ABSTRACT Parkinson's disease (PD) increasingly relies on non‐invasive digital biomarkers and artificial intelligence (AI) methods for early diagnosis, symptom monitoring, and disease management. However, the growing use of complex machine learning and deep learning models introduces challenges related to model opacity, limiting clinical interpretability and potentially hindering adoption in healthcare settings. Explainable artificial intelligence (XAI) has emerged as a potential approach for improving transparency and supporting understanding of model behavior. This scoping review examined the application of XAI techniques in AI systems based on low‐cost, non‐invasive digital biomarkers for PD. Following PRISMA‐ScR guidelines, literature searches were performed in Scopus, PubMed, and Web of Science, identifying 36 eligible studies published between 2021 and 2025. Extracted data included clinical application, data modality, AI models, XAI approaches, explanation characteristics, explainability application, and explanation evaluation strategies. The reviewed studies showed widespread adoption of post hoc attribution‐based methods, particularly SHAP and LIME, across applications including diagnosis, symptom assessment, disease progression monitoring, and treatment‐state identification. XAI was commonly employed for feature ranking, biomarker exploration, model optimization, and explanatory analysis. However, explainability evaluation was highly heterogeneous, with most studies relying primarily on descriptive interpretation or plausibility assessments. Formal evaluation of explanation fidelity, clinician‐centered usability studies, and prospective clinical validation were largely absent. Current evidence suggests that XAI in PD research provides methodological tools that may support transparency and model understanding, but evidence for clinical trust, usability, and real‐world deployment remains limited. The field appears methodologically active but still immature for routine clinical implementation.
ABSTRACT The widespread adoption of Android and the Internet of Things (IoT) devices has led to a substantial increase in malware threats, challenging the effectiveness of traditional signature‐based security mechanisms. The dynamic and evolving nature of modern malware requires intelligent and adaptive detection techniques, making machine learning (ML) and deep learning (DL) approaches increasingly important. This paper presents a systematic survey of ML and DL‐based malware detection techniques, analyzing existing studies based on learning models, feature selection, datasets, and evaluation metrics. Conventional ML algorithms and advanced DL architectures including CNNs, RNNs, autoencoders, and hybrid models are critically compared. Our analysis confirms that DL and hybrid approaches outperform traditional ML methods for complex threats, though their accuracy and F1‐score depend heavily on dataset quality, feature engineering, and tuning. We synthesize widely used datasets and evolving malware families. However, despite strong experimental results, significant challenges remain, including dataset imbalance, IoT resource constraints, limited interpretability, and generalization gaps that hinder real‐world deployment. This study concludes that future work should prioritize developing lightweight, adaptive, and explainable detection frameworks that are specifically designed to operate under the operational constraints of Android and IoT platforms.
ABSTRACT Recent studies have shown that voice analysis provides a feasible, non‐invasive diagnostic alternative for the early detection of neurodegenerative diseases (NDs), such as Parkinson's disease (PD) and Amyotrophic Lateral Sclerosis (ALS), using Artificial Intelligence (AI), primarily machine learning models (MLs) that can be integrated into complex clinical decision support systems (CDSSs). Although significant progress has been made, several computational challenges remain, including dataset variability, noisy and incomplete input data, and the need for robust generalization across linguistic and demographic differences. To overcome these issues, system and model reliability must be improved through extensive testing under realistic conditions and integration with clinical workflows. This comprehensive study reviews the potential of AI to analyze speech biomarkers for accurate early detection and diagnosis. Improvements over traditional methods have been achieved through the use of Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Ensemble Methods (EMs), and diverse hybrid models. However, adoption of these models in real clinical practice is limited by challenges in standardization and effective real‐world applicability. Effective adoption of highly efficient AI‐based voice analysis in clinical practice requires standardized methods, collaboration among specialists from different disciplines, and evidence from real clinical trials. Physicians require comprehensive explanations of AI models to develop confidence in AI‐powered diagnostic tools. Explainable Artificial Intelligence (XAI) applications for neurodegenerative disease diagnosis remain underexplored in research, despite the need for healthcare practitioners to understand diagnostic predictions. Our study provides insights for researchers, software developers, and clinical practitioners, highlighting the need for future research on the development of robust, adaptive models that can accommodate data errors and remain fair and private across AI‐based diagnostics. Future research must build transparent AI models able to achieve high accuracy alongside interpretability standards for standard medical practice deployment. The review also systematically evaluates the quality of the data underlying the models that are reported, including the adequacy of the sample size, class balance, linguistic diversity, and recording conditions. This article is categorized under: Algorithmic Development > Biological Data Mining Application Areas > Health Care Technologies > Computational Intelligence
ABSTRACT Interactive new media arts (INMA) are technologically oriented artistic practices that share with knowledge management (KM) a concern for managing data, information, and knowledge (DIK). Despite this overlap, KM has largely overlooked INMA's practices—a gap compounded by the absence of a KM framework addressing INMA's contributions to KM in the public domain. Drawing on three interconnected theories, this review proposes a KM framework that examines how INMA applications' organizational interactions manage DIK across generative, participatory, and integrative management contexts. The framework is structural rather than normative: it identifies the operational conditions under which INMA manages DIK, not whether those outcomes are epistemically superior. A PRISMA‐guided analysis of 239 INMA studies (2000–2024) examined these contexts to characterize how INMA broadens KM in the public domain. The analysis shows that INMA's organizational interactions create dialogues between explicit and tacit DIK in the generative context and amplify them through expanding interactions in the participatory and integrative contexts. These operations consolidate into three aligned triads: tiers of interaction (passive, retroactive, and active), knowledge outcomes (receptive, dialogic, and co‐constitutive), and DIK relationship models (separate, interconnected, and unified). Synthesizing across the models yields a paradigm in which the active tier corresponds to the unified model to produce co‐constitutive knowledge—the configuration most capable of extending KM into open, public settings, yet the least frequently realized in the corpus. Compared with traditional KM, the paradigm characterizes INMA's distinctive approach to open DIK acquisition, flow, and codification and outlines the associated benefits and challenges.
ABSTRACT The increasing digitalization of healthcare systems demands secure, transparent, and interoperable data management solutions. Conventional centralized healthcare infrastructures have several significant drawbacks, such as the risk of breaches, poor interoperability between different applications, little or no control for patients over access to their own data, and reliance upon third‐party providers who can be untrustworthy. Due to the decentralized nature of its architecture and its resistance to tampering, blockchain technology offers the potential to create healthcare solutions that address each of these issues by enhancing trust and accountability within the healthcare community. This paper is a systematic review of recent research into blockchain‐based healthcare solutions which includes an examination of recent research development, technological frameworks, and current trends. In addition to classifying existing research according to application domains (such as Electronic Medical Record Management, IoT based Health Monitoring, Pharmaceutical Supply Chain Traceability, Insurance Fraud Detection, Disease Prediction) this paper also examines various aspects of blockchain technologies including platforms, consensus mechanisms, cryptographic techniques, and architectural models, while highlighting their strengths and limitations. Finally, this paper highlights several open challenges including scalability, interoperability, regulatory compliance, and limited real‐world deployments, and presents several possible research directions toward creating patient‐centric healthcare systems.
ABSTRACT Large language models (LLMs) generate fluent text but remain vulnerable to hallucinations, outdated knowledge, weak provenance, and costly model updates in knowledge‐intensive settings. Retrieval‐augmented generation (RAG) addresses these limitations by grounding generation in external evidence retrieved at inference time. This Advanced Review presents a PRISMA‐guided synthesis of 91 studies published between January 2020 and December 2025. It traces the evolution of RAG from early retrieve‐then‐generate pipelines to a broader design space that includes sparse, dense, and hybrid retrieval; multi‐stage reranking; evidence fusion; retrieval‐aware pre‐training; non‐parametric memory; iterative and agentic workflows; GraphRAG; multimodal RAG; multilingual retrieval; and privacy‐preserving deployment. The review examines how retrieval granularity, indexing, reranking, context construction, and retriever‐generator integration affect recall, precision, latency, robustness, and governance. It also compares explicit retrieval with long‐context prompting, vector‐based retrieval with graph‐based retrieval, and fixed pipelines with adaptive or agentic architectures. Domain requirements shape these choices: healthcare systems require evidence quality and privacy protection; legal and policy applications prioritize authoritative sources and citation fidelity; enterprise systems require freshness and access control; and educational systems must combine factual grounding with pedagogically appropriate explanations. Evaluation remains a central challenge. End‐task accuracy alone does not capture retrieval quality, faithfulness, attribution, calibration, abstention, efficiency, or resilience to prompt injection and knowledge‐base poisoning. The review therefore synthesizes current evaluation practices, promotes reproducibility‐oriented reporting through RAG Cards, and identifies open research directions for secure, efficient, transparent, and domain‐aware RAG systems.
ABSTRACT Recent advancements in deep learning have transformed medical image processing, particularly in integrating multimodal data and optimizing complex tasks, positioning fusion models as the leading paradigm in contemporary research. Historically viewed as mere pixel‐level blending, this field has fundamentally evolved into a broader paradigm of Medical Information Fusion. This review systematically explores deep learning‐driven advancements, focusing on their theoretical foundations, key technological breakthroughs and typical application scenarios. First, we systematically categorize fusion architectures into data‐level, feature‐level, and decision‐level hierarchies across diverse modalities. Second, we address key technical challenges and methodological breakthroughs, focusing on advanced solutions for domain shift and semantic alignment, handling data scarcity via missing modality completion, and task‐specific optimization through multi‐task learning. Furthermore, we critically examine the clinical transitional potential of these models in key clinical scenarios, including advanced tumor diagnosis, smart surgery, personalized therapy, and brain function analysis. By synthesizing these insights, this work aims to provide a comprehensive understanding of the current landscape and future directions of medical information fusion, paving the way for advancements in precision medicine and improved healthcare outcomes.
ABSTRACT This work aims to review the state of the art in the design, generation, and labeling of attack pattern datasets for the training of detection systems based on machine learning. A comparative study of different proposals will be carried out to detect shortcomings and areas for improvement in this field, which will serve as a starting point for future work. To this end, search and quality criteria have been defined to select a suitable set of works to be reviewed. A detailed analysis of the publications under study reveals the heterogeneity of criteria when it comes to research on attack pattern datasets. The variety of datasets analyzed, as well as the disparity in the selection of metrics and classification algorithms, makes it very difficult to establish clear comparative criteria. Further research is needed in this area to have updated datasets that include a catalogue of attack patterns as broad as possible. It is also worth highlighting the convenience of using new techniques and algorithms for the identification of attack patterns, such as those related to Deep Learning. This article is categorized under: Technologies > Classification Technologies > Machine Learning Fundamental Concepts of Data and Knowledge > Information Repositories
ABSTRACT Skeleton‐based hand gesture recognition (SGR) has emerged as a prominent research area in computer vision (CV) due to the unique advantages of skeletal data. Although gesture recognition (GR) has traditionally relied on video streams and RGB image data, skeleton‐based approaches remain underexplored. Despite advances in deep learning techniques, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), graph convolutional networks (GCNs), attention mechanisms, and multimodal fusion, a systematic and comprehensive review of these methods within the context of SGR is lacking. This paper addresses this gap by first underscoring the significance of GR and the pivotal role of skeletal data in its analysis. We then present the data acquisition techniques for SGR and examine its key applications, such as human‐robot interaction (HRI), sign language recognition, gaming and entertainment, and healthcare and rehabilitation. Furthermore, we analyze state‐of‐the‐art methodologies and compare their strengths and limitations in detail. Based on a thorough evaluation of existing research, we identify current challenges and unresolved bottlenecks. Finally, we propose future research directions to advance SGR, focusing on model robustness, multimodal fusion, and dataset expansion.
ABSTRACT Artificial intelligence (AI) is revolutionizing the field of epidemiology, offering transformative capabilities in disease surveillance, predictive analytics, and precision public health. By harnessing vast datasets from electronic health records, genomic sequences, and real‐time mobility data, AI‐driven algorithms enable early detection of outbreaks, accurate disease modeling, and personalized intervention strategies. This paper explores the multifaceted applications of AI in epidemiology, highlighting its role in enhancing disease prediction, exposure assessment, and outcome classification. AI‐powered platforms have demonstrated their efficacy in identifying emerging infectious threats, including the early detection of COVID‐19 outbreaks. Additionally, AI‐driven simulations provide policymakers with actionable insights, optimizing resource allocation and evaluating the effectiveness of public health interventions. However, challenges related to data privacy, algorithmic bias, and model transparency remain significant barriers to widespread adoption. This paper emphasizes the need for robust regulatory frameworks and interdisciplinary collaboration to ensure the responsible deployment of AI technologies in public health. By balancing innovation with ethical considerations, AI has the potential to drive a new era of precision epidemiology, transforming how diseases are monitored, prevented, and controlled on a global scale. This article is categorized under: Application Areas > Health Care Commercial, Legal, and Ethical Issues > Ethical Considerations Technologies > Artificial Intelligence
ABSTRACT Driver fatigue has been identified as one of the primary causes of traffic accidents. As long‐duration and high‐load driving becomes increasingly common, the risks of delayed reactions and impaired distance judgment continue to rise. Traditional behavior‐based methods for detecting driver fatigue often exhibit limited stability in complex driving environments. In contrast, electroencephalography (EEG) offers a more reliable detecting method by directly capturing central nervous system activity. This work focuses on fatigue driving detection based on deep learning and EEG, which outlines commonly used public datasets, key preprocessing techniques, feature extraction techniques, performance evaluation metrics, and mainstream deep learning architectures. Based on research progress over the past three years, the use of datasets, published journals, research challenges, and limitations of current methods were analyzed. Future research should improve the model's generalization ability and robustness, introduce richer brain network features, and construct a larger‐scale, high‐quality dataset that closely resembles the real driving environment.
ABSTRACT Gastrointestinal (GI) diseases remain among the leading causes of global mortality, with early detection directly linked to survival outcomes. While previous reviews have focused on single imaging modalities, this systematic review uniquely examines artificial intelligence applications across endoscopic, radiological, and histological approaches, reflecting actual clinical diagnostic pathways. This systematic review analyzes 76 high‐quality studies (2016–2024) and provides the first comprehensive assessment of how AI performs across different imaging techniques for GI abnormality detection. This multi‐modal perspective is particularly timely as healthcare systems move toward integrated diagnostic workflows. Our analysis reveals endoscopy as the most widely used modality ( n = 44), particularly for Helicobacter pylori , colorectal polyps, and ulcerative colitis detection. Histological analysis emerges as the second most common approach ( n = 25), especially for celiac disease and ulcerative colitis, while CT imaging ( n = 10) primarily supports colorectal polyp detection. Deep learning methods significantly outnumber traditional machine learning techniques (68 vs. 8 studies), consistently achieving 90%–99% diagnostic accuracy across multiple disease categories. However, these systems face significant implementation barriers to clinical adoption. Most validation is still conducted in controlled, single‐center settings using curated datasets that poorly reflect clinical complexity. Future studies must prioritize multicenter validation, standardized imaging protocols and preprocessing pipelines, and the integration of interpretable AI models capable of providing transparent diagnostic rationale. This review maps the current technical landscape while highlighting critical translational challenges that must be addressed to enable real‐world impact. This article is categorized under: Technologies > Data Preprocessing Technologies > Artificial Intelligence
ABSTRACT It is estimated that the volume of data on the digital fronts will grow exponentially to reach a volume of 180 zettabytes by 2025, and more than 90% of this data will be of unstructured forms. The unimodal to multimodal text analytics (MTA) has been triggered by this phenomenon. The early introduction of the multimodal text were observed in scholarly literature and industrial use‐cases during the early 2010s. Since then, it has greatly expanded its horizons in other sectors such as healthcare, e‐commerce, education and public safety. This survey presents a task‐oriented, modality‐inclusive, and dataset‐aware synthesis of recent advancements in MTA, which offers an in‐depth review of 10 core text analytics tasks through a multimodal lens. We systematically analyze over 160 research studies and categorize more than 120 state‐of‐the‐art models, spanning fusion strategies, representation learning, transformer architectures, and pretrained vision‐language frameworks (e.g., CLIP, ViLBERT). In a variety of datasets including CMU‐MOSI, CMU‐MOSEI, IEMOCAP, and MAViT‐Bangla, multimodal models achieve up to 18%–25% F 1‐score improvements over text‐only baselines, captured in the standardized task‐wise comparison tables that are part of this survey. Moreover, this survey discusses seven under‐explored tasks, including personality detection, satire detection, and author profiling, and elaborates gaps in research in modality fusion, diversity of data sets, and social inclusivity in these tasks. It does not only fill gaps in the current literature by unifying knowledge in different fields, but also offers researchers working on MTA a future path. It is the first survey that puts all the key tasks within multimodal text analytics into a contiguous and consistent overview compared to other surveys that either refer to multimodal computing at an administrative level or concentrate on a specific task. This article is categorized under: Algorithmic Development > Text Mining Algorithmic Development > Web Mining Application Areas > Society and Culture
ABSTRACT Generative artificial intelligence (GAI) methods have shown strong potential to accelerate drug discovery projects and de novo drug design. Yet, the fast pace of new proposals and the complexity of existing GAI methods configures a difficult scenario for professionals that aim to understand and effectively exploit these new tools. These issues are further exacerbated due to the plethora of different architectures for generative models, molecular representations, generation objectives and evaluation metrics. Moreover, other relevant aspects for understanding the inner workings of these models and their applications are scarcely addressed in the literature. For this reason, we propose different dimensions to taxonomically organize the wide range of GAI approaches used for drug discovery. These dimensions include the underlying computational model, the representation of molecules, and the building block for molecular generation. We also describe advantages and limitations of each family of methods, along with a comprehensive survey of metrics to evaluate performances from multiple perspectives. Besides, we explore how the concept of applicability domain applies to conditional generative methods, particularly in terms of quantifying the uncertainty associated with the generated molecules. As a final contribution, we describe and classify different approaches currently used in GAI‐driven drug design to enhance explainability. Finally, we discuss open challenges to strengthen the adoption of these generative models within research and industry. This article is categorized under: Technologies > Machine Learning Application Areas > Health Care
ABSTRACT Feature selection poses a common challenge in data mining and machine learning. This paper offers a comprehensive overview of the field, encompassing its historical context, recent advancements, and associated challenges. Topics covered include categorization of feature selection methods, development of ensemble frameworks, and assessment of selection stability. Current research trends in feature selection and computational intelligence are examined, alongside practical applications in ensemble configuration, threshold detection, evaluation criteria, and stability enhancement techniques. This article is categorized under: Algorithmic Development > Ensemble Methods
ABSTRACT Depression is a common and devastating mental health illness with serious personal and societal consequences. Despite advancing treatment techniques, there are still hurdles in the effective diagnosis and treatment of depression, such as prompt diagnosis, personalized medication, and continuous monitoring. In recent years, artificial intelligence (AI) has emerged as a potential tool in mental health treatment, providing novel solutions to these difficulties. This systematic study aims to comprehensively assess the existing AI systems for depression detection and treatment. The paper presents a systematic and comprehensive review of the last decade for depression detection, prediction, and treatment. One hundred eighty journal articles fulfilling preset inclusion criteria were found and analyzed using Preferred Reporting Items for Systematic Reviews and Meta‐Analyses from major academic databases. This review used a variety of detection modalities (physical, physiological, repetitive transcranial magnetic stimulation, and pharmacological treatment response) and AI approaches, including machine learning (ML) and deep learning (DL), to address various areas of depression care, including detection, diagnosis, prediction, and treatment. Key findings demonstrate that AI offers tremendous promise in boosting depression care across the continuum, from early identification to individualized therapy optimization and remote monitoring. ML and DL models demonstrate promising accuracy in predicting depression onset, severity, and treatment response based on diverse data sources, including electroencephalogram, electrocardiogram, photoplethysmography, electrodermal activity, electronic healthcare records, facial, speech, text, and pharmaceutical data. The paper highlights the important research challenges in current automated depression decision‐making models. Finally, we emphasize the prospects for developing effective and robust AI‐based depression models incorporating data and model fusion, the model's trust, portability, privacy preservation, and security features. This article is categorized under: Fundamental Concepts of Data and Knowledge > Explainable AI Technologies > Machine Learning Technologies > Artificial Intelligence
ABSTRACT The remarkable success of deep learning across computer vision, natural language processing, and medical diagnosis has largely depended on manually designed neural architectures—a labor‐intensive process lacking transferability. This has motivated the development of automated neural architecture search (NAS). This review systematically examines NAS through four methodological pillars: search space design, architecture optimization, hyperparameter optimization, and performance evaluation. For each pillar, we analyze representative approaches, identify trade‐offs (e.g., efficiency vs. reliability), and synthesize key insights, including the dominance of cell‐based spaces, the emerging trend of hybridization among optimization strategies, and the synergistic gains of multi‐fidelity hyperparameter methods. Beyond these pillars, we discuss quantum NAS as an emerging paradigm and outline seven open challenges, including benchmark expansion, interpretability, human bias, and adversarial robustness. This review provides a comprehensive reference for researchers and practitioners advancing automated deep learning.
ABSTRACT The Right to Erasure has facilitated the erasure of data as part of ethical and legal compliance for all users. In this regard, Machine Unlearning (MUL) is an emerging tool that transforms the existing trained model to execute comprehensive data erasure. Although beneficial to a few, the rapid development of unlearning algorithms has hindered a beginner's ability to recognize the relationship between the algorithm's productivity and model effectiveness. Though machine unlearning can be quite instrumental in building ethical and trustworthy AI, it is not beyond corrections. Substantive gaps in the form of standardization, formal guarantees, and verifiable implementation do exist. The aim of this paper, therefore, is to present a comprehensive understanding of the machine unlearning field literature and practices. Over 100 peer‐reviewed articles were reviewed, available from 2019 to 2025, including those on federated learning, continual learning, graphical neural networks, and rapidly growing models like generative and large language models. The paper presents a critical analysis of evaluation metrics, unlearning verification, efficiency, model utility on retained data, scalability, and handling of interdependent and multimodal data. Furthermore, it proposes future research directions, pressing the need for effective and efficient unlearning mechanisms. This article is categorized under: Technologies > Machine Learning
ABSTRACT High‐dimensional datasets with missing values present substantial challenges for statistical modeling and inference because conventional multiple imputation (MI) methods may become unstable or nonidentifiable under severe dimensional imbalance. Regularized regression techniques, including lasso, ridge, and related Bayesian formulations, offer a promising way to improve feasibility by enabling variable selection and stabilizing estimation in high‐dimensional settings. This overview provides a structured synthesis of regularized regression‐based MI methods, organized into direct and indirect integration strategies, Bayesian sparsity‐inducing approaches, optimization‐based and sequential penalization methods, and post‐imputation resampling or adjustment techniques. For each category, we summarize the underlying methodological logic, review empirical findings from simulation studies, and highlight practical considerations for implementation. Particular attention is given to methods developed for missing not at random (MNAR) and to the inferential challenges that arise after variable selection. Overall, this overview aims to help researchers choose regularized imputation strategies that are methodologically appropriate, practically feasible, and aligned with the analytical goals of high‐dimensional incomplete‐data studies. This article is categorized under: Technologies > Machine Learning Algorithmic Development > Statistics
ABSTRACT As a novel paradigm to overcome the problems of limited data and model generalization in the era of big data, transfer learning (TL) methods have become an essential branch of machine learning. However, the academic community still needs a data‐driven, dynamic, and quantitatively structured review of TL. This paper visually summarizes the literature growth and associated disciplinary trends of TL topics from a knowledge mapping perspective using 4496 documents collected in the WOS core collection, combined with bibliometric. The most essential 30 TL research results were identified and mapped spatial–temporally. This was followed by a structured analysis of the most influential research vectors including countries, institutions, authors, and publication sources. Following this, the knowledge base, core topic distribution, hot topics, and knowledge evolution process of TL research are systematically analyzed around the construction of a co‐cited literature network and keyword co‐occurrence network. The challenges of TL research are thereby summarized and future directions are proposed. This study provides researchers in the area of TL with a holistic insight into research trends, as well as conclusive and comprehensive analysis results that may help relevant scholars better grasp its dynamic direction. This article is categorized under: Technologies > Machine Learning Technologies > Artificial Intelligence
