Introduction: Vision impairment caused by retinal diseases stands out as one of the significant sources of irreversible blindness, highlighting the importance of developing robust automated screening tools to detect eye diseases at an early stage. Despite the promising results of recent deep learning approaches to the automatic classification of retinal images, most existing solutions focus on image-based representations and do not consider the importance of vascular topology. To overcome the limitations, this study proposes VG-RETFound - a hybrid framework based on a retinal foundation model and graph-based vascular-topology modeling that can be used to classify multiple retinal categories and detect vision impairment. Methods: The authors use RETFound-a retinal foundation model-pretrained on a vast number of retinal images, as well as the Graph Attention Network (GAT), to learn vascular topology features by representing retinal vessels as graphs. The presented approach employs a cross-modal multi-head attention fusion method to integrate features from both modalities in real time. To evaluate the efficiency of VG-RETFound, experiments were conducted using the RFMiD and ODIR-5 K datasets for model training and testing across six retinal categories: Normal, Diabetic Retinopathy, Pathological Myopia, Age-related Macular Degeneration, Hypertensive Retinopathy, and Other Retinal Diseases. Results: The accuracy of 96.39%, F1-score of 95.84%, MCC of 94.87%, and Kappa of 94.71% were attained on an internal test dataset, while cross-dataset external validation on the BRSET dataset revealed accuracy of 94.81%, F1-score of 94.25%, MCC of 93.09%, and Kappa of 92.94%. Discussion: These findings demonstrate that integrating retinal foundation models with anatomically informed vessel-topology learning significantly improves classification accuracy, interpretability, and generalization, highlighting the potential of VG-RETFound as a tool for large-scale vision impairment. screening and retinal disease diagnosis.
Background The increased use of Machine Learning (ML) systems has posed critical Quality Assurance (QA) issues in areas of security, privacy and performance. Despite advances, QA frameworks for ML remain disjointed. The work examines QA through the Technological, Organizational, and Environmental (TOE) model and proposes human-agentic AI actions to enhance assurance mechanisms. This paper aims to identify, compare and model QA problems in ML software systems and recommend a SWOT-TOE (Strengths, Weaknesses, Opportunities, and Threats) matrix to enhance human-agentic AI-based QA problems. Methodology The study uses a mixed-method design, which includes systematic literature review (SLR), a quantitative survey, and comparative data modelling. QA issues are coded using the TOE structure and the association tests are linear-by-linear tests to investigate the relationship between the literature and survey results. A SWOT-TOE matrix and mathematical modelling are used to assess key QA issues and to solve them. Results The analysis shows that most of the QA problems revolve around interpretability of models, integrity of data, organizational readiness and regulatory limits. The linear-by-linear association results suggest that there is a strong correlation of the patterns of QA that are identified by the literature and those documented in the survey. Comparative data modeling shows the differences in priorities of QA between the regions. The suggested SWOT-TOE framework combines human-agentic solutions (e.g., adaptive auditing and explainable AI) to enhance QA and security. Conclusion The combination of the TOE framework and human-agentic AI practices and quantitative models allows a systematic approach to solving complex QA problems in ML systems. The research leads to the creation of human-centered, adaptive QA strategies, which foster trust, transparency, and sustainability in AI-powered settings. The framework assists developers, QA engineers, and policymakers to increase security, reliability, and accountability in ML ecosystems.
With IoT networks expected to exceed 29 billion connected devices by 2030, the risk of cyberattacks has never been higher. As more devices come online, the attack surface for hackers continues to expand, making cybersecurity a pressing concern. Intrusion Detection Systems (IDS) are essential for identifying and mitigating these threats in real-time. However, a significant challenge IDS faces is dealing with imbalanced datasets, where attack instances are significantly underrepresented compared to normal traffic. Training models on such skewed data leads to a bias toward majority-class patterns, reducing their ability to detect intrusions effectively. To address this issue, this work introduces CSMCR (Cosine Similarity-based Majority Class Reduction), a novel technique that selectively removes redundant majority-class samples while preserving dataset integrity. Unlike traditional approaches like SMOTE (oversampling) or random undersampling, CSMCR ensures that the retained majority instances remain diverse by analyzing feature-wise similarity. This prevents unnecessary data duplication and minimizes information loss. Additionally, we developed a hybrid deep learning model integrating RegNet and FBNet architectures to enhance feature extraction and classification performance. Experimental results on multiple IDS datasets confirm that balancing the dataset to a 1:1 ratio optimally prevents overfitting and improves model interpretability. The proposed model achieved an F1-score of 0.9758 on RT-IoT2022 and 0.9275 on UNSW Bot-IoT, outperforming SMOTE-based methods in accuracy and computational efficiency. Notably, CSMCR reduced training time by 53% compared to conventional oversampling techniques. Incremental training evaluations reveal that bias formation reduces performance beyond a 1:2 majority-to-minority ratio. These findings establish CSMCR as a robust, scalable, and computationally efficient IDS balancing strategy tailored for IoT network security.
Chronic obstructive pulmonary disease (COPD) is a progressive respiratory condition, contributing significantly to global morbidity and mortality. Traditional diagnostic tools are effective in diagnosing COPD. However, these tools demand specialized equipment and expertise. Advances in artificial intelligence (AI) provide a platform for enhancing COPD diagnosis by leveraging diverse data modalities. The existing reviews primarily focus on single modalities and lack information on interpretability and explainability. Thus, this review intends to synthesize the AI-powered frameworks for COPD identification, focusing on data modalities, methodological innovation, evaluation strategies, and reporting limitations and potential biases. By adhering to Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, a systematic search was conducted across multiple repositories. From an initial pool of 1978 records, 22 studies were included in this review. The included studies demonstrated exceptional performance in specific settings. Most studies were retrospective and limited in diversity, lacking generalizability and external or prospective validation. This review presents a roadmap for advancing AI-assisted COPD detection. By highlighting the strengths and limitations of existing studies, it supports the development of future research. Future studies can utilize the findings to build models using prospective, multicenter, and multi-ethnic validations, ensuring generalizability and fairness.
Speech disorders affect an individual's ability to generate sounds or utilize the voice appropriately. Neurological, developmental, physical, and trauma may cause speech disorders. Speech impairments influence communication, social interaction, education, and quality of life. Successful intervention entails early and precise diagnosis to allow for prompt treatment of these conditions. However, clinical examinations by speech-language pathologists are time-consuming, subjective, and demand an automated speech disorder detection (SDD) model. Mel-spectrogram images present a visual representation of multiple speech disorders. By classifying Mel-Spectrogram, various speech disorders can be identified. In this study, the authors proposed an image classification-based automated SDD model to classify Mel-Spectrograms to identify multiple speech disorders. Initially, Wavelet Transform (WT) hybridization technique was employed to generate Mel-Spectrogram using the voice samples. A feature extraction approach was developed using an enhanced LEVIT transformer. Finally, the extracted features were classified using an ensemble learning (EL) approach, containing CatBoost and XGBoost as base learners, and Extremely Randomized Tree as a meta learner. To reduce the computational resources, the authors used quantization-aware training (QAT). They employed Shapley Additive Explanations (SHAP) values to offer model interpretability. The proposed model was generalized using Voice ICar fEDerico II (VOICED) and LANNA datasets. The exceptional accuracy of 99.1 with limited parameters of 8.2 million demonstrated the significance of the proposed approach. The proposed model enhances speech disorder classification and offers novel prospects for building accessible, accurate, and efficient diagnostic tools. Researchers may integrate multimodal data to increase the model's use across languages and dialects, refining the proposed model for real-time clinical and telehealth deployment.
Agriculture sector is faced with perennial challenges that threaten both its productivity and sustainability. Among the greatest threats to cereal crops is disease, especially on cereal grains such as maize. Maize is an important grain that has grown globally, yet it often falls prey to maize leaf disease, a destructive and prevalent disorder. The consequences of these diseases go beyond individual farmers; reduced yields destabilize supply chains, market stability, and global efforts towards creating sustainable food systems. The prevalence of leaf diseases adversely affects crop productivity, which directly impacts the objective of sustainable agriculture. In order to address this problem, technology, more specifically artificial intelligence, has been a game-changer. Adopting cutting-edge research for maize disease detection not only raises diagnostic accuracy but also supports sustainable agricultural practices. This approach encourages effective input use, supports food security, minimizes environmental degradation, and provides farmers with accurate tools, all of which contribute to long-term agricultural resilience and sustainability. However, the incorporation of AI in farming is confronted by a number of challenges. Numerous hindrances hinder precise detection and categorization of maize leaf diseases via artificial intelligence methods. In tackling the limitations, the current study introduces an AI-based approach. It applied Multi-scaled Xception pre-trained models to extract deep features from images. The models were fine-tuned with varying weights for advancing the feature extraction so as to enhance the likelihood of correct visual classification. In addition to its strong accuracy, the research provides a formal and strict mathematical formulation, and different optimization methods further establish the effectiveness of the model. Additionally, the examination of fusion operators helps to improve the interpretability of the model. The expected model was tested with a confidence interval, showing that its performance remains within set limits.
Background As a crucial cognitive and motor function, handwriting can reveal dyslexia, dysgraphia, and other neurodevelopmental conditions. With the emergence of AI, deep learning (DL) models can analyze handwriting patterns for early identification and categorization of these conditions with high level of accuracy. Objective This study presents a literature review by examining DL models in classifying learning disabilities using handwriting data. It identifies models’ shortcomings and suggests future research directions, improving model accuracy, fairness, and usability. Methodology By employing the standard literature guidelines, the authors identified 24 studies from a pool of 502 studies. They extracted dataset details, methodologies, and key findings from the studies. Findings The findings indicate the dominance of convolutional neural networks in improving learning disabilities classification through the extraction of fine-grained spatial handwritten patterns. Vision transformers (ViTs) show potential for long-range contextual feature extraction to enhance the model’s generalization capabilities. Despite these advances, dataset size, model generalizability, computational efficiency, interpretability, and real-world deployment continue to be significant challenges. Conclusions The review findings underscore the transformative potential of DL-based handwriting analysis in educational and clinical settings, providing early detection and intervention in learning disabilities.
Down syndrome (DS) is a genetic condition characterized by distinct facial features and developmental challenges, making early and accurate detection crucial for timely intervention and support. Traditional diagnostic methods often require specialized expertise, which can be resource intensive and inaccessible in many settings. In this study, we propose a novel artificial intelligence (AI)-driven approach to detecting DS using facial images, leveraging advancements in deep learning and machine learning for accurate and scalable diagnostics. A dataset of 168 facial images was curated from publicly accessible platforms, comprising 95 images of individuals without DS and 73 with the condition. The model was implemented using TensorFlow/PyTorch and employs RegNet X and Shifted Window Transformer (SWIN Transformer) for feature extraction, capturing both local and global facial features. To enhance feature representation, an attention-based fusion mechanism was applied, prioritizing the most discriminative patterns while filtering redundant information. For classification, an Extremely Randomized Tree (ExtraTrees) model was employed, known for its ability to efficiently handle high-dimensional data with strong generalization capabilities. The model was fine-tuned using Bayesian Optimization and Hyperband (BOHB), which optimized hyperparameters by balancing exploration and exploitation, thereby improving predictive performance. The final classification achieved a 99.2% accuracy with an area under the curve (AUC) of 1.00, demonstrating exceptional reliability in distinguishing DS from normal cases. By combining advanced feature extraction, fusion, and classification techniques, this study highlights the potential of AI to revolutionize DS diagnostics, offering a non-invasive, efficient, and scalable solution for early detection.
Down syndromeface classificationearly diagnosismodel interpretability
Across the globe, chronic obstructive pulmonary disease (COPD) continues to be a critical healthcare burden due to its progressive and undetected decline in pulmonary function. In recent times, convolutional neural networks (CNNs) and vision transformers (ViTs) have been widely applied to lung sounds (LS) analysis, leading to the development of improved non-invasive, accessible, and scalable COPD screening methods. However, the shortcomings of CNNs and ViTs architectures restrict the COPD classifier’s discriminative power in distinguishing normal and COPD samples. In this study, a unique and interpretable COPD classification is proposed. This study introduces a hybrid feature extraction strategy integrating CNNs and ViTs architecture to capture the crucial COPD patterns from the LS’s Mel-spectrogram representation. The proposed CNNs-ViTs architecture addresses the limitations of standalone CNNs or ViTs architectures by capturing fine-grained acoustic features and long-range temporal dependencies while efficiently modeling non-linear relationships. A novel feature fusion is used to combine the heterogeneous features into a unified and highly discriminative feature representation space. A tailored hyperparameter optimization strategy is developed to fine-tune the Kolmogorov-Arnold Networks (KANs) for classifying the features into normal and COPD classes. By incorporating internal (552 Healthy and 1324 COPD LS recordings) and external (35 Healthy and 77 COPD LS recordings) datasets, the authors evaluate the model’s performance. The model achieves a generalization accuracy of 97.62% and specificity of 96.51% on the external dataset. Through Shapley Additive Explanations (SHAP) overlays, the model’s interpretability is enhanced. The experimental findings highlight the effectiveness of the proposed COPD classification architecture in the domain of biomedical acoustic signal processing. Operating with minimal reliance on specialized diagnostic infrastructure renders the proposed model ideal for primary care and resource-limited environments.
Autism spectrum disorder (ASD) is a diverse neurodevelopmental condition with a significant genetic basis. Machine learning (ML) and deep learning (DL) models are emerging as powerful tools for identifying potential ASD risk genes and biomarkers to improve diagnostic accuracy. This narrative systematic review synthesizes evidence on the use of ML and DL models for ASD prediction and diagnosis, focusing on studies that utilized genetic variants and biomarkers. We conducted a systematic search across six databases [Web of Science, PubMed, Science Direct, Semantic Scholar, Institute of Electrical and Electronics Engineers (IEEE Xplore), and Association for Computing Machinery Digital Library (ACM-DL)] for articles published between 2009 and 2024, yielding 483 initial records. Multiple inclusion criteria were applied, and only peer-reviewed articles that used artificial intelligence, ML, or DL to predict ASD based on genetic biomarkers were included. After deduplication and applying inclusion and exclusion criteria, 14 studies were included in this review. These studies focused on genetic variants based on whole-genome sequencing, whole-exome sequencing, RNA sequencing, single-nucleus RNA sequencing, mRNA expression, long non-coding RNA, or single nucleotide polymorphism data. Our findings indicate that various ML models achieved robust performance, while DL architectures such as DeepASDPred outperformed ML models, with an accuracy of 93.8%. Most of the studies used datasets such as the Simons Foundation Autism Research Initiative (SFARI), the Simons Simplex Collection (SSC), the Autism Genetic Resource Exchange (AGRE), Atlas of the Developing Human Brain (BrainSpan), and the Simons Foundation Powering Autism Research for Knowledge (SPARK). The reviewed studies demonstrate that ML can help identify genetic risk factors and biomarkers that contribute to ASD. However, challenges remain, including data heterogeneity, small sample sizes, and the need for greater model interpretability. Addressing these challenges is crucial for translating ML models into clinically useful tools. In conclusion, DL substantially enhances ASD prediction, leveraging genetic data to identify novel candidate genes and support biological discovery of cell type-specific gene expression and roles of non-coding mutations in ASD.
Hematological biomarkers have emerged as powerful tools in diagnosing Acute Heart Failure (AHF). This study introduces a novel diagnostic framework that integrates Explainable Artificial Intelligence (XAI) with Morris Sensitivity Analysis (MSA) to enhance both the interpretability and performance of machine learning models in AHF detection. A dataset consisting of 425 AHF patients and 430 controls was analyzed using eight machine learning models, including XGBoost, Histogram-based Gradient Boosting (histGB), Explainable Boosting Machine (EBM), and Random Forest. Model performance was evaluated through metrics such as AUC, accuracy, precision, recall, and Brier score. Hyperparameters were optimized via Bayesian optimization. Feature importance was assessed using MSA to identify variables with the highest predictive influence. The histGB model achieved the highest performance with an AUC of 87.93%. Both MSA and EBM consistently identified PDW, RDW-CV, NEU, NEU/LY ratio, age, and WBC as top predictive features across multiple models. These hematological markers demonstrated strong potential for early diagnosis and risk stratification in AHF patients. This study presents a clinically relevant, interpretable, and cost-effective diagnostic strategy that combines XAI with MSA for AHF prediction. The framework enhances clinical trust and provides a pathway toward personalized treatment by identifying accessible hematological biomarkers. The integration of explainability into AI models improves their transparency and applicability in real-world clinical settings.
A robust, reproducible pipeline to (i) generate realistic synthetic RNA-seq data for visual-impairment research using a Negative Binomial model, (ii) build and train a 1D-CNN classifier, (iii) evaluate performance via ROC-AUC and SHAP interpretability, and (iv) publish both code and datasets to public repositories (Protocols.io & NCBI GEO).
visual impairmentnegative binomial regressionSHAParea under the receiver operating characteristic
Vaccine acceptance is a crucial component of a viable immunization program in healthcare system, yet the disparities in new and existing vaccination adoption rates prevail across regions. Disparities in the rate of vaccine acceptance result in low immunization coverage and slow uptake of newly introduced vaccines. This research presents an innovative AI-driven predictive model, designed to accurately forecast vaccine acceptance within immunization programs, while providing high interpretability. Primarily, the contribution of this study is to classify vaccine acceptability into Low, Medium, Partial High, and High categories. Secondly, this study implements the Feature Importance method to make the model highly interpretable for healthcare providers. Thirdly, our findings highlight the impact of demographic and socio-demographic factors on vaccine acceptance, providing valuable insights for policymakers to improve immunization rates. A sample dataset containing 7150 data records with 31 demographic and socioeconomic attributes from PDHS (2017-2018) is used in this paper. Using the LightGBM algorithm, the proposed model constructed on the basis of different machine-learning procedures achieved 98% accuracy to accurately predict the acceptability of vaccines included in the immunization program. The association rules suggest that higher SES, region, parents' occupation, and mother's education have an association with vaccine acceptability.
IEEE Transactions on Computational Social Systems · مجلد 12 (3) · 1307-1318
The interactive attention graph convolution network (IAGCN), a novel model proposed in this article, will revolutionize aspect-level sentiment analysis (SA). IAGCN effectively addresses these key features, in contrast to prior research that ignored the meaning of aspect terms and their relationship with context. The model combines a modified dynamic weighting layer with bidirectional long short-term memory (BiLSTM) to accurately acquire context. It takes use of graph convolutional networks (GCNs) to encrypt syntactic information from the syntactic dependency tree. Furthermore, a method for interactive attention is employed to discover the intricate relationships between context and aspect terms, which results in the reconstruction of those terms’ representations. Comparing the proposed IAGCN model to baseline models, impressive gains are made. Across five datasets, the model beats previous methods with an amazing improvement in F1 scores that ranges from 1.34% to 4.04% and an impressive improvement in accuracy that ranges from 0.56% to 1.75%. Additionally, the IAGCN model outperforms the global vectors (GloVe)-based strategy when the potent pretrained model bidirectional encoder representations from transformers (BERT) is included in the challenge, resulting in even greater improvements. The F1 score considerably increases from 2.59% to 7.55%, and accuracy increases from 1.47% to 3.95%, making the IAGCN model a standout performer in aspect-level SA.