies among tabular features and detect abnormal attack behaviors. Precisely, each numerical feature is mapped into a learnable embedding vector and augmented with positional encodings to preserve feature identity and inter-feature relationships within the embedding space. The main task for the proposed model is to achieve binary classification tasks the model should classify the traffic data as either normal or abnormal. Furthermore, the model utilized a benchmark dataset such as the CICIoMT2024. Furthermore, this type of dataset faces issues, such as imbalance. So, the system integrates SMOTE-based data balancing, Stratified K-Fold Cross-Validation, and threshold optimization to ensure fairness and reproducibility to accomplish a binary classification task. As a consequence, experiments on the CICIoMT2024 dataset yield superior results, achieving a mean accuracy of 99.85. Through SHAP-based interpretability, key features influencing model predictions are identified, confirming the framework’s transparency, robustness, and suitability for real-world ARP intrusion detection.

Credit card fraud poses a significant challenge for both consumers and organizations worldwide, particularly with the increasing reliance on credit cards for financial transactions. Therefore, it is crucial to establish effective mechanisms to detect credit card fraud. However, the uneven distribution of instances between the two classes in the credit card dataset hinders traditional machine learning techniques, as they tend to prioritize the majority class, leading to inaccurate fraud pre- dictions. To address this issue, this paper focuses on the use of the Elbow Fuzzy Noise Filtering SMOTE (EFN-SMOTE) technique, an oversampling approach, to handle unbalanced data. EFN-SMOTE partitions the dataset into multiple clusters using the Elbow method, applies noise filtering to each cluster, and then employs SMOTE to synthesize new minority instances based on the nearest majority instance to each minority instance, thereby improving the model’s ability to perceive the decision boundary. EFN-SMOTE’s performance was evaluated using an Artificial Neural Network model with four hidden layers, resulting in significant improvements in classification performance, achieving an accuracy of 0.999, precision of 0.998, sensitivity of 0.999, specificity of 0.998, F-measure of 0.999, and G-Mean of 0.999.

Early identification of individuals at high risk of stroke is essential for improving prognosis, guiding preventive strategies, and reducing the burden of stroke-related disability. This study presents a comparative machine learning approach for stroke risk prediction using structured healthcare data containing demographic, lifestyle, and clinical variables. The proposed approach evaluates five supervised classification models, namely Random Forest, Support Vector Machine, XGBoost, LightGBM, and CatBoost, to identify the most suitable model for stroke risk classification. A systematic preprocessing pipeline was applied, including median imputation for missing BMI values, categorical feature encoding, stratified data splitting, feature scaling, and Synthetic Minority Over-sampling Technique (SMOTE) to address severe class imbalance in the training data. Model performance was assessed using accuracy, precision, recall, and F1-score Among the evaluated models, CastBoost achieved the most balanced overall performance, while SVM showed higher sensitivity in detecting stroke cases. These findings indicate that model selection should depend on the intended clinical purpose, particularly whether the goal is balanced prediction or improved detection of possible stroke cases. Overall, this study highlights the potential of machine learning to support early stroke risk assessment and preventive healthcare decision-making using routinely available structured health data.