Boulbaba Ben Ammar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Boulbaba Ben Ammar has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
Mohamed Ben Aouicha
3 shared publications- Zainab Saad Rubaidi
College of Agriculture, Al-Muthanna University, Samawah 66001, Iraq
3 shared publications - Mohamed Ben Aouicha
University of Sfax
2 shared publications Mohamed Tahar Bhiri
2 shared publications- Walid Karamti
Qassim University
2 shared publications Zainab Saad Rubaidi
2 shared publications
Latest Publications
- Open Access
- 2 November 2026
- Published by MDPI AG in Diagnostics
- Vol. 16 (21) , 3622
Toward Clinically Trustworthy Deep Learning in Cardiology: A Unified Mathematical Synthesis of Representation, Optimization, and Reliability
Background: Deep learning (DL) has revolutionized cardiovascular medicine by allowing for automated analysis of electrocardiography, imaging of the heart, and multimodal clinical datasets. Although there have been impressive advancements in terms of predictive performance, the deployment of DL methods in a clinical setting remains limited due to lack of sufficient mathematical understanding around the concepts of uncertainty, robustness, fairness, explainability, and generalizability. Methods: In this work, a systematic review is conducted using PRISMA principles on 79 peer-reviewed articles published from 2017 until 2025, with a unifying mathematical perspective on the notion of trustworthy deep learning in cardiology. Results: We find that there is significant discrepancy between prediction accuracy and clinical reliability, with little emphasis placed on mathematical guarantees, reproducibility of evaluation procedures, calibration, or robustness verification. To address this challenge, we propose a conceptual mathematical framework which brings together the elements of representation learning, optimization, and trustworthiness into a unified audit of AI in cardiac care. We note that the presented indices are purely theoretical and are meant to provide guidance for future standardization, and as such do not represent actual clinical metrics. Conclusions: Although our paper does not propose any new predictive algorithms, we do contribute a strong theoretical basis for the evaluation and development of clinically reliable deep learning algorithms in cardiology.
- 2 November 2026
- Published by MDPI AG in Diagnostics
- Vol. 16 (21) , 3622
Open AccessToward Clinically Trustworthy Deep Learning in Cardiology: A Unified Mathematical Synthesis of Representation, Optimization, and Reliability
Background: Deep learning (DL) has revolutionized cardiovascular medicine by allowing for automated analysis of electrocardiography, imaging of the heart, and multimodal clinical datasets. Although there have been impressive advancements in terms of predictive performance, the deployment of DL methods in a clinical setting remains limited due to lack of sufficient mathematical understanding around the concepts of uncertainty, robustness, fairness, explainability, and generalizability. Methods: In this work, a systematic review is conducted using PRISMA principles on 79 peer-reviewed articles published from 2017 until 2025, with a unifying mathematical perspective on the notion of trustworthy deep learning in cardiology. Results: We find that there is significant discrepancy between prediction accuracy and clinical reliability, with little emphasis placed on mathematical guarantees, reproducibility of evaluation procedures, calibration, or robustness verification. To address this challenge, we propose a conceptual mathematical framework which brings together the elements of representation learning, optimization, and trustworthiness into a unified audit of AI in cardiac care. We note that the presented indices are purely theoretical and are meant to provide guidance for future standardization, and as such do not represent actual clinical metrics. Conclusions: Although our paper does not propose any new predictive algorithms, we do contribute a strong theoretical basis for the evaluation and development of clinically reliable deep learning algorithms in cardiology.
- Open Access
- 30 June 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1881829
Label leakage unmasked: a trustworthy-AI audit of autism screening models using the CLEAR-RD framework
Trustworthy pediatric screening AI needs more than accuracy: label validity, calibration, subgroup robustness, and clinical-scope limits. We propose Clear-Rd ( Clinical Leakage Evaluation and Audit Routine for Rule-Derived labels ), a five-stage framework—Configure, Lift, Evaluate, Audit, Report—and apply it to two public autism-screening datasets—Q-CHAT-10 toddler ( N = 1054) and AQ-10 child/adolescent/adult ( N = 6075)—merged into a unified cohort ( N = 7129); prior studies report accuracies above 95%. Three configurations— S 1 (full, with score), S 2 (items, no score), and S 3 (demographics only)—separate label leakage from demographic signal. A leakage audit confirms the Q-CHAT-10 label is exactly determined by score>3, the AQ-10 label approximately by score≥6. Sixteen models from three families (classical, deep-tabular, and prompt-structure simulators) are evaluated under 5-fold stratified cross-validation. On the leakage-free S 3 , Gradient Boosting attains ROC-AUC 0.766 ± 0.005 and all families converge within 0.04 ROC-AUC, indicating the binding constraint is the feature set, not model capacity. Calibration, subgroup-parity, and cost-sensitive analyses complete the audit. The results expose a gap between reported and clinically interpretable performance; Clear-Rd is a reusable template for rule-derived medical labels. No such model should be interpreted as predicting clinical autism diagnosis without external validation against gold-standard outcomes.
- 30 June 2026
- Published by Frontiers Media SA in Frontiers in Public Health
- Vol. 14, 1881829
Open AccessLabel leakage unmasked: a trustworthy-AI audit of autism screening models using the CLEAR-RD framework
Trustworthy pediatric screening AI needs more than accuracy: label validity, calibration, subgroup robustness, and clinical-scope limits. We propose Clear-Rd ( Clinical Leakage Evaluation and Audit Routine for Rule-Derived labels ), a five-stage framework—Configure, Lift, Evaluate, Audit, Report—and apply it to two public autism-screening datasets—Q-CHAT-10 toddler ( N = 1054) and AQ-10 child/adolescent/adult ( N = 6075)—merged into a unified cohort ( N = 7129); prior studies report accuracies above 95%. Three configurations— S 1 (full, with score), S 2 (items, no score), and S 3 (demographics only)—separate label leakage from demographic signal. A leakage audit confirms the Q-CHAT-10 label is exactly determined by score>3, the AQ-10 label approximately by score≥6. Sixteen models from three families (classical, deep-tabular, and prompt-structure simulators) are evaluated under 5-fold stratified cross-validation. On the leakage-free S 3 , Gradient Boosting attains ROC-AUC 0.766 ± 0.005 and all families converge within 0.04 ROC-AUC, indicating the binding constraint is the feature set, not model capacity. Calibration, subgroup-parity, and cost-sensitive analyses complete the audit. The results expose a gap between reported and clinically interpretable performance; Clear-Rd is a reusable template for rule-derived medical labels. No such model should be interpreted as predicting clinical autism diagnosis without external validation against gold-standard outcomes.
- Open Access
- 27 April 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (9) , 1467
Integrating Sequential Hybrid Oversampling with Decision-Theoretic Threshold Design for Credit Risk Assessment
Credit risk assessment under severe class imbalance requires both structured imbalance correction and principled decision rules, yet most studies treat these as independent steps. This study develops a general integrated three-layer framework for credit risk assessment under class imbalance. The first layer introduces Sequential Hybrid Data Oversampling (SHDO), which sequentially applies five complementary oversampling techniques to enrich minority-class representation in mixed-type credit data. The second layer formulates credit approval as a decision-theoretic optimisation problem: a closed-form optimal threshold is derived under asymmetric costs, extended to constrained portfolios via a Lagrangian formulation with Karush–Kuhn–Tucker conditions, and further extended to minimax-robust decision making under estimation uncertainty. The third layer compares eleven classifiers under a unified evaluation protocol with an ablation isolating the effect of SHDO. The framework is empirically validated on the Home Credit Default Risk dataset, which is used as an illustrative case study rather than defining the scope of the contribution. On the held-out test set, XGBoost trained with SHDO achieves the highest minority-class F1 (0.254), while gradient-boosted models collectively attain ROC-AUC values of 0.713–0.750, outperforming classical baselines (0.540–0.620). The ablation confirms that without SHDO, all models exhibit near-zero minority-class recall despite adequate ranking ability. SHAP analysis on XGBoost confirms that the learned risk structure aligns with established creditworthiness determinants. The decision framework converts these probability estimates into analytically justified approval thresholds responsive to economic parameters, institutional constraints, and estimation uncertainty.
- 27 April 2026
- Published by MDPI AG in Mathematics
- Vol. 14 (9) , 1467
Open AccessIntegrating Sequential Hybrid Oversampling with Decision-Theoretic Threshold Design for Credit Risk Assessment
Credit risk assessment under severe class imbalance requires both structured imbalance correction and principled decision rules, yet most studies treat these as independent steps. This study develops a general integrated three-layer framework for credit risk assessment under class imbalance. The first layer introduces Sequential Hybrid Data Oversampling (SHDO), which sequentially applies five complementary oversampling techniques to enrich minority-class representation in mixed-type credit data. The second layer formulates credit approval as a decision-theoretic optimisation problem: a closed-form optimal threshold is derived under asymmetric costs, extended to constrained portfolios via a Lagrangian formulation with Karush–Kuhn–Tucker conditions, and further extended to minimax-robust decision making under estimation uncertainty. The third layer compares eleven classifiers under a unified evaluation protocol with an ablation isolating the effect of SHDO. The framework is empirically validated on the Home Credit Default Risk dataset, which is used as an illustrative case study rather than defining the scope of the contribution. On the held-out test set, XGBoost trained with SHDO achieves the highest minority-class F1 (0.254), while gradient-boosted models collectively attain ROC-AUC values of 0.713–0.750, outperforming classical baselines (0.540–0.620). The ablation confirms that without SHDO, all models exhibit near-zero minority-class recall despite adequate ranking ability. SHAP analysis on XGBoost confirms that the learned risk structure aligns with established creditworthiness determinants. The decision framework converts these probability estimates into analytically justified approval thresholds responsive to economic parameters, institutional constraints, and estimation uncertainty.
- 16 March 2026
- Published by Taylor & Francis
- p. 215-234
Developing SHDO
Data augmentation is a frequently used technique to address class imbalance in classification tasks such as fraud detection. This is enhancing machine learning (ML) model accuracy by generating new, diverse examples. This study introduces Sequential Hybrid Data Oversampling (SHDO), a novel approach that integrates several advanced oversampling methods, including SMOTE, Random Oversampling, SVMSMOTE, Borderline SMOTE, and SMOTEN. Our systematic approach involves addressing a problem in a step-by-step manner, progressing linearly from one stage to the next. Each step builds on the previous one. SHDO effectively mitigates bias toward the majority class by creating synthetic minority class samples, reducing overfitting associated with traditional oversampling. By integrating techniques like SMOTE and SVMSMOTE, SHDO adapts to different data complexities, while Borderline SMOTE improves boundary handling between classes. SMOTEN is employed to manage categorical data, improving recall and detecting rare examples. SHDO was assessed on three datasets, including vehicle insurance fraud, synthetic, and predictive maintenance, by using a 75% training and 25% testing split. The vehicle insurance fraud dataset was specifically used to evaluate SHDO’s effectiveness in fraud detection scenarios. The results demonstrate that the proposed SHDO approach outperforms other oversampling techniques, particularly in terms of F1-score, with SVMSMOTE ranking second in performance. Both Random Forest and XGBoost classifiers showed strong performance across datasets. However, the results and performance of oversampling methods and classifiers varied based on the dataset and evaluation metrics. This study contributes to the literature on data augmentation techniques, offering valuable insights for improving classification performance in imbalanced binary classification scenarios.
- 16 March 2026
- Published by Taylor & Francis
- p. 215-234
Developing SHDO
Data augmentation is a frequently used technique to address class imbalance in classification tasks such as fraud detection. This is enhancing machine learning (ML) model accuracy by generating new, diverse examples. This study introduces Sequential Hybrid Data Oversampling (SHDO), a novel approach that integrates several advanced oversampling methods, including SMOTE, Random Oversampling, SVMSMOTE, Borderline SMOTE, and SMOTEN. Our systematic approach involves addressing a problem in a step-by-step manner, progressing linearly from one stage to the next. Each step builds on the previous one. SHDO effectively mitigates bias toward the majority class by creating synthetic minority class samples, reducing overfitting associated with traditional oversampling. By integrating techniques like SMOTE and SVMSMOTE, SHDO adapts to different data complexities, while Borderline SMOTE improves boundary handling between classes. SMOTEN is employed to manage categorical data, improving recall and detecting rare examples. SHDO was assessed on three datasets, including vehicle insurance fraud, synthetic, and predictive maintenance, by using a 75% training and 25% testing split. The vehicle insurance fraud dataset was specifically used to evaluate SHDO’s effectiveness in fraud detection scenarios. The results demonstrate that the proposed SHDO approach outperforms other oversampling techniques, particularly in terms of F1-score, with SVMSMOTE ranking second in performance. Both Random Forest and XGBoost classifiers showed strong performance across datasets. However, the results and performance of oversampling methods and classifiers varied based on the dataset and evaluation metrics. This study contributes to the literature on data augmentation techniques, offering valuable insights for improving classification performance in imbalanced binary classification scenarios.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 142262-142276
Calibrated RoBERTa With Targeted Augmentation for Financial Risk-Tone Detection: A Decision- Theoretic Intelligent System for Short-Form Financial Text Monitoring
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 142262-142276
Open AccessCalibrated RoBERTa With Targeted Augmentation for Financial Risk-Tone Detection: A Decision- Theoretic Intelligent System for Short-Form Financial Text Monitoring
- Open Access
- 8 December 2025
- Published by Engineering, Technology & Applied Science Research in Engineering, Technology & Applied Science Research
- Vol. 15 (6) , 29872-29879
Classification of Requirements from Software Requirement Specification Documents
Accurate classification of software requirements is a crucial task in Software Engineering (SE) that prioritizes development efforts and ensures the holistic quality of the system, encompassing both Functional Requirements (FRs) and Non-Functional Requirements (NFRs). While the majority of requirements classification research has so far focused on binary classification, fine-grained multi-class classification still encounters the challenge of extreme class imbalance in requirements datasets. To mitigate this limitation, we present TextAttack Oversampling (TAOS), a novel method that utilizes a Natural Language Processing (NLP)-based text augmentation technique to address this imbalance, thus reducing dependence on expensive expert labeling. An empirical assessment of a 12-class requirements dataset indicates that TAOS considerably outperforms standard classification techniques. Our approach achieves a 24% gain in F1-score, increasing from 0.75 to 0.93, and effectively improves the performance on minority requirement classes that were previously undetectable. This study demonstrates the effectiveness of context-aware text augmentation in improving multi-class requirements classification, thus providing a proven method to enhance the reliability and usefulness of automated requirements analysis and management tools for software engineers.
- 8 December 2025
- Published by Engineering, Technology & Applied Science Research in Engineering, Technology & Applied Science Research
- Vol. 15 (6) , 29872-29879
Open AccessClassification of Requirements from Software Requirement Specification Documents
Accurate classification of software requirements is a crucial task in Software Engineering (SE) that prioritizes development efforts and ensures the holistic quality of the system, encompassing both Functional Requirements (FRs) and Non-Functional Requirements (NFRs). While the majority of requirements classification research has so far focused on binary classification, fine-grained multi-class classification still encounters the challenge of extreme class imbalance in requirements datasets. To mitigate this limitation, we present TextAttack Oversampling (TAOS), a novel method that utilizes a Natural Language Processing (NLP)-based text augmentation technique to address this imbalance, thus reducing dependence on expensive expert labeling. An empirical assessment of a 12-class requirements dataset indicates that TAOS considerably outperforms standard classification techniques. Our approach achieves a 24% gain in F1-score, increasing from 0.75 to 0.93, and effectively improves the performance on minority requirement classes that were previously undetectable. This study demonstrates the effectiveness of context-aware text augmentation in improving multi-class requirements classification, thus providing a proven method to enhance the reliability and usefulness of automated requirements analysis and management tools for software engineers.
- Open Access
- 6 October 2025
- Published by Engineering, Technology & Applied Science Research in Engineering, Technology & Applied Science Research
- Vol. 15 (5) , 27852-27857
SQL Injection Detection Using Fine-Tuned CodeBERT
SQL injection attacks continue to pose a serious threat to web application security, with millions of systems vulnerable worldwide despite ongoing research and mitigation efforts. This study addresses the limitations of current SQLi detection techniques, particularly their inability to adapt to evolving attack patterns and their tendency to high false positive rates. A systematic review of machine learning- and deep learning-based SQL injection detection studies, published between 2017 and 2023, examined performance metrics, methods, and common shortcomings. Building on these insights, this study proposes a novel transformer-based approach utilizing a fine-tuned CodeBERT model trained on 30,919 SQL queries. This method includes extensive preprocessing and rigorous evaluation to ensure robustness and applicability. The results show that while traditional machine learning approaches reach between 73.5% and 99% accuracy, and deep learning models achieve 85% to 98%, the proposed CodeBERT-based system significantly outperforms them, attaining 99.90% accuracy, 99.96% precision, 97.75% recall, and 99.86% F1-score. These findings underscore the effectiveness of transformer models trained on code for SQL injection detection, setting new benchmarks and offering a deployable solution to enhance web application security.
- 6 October 2025
- Published by Engineering, Technology & Applied Science Research in Engineering, Technology & Applied Science Research
- Vol. 15 (5) , 27852-27857
Open AccessSQL Injection Detection Using Fine-Tuned CodeBERT
SQL injection attacks continue to pose a serious threat to web application security, with millions of systems vulnerable worldwide despite ongoing research and mitigation efforts. This study addresses the limitations of current SQLi detection techniques, particularly their inability to adapt to evolving attack patterns and their tendency to high false positive rates. A systematic review of machine learning- and deep learning-based SQL injection detection studies, published between 2017 and 2023, examined performance metrics, methods, and common shortcomings. Building on these insights, this study proposes a novel transformer-based approach utilizing a fine-tuned CodeBERT model trained on 30,919 SQL queries. This method includes extensive preprocessing and rigorous evaluation to ensure robustness and applicability. The results show that while traditional machine learning approaches reach between 73.5% and 99% accuracy, and deep learning models achieve 85% to 98%, the proposed CodeBERT-based system significantly outperforms them, attaining 99.90% accuracy, 99.96% precision, 97.75% recall, and 99.86% F1-score. These findings underscore the effectiveness of transformer models trained on code for SQL injection detection, setting new benchmarks and offering a deployable solution to enhance web application security.
- Open Access
- 23 August 2025
- Published by CV Media Inti Teknologi in International Journal of Advances in Artificial Intelligence and Machine Learning
- Vol. 2 (3) , 160-169
Enhancing SQL Code Security and Maintainability: A Deep Learning Based Approach
Background of study: SQL injection attacks continue to pose a significant risk to online systems. Traditional rule-based detection regularly fails to identify emerging or disguised attack vectors. Deep learning holds significant promise for robust detection, yet few studies rigorously compare model types or examine how to convey detection results as actionable security advice for developers.Aims and scope of paper: Building on this gap in existing research, this study tests three deep learning models for detecting SQL injection: Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and DistilBERT. The best model is then utilized in a tool that provides developers with risk assessments, warnings about unsafe patterns, and examples of secure queries.Methods: To achieve this, a dataset of 30,919 labeled SQL queries was preprocessed using normalization, syntax validation, and stratified splitting (70/15/15). A dual tokenization approach enabled fair comparisons between architectures. Models were trained using Adam/AdamW optimizers and evaluated for accuracy, precision, recall, F1-score, AUC-ROC, and MCC.Result: Among the tested models, DistilBERT set the performance benchmark, achieving 99.8% accuracy, 99.9% precision, 99.5% recall, and a false positive rate of just 0.1%. CNN and BiLSTM showed strong results, but proved weaker against obfuscated or distributed attacks. The SQL Security Advisor system converts model predictions directly into actionable guidance for developers.Conclusion: In conclusion, our findings indicate that DistilBERT detects SQL injections more effectively than CNN and BiLSTM, particularly when attacks are complex or hidden. By combining detection, explanation, and repair, this approach helps bring research closer to real-world use and supports developers in building more secure systems.
- 23 August 2025
- Published by CV Media Inti Teknologi in International Journal of Advances in Artificial Intelligence and Machine Learning
- Vol. 2 (3) , 160-169
Open AccessEnhancing SQL Code Security and Maintainability: A Deep Learning Based Approach
Background of study: SQL injection attacks continue to pose a significant risk to online systems. Traditional rule-based detection regularly fails to identify emerging or disguised attack vectors. Deep learning holds significant promise for robust detection, yet few studies rigorously compare model types or examine how to convey detection results as actionable security advice for developers.Aims and scope of paper: Building on this gap in existing research, this study tests three deep learning models for detecting SQL injection: Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and DistilBERT. The best model is then utilized in a tool that provides developers with risk assessments, warnings about unsafe patterns, and examples of secure queries.Methods: To achieve this, a dataset of 30,919 labeled SQL queries was preprocessed using normalization, syntax validation, and stratified splitting (70/15/15). A dual tokenization approach enabled fair comparisons between architectures. Models were trained using Adam/AdamW optimizers and evaluated for accuracy, precision, recall, F1-score, AUC-ROC, and MCC.Result: Among the tested models, DistilBERT set the performance benchmark, achieving 99.8% accuracy, 99.9% precision, 99.5% recall, and a false positive rate of just 0.1%. CNN and BiLSTM showed strong results, but proved weaker against obfuscated or distributed attacks. The SQL Security Advisor system converts model predictions directly into actionable guidance for developers.Conclusion: In conclusion, our findings indicate that DistilBERT detects SQL injections more effectively than CNN and BiLSTM, particularly when attacks are complex or hidden. By combining detection, explanation, and repair, this approach helps bring research closer to real-world use and supports developers in building more secure systems.
- 15 July 2025
- Published by Springer Nature
- p. 199-209
Handling Imbalance Functional and Non-Functional Software Requirement Classification Based on Machine Learning Algorithms
- 15 July 2025
- Published by Springer Nature
- p. 199-209
Handling Imbalance Functional and Non-Functional Software Requirement Classification Based on Machine Learning Algorithms
- 6 January 2024
- Published by Springer Nature in SN Computer Science
- Vol. 5 (1) , 1-10
Machine Learning Models for Early Prediction of COVID-19 Infections Based on Clinical Signs
- 6 January 2024
- Published by Springer Nature in SN Computer Science
- Vol. 5 (1) , 1-10
Machine Learning Models for Early Prediction of COVID-19 Infections Based on Clinical Signs