Field service technicians typically rely on their experience and fault descriptions to identify and select the propriate spare parts for repairing an appliance. However, missing parts often lead to multiple service increasing operational costs and reducing customer satisfaction. To address this issue, we propose DL-SPP.sbert, a Deep Learning-based Spare Part Prediction system that leverages NLP and structured data to accurately predict the spare parts needed for repairs. DL-SPP.sbert processes historical repair data, appliance metadata, and descriptions by leveraging SBERT sentence embeddings to capture the semantic meaning of textual inputs. retrieval-based methods, our system directly learns the relationships between input features and spare parts through a multi-label classification model. DL-SPP.sbert was evaluated on two real-world case studies, where achieved up to a 64 % reduction in multiple-visit interventions, surpassing the performance of state-of-the-art baselines. A detailed explainability study highlights the key role of fault descriptions and appliance metadata prediction accuracy, confirming DL-SPP.sbert as an effective tool for optimizing spare parts provisioning. Finally, experimental results indicate that updating the model with newly collected data is fundamental to preserving predictive performance over time in continuously evolving real-world industrial settings.

The automatic classification of cell nuclei in histopathological images constitutes a fundamental component in the development of computer-aided diagnosis systems, offering valuable support in clinical decision-making and treatment planning. Despite the notable performance achieved by Deep Learning (DL) models in this domain, their limited interpretability remains a significant barrier to their adoption, especially in safety-critical fields such as healthcare. This study presents an explainable Nucleus Classification (NuC) system based on Fuzzy Decision Tree (FDT). The proposed approach leverages as input a set of human-interpretable numerical features extracted from the images of segmented nuclei. The system is evaluated on the PanNuke dataset, with a specific focus on the testicular tissue subset, and benchmarked against a Multi-Layer Perceptron (MLP) employed as a reference opaque model. Experimental results indicate that the FDT-based system attains competitive classification performance while offering intrinsically interpretable, rule-based outputs. In addition, we perform an explainability analysis demonstrating the proposed model’s capacity to generate linguistically meaningful rules that are consistent with domain-specific histopathological knowledge.

Neuroaesthetics investigates the neural activities during aesthetic experiences, using EEG recordings or fMRI images to decode the perception of visual art. However, studies in this domain are hindered by the limited and imbalanced nature of datasets, which is due to the subjective and resource-intensive nature of data collection. This study examines the effectiveness of various data augmentation strategies in enhancing EEG classification performance for neuroaesthetic analysis. We experiment three different EEG augmentation techniques, namely Signal Segmentation and Recombination, Temporal and Spatial Reconstruction Data Augmentation, and Gaussian Noise Addition. Furthermore, once extracted the features from the signals, we applied an instance-level data augmentation algorithm, namely SMOTE. We tested the four augmentation techniques individually, as well as SMOTE applied in cascade with the three EEG-specific augmentation methods, using stratified ten-fold cross-validation and leave-one-subject-out validation strategies. Results show that Gaussian Noise Addition, particularly when combined with SMOTE for generalization, yields consistent performance improvements in both accuracy and F-score. Conversely, Temporal and Spatial Reconstruction Data Augmentation often degrades classification performance (up to −9.95

Artificial Intelligence is widely recognised as a driver of adaptive autonomy in robotics. Yet, the extent to which AI techniques truly permeate the functional architecture of autonomous systems is still only partially characterised. Existing bibliometric analyses typically map research themes, keywords or algorithms but provide limited insight into how contributions distribute across the functional logic of autonomous systems. This raises a fundamental question: is AI really pervasive across the functions that enable robots to act adaptively in complex environments? Which areas are mature or under-explored? To achieve this outcome, the paper adopts a functional, control-loop-oriented perspective that avoids the bias of vertical domains or robot-specific applications, More than 2500 scientific works, published in the last 25 years, were mapped across the 13 functional modules, using a multi-label neural classification pipeline, and analysed via co-occurrence and structural techniques. This approach allowed to highlight not only areas where AI is already known to be central and consistently confirmed, but also those where its impact would be expected to be significant yet remains surprisingly limited. By combining architectural reasoning with bibliometric evidence, the study provides a broader lens for assessing research gaps and for situating current advances within the long-term agenda of adaptive and human-centred autonomy.

Objectives: This protocol presents the AI-SSI Project, which aims to develop, implement and validate an artificial intelligence (AI)-based system for the surveillance of surgical site infections (SSIs), to improve detection, support infection prevention and reduce the burden of healthcare-associated infections. Methods: At a tertiary university hospital (AOUP, Pisa, Italy), hospital discharge letters (HDLs) from 2020–2023 will be filtered using surgical and infection-related keywords and manually labelled with the involvement of public-health residents; candidate cases will be confirmed against other electronic health records (EHRs). Natural language processing and machine-learning models will be developed and internally validated on the labelled corpus, externally validated on data from other facilities, and re-trained through a federated learning approach that avoids sharing patient-level data. A parallel phase will estimate baseline SSI incidence and standardized structure and process indicators. Results: Anticipated outputs include a validated semi-automated SSI-detection pipeline integrating structured and unstructured data, baseline incidence estimates, a demonstration of federated learning feasibility across settings, and an assessment of EHR data quality. Conclusions: By exploiting under-used free-text data, this AI-based approach may complement traditional surveillance, strengthen the detection of post-discharge SSIs, and provide a transferable model for proactive, standards-based infection control.

Automatic detection of Personal Protective Equipment (PPE) can improve road construction safety through continuous visual monitoring in urban environments. However, roadside deployment poses two challenges: limited computational resources on embedded devices and degraded visual conditions due to lighting variability and blurring. To date, the extent to which detection models maintain their performance under such visual degradations has not been systematically quantified. This work presents a systematic evaluation of YOLOv8-based PPE detection on edge devices. Multiple model variants with progressively increasing complexity are deployed on resource-constrained hardware and compared across heterogeneous inference formats in terms of detection accuracy, memory footprint, and end-to-end latency. A controlled sensitivity analysis further quantifies performance degradation under selected brightness variations and Gaussian blur designed to emulate real-world roadside conditions. The results indicate that lightweight configurations of the model provide the most favourable accuracy-latency tradeoff. The Nano variant achieves 88.9% mAP50 at 10.2 FPS on Edge TPU hardware. Severe blur reduces accuracy by more than 27 percentage points, whereas brightness shifts induce moderate degradation. Increasing model complexity does not yield proportional robustness improvements.

In recent years, Deep Learning (DL) models have shown excellent results in various fields, including healthcare, but their “opaque” nature limits their trust and adoption in critical contexts. In this work, we present XAIMed2.0, an application designed to support healthcare professionals in interpreting the decisions of CNN models applied to medical image analysis. The application integrates several eXplainable AI (XAI) techniques to offer different visual explanations. The application integrates multiple eXplainable AI (XAI) techniques to generate visual explanations. Specifically, it employs several saliency map methods (Grad-CAM, Grad-CAM++, RISE, and LIME) and merges them into a cumulative visualization that highlights the regions of highest consensus across methods. To complement the visual explanations, the system uses a Large Language Model (LLM) to generate textual explanations in natural language, capable of contextualizing the results, highlighting discrepancies, and improving clinician understanding.

Brain Tumor Classification (BTC) using Magnetic Resonance Imaging (MRI) has achieved remarkable progress through Deep Learning (DL) models, particularly Convolutional Neural Networks (CNNs). However, the opaque nature of these models raises concerns regarding explainability, which is critical in clinical decision support. To address this, most research has focused on post-hoc Explainable AI (XAI) methods that provide after-the-fact interpretations of CNN predictions. In contrast, this work investigates an inherently explainable alternative based on Fuzzy Decision Trees (FDTs), which combine the interpretability of rule-based reasoning with the expressiveness of fuzzy logic. Moreover, we enhance model transparency by integrating radiomic features that capture clinically meaningful tumor characteristics such as shape, texture, and intensity. To the best of our knowledge, this is among the first studies to apply FDTs to brain tumor classification from MRI, explicitly coupling radiomics with multi-way FDT architectures. We perform a comprehensive evaluation comparing FDTs against four state-of-the-art CNNs, namely ConvNeXt, ResNet18, ResNet50, and EfficientNetB0, as well as classical binary Decision Trees (DTs). We provide an explicit analysis of the trade-off between accuracy, complexity, and interpretability of the models. Results show that FDTs achieve competitive performance (overall F1-score ≈ 0.84) compared to the best CNN baseline (ResNet50, F1-score ≈ 0.86), while offering substantially higher explainability and interpretability. Overall, this study demonstrates that FDTs can bridge the gap between accuracy and explainability, offering a viable explainable-by-design alternative to deep learning in medical imaging. Future work will focus on validating this generalizability across different imaging domains and dataset variations.

The forthcoming EU Machinery Regulation represents a significant evolution in industrial safety by enabling the use of Artificial Intelligence (AI) and autonomous technologies for safety-related functions in collaborative human-machine environments. However, these environments introduce new safety challenges related to human behaviour, improper machine use, and potential manumission of traditional safety measures. The aim of this study is to propose an integrated AI-based supervision system, named AISAFETY, designed to enhance operator safety in production processes through continuous monitoring and intelligent decision support. The proposed system is based on a layered architecture that integrates Radio-Frequency Identification for realtime operator localization, Computer Vision for visual monitoring, and a machine control layer governed by a rule-based expert system for managing safety-critical scenarios. The combination of heterogeneous sensing technologies provides redundancy and robustness in detecting hazardous conditions, such as unauthorized operator presence, tampering with safety guards, and improper collaborative robot operation. To support future validation, a prototype system has been implemented on an experimental setup consisting of a multimodal CNC lathe and a collaborative robot. A set of representative use cases has been defined to model potential hazardous situations and to guide the planned experimental evaluation of the system's safety supervision capabilities. The proposed approach highlights the potential of AI- and IoT-based supervision systems to support the implementation of upcoming regulatory frameworks and to advance human-centred safety management in collaborative industrial environments.

Federated Learning (FL) enables collaborative and privacy-preserving model training across distributed clients, but most existing FL systems implicitly assume data stationarity. In real-world settings—such as healthcare, industrial IoT (IIOT), cybersecurity, and smart cities—data streams are inherently non-stationary, leading classical FL methods to suffer from performance degradation, instability, and catastrophic forgetting.Continual Learning (CL) addresses learning under evolving data distributions but has been largely studied in centralized settings, overlooking key constraints of federated systems, including privacy, limited communication, and client heterogeneity. Federated Continual Learning (FCL) emerges at the intersection of FL and CL, aiming to support lifelong, adaptive, and privacy-aware learning over distributed and non-stationary data.This survey provides a comprehensive and systematic overview of FCL. We first present a formal definition of the FCL problem and clarify its distinctive characteristics. We then analyze the limitations of classical FL under non-stationary conditions, highlighting how CL principles support long-term adaptation. To organize the rapidly growing literature, we propose a multi-dimensional taxonomy of FCL approaches. Furthermore, we review representative application domains and data modalities, summarize commonly used evaluation metrics, and discuss experimental perspectives for assessing long-term performance and forgetting. Finally, we highlight key open challenges, including handling extreme heterogeneity under temporal drift, designing scalable and privacy-preserving memory mechanisms, and establishing standardized benchmarks. This survey aims to serve as a reference and a roadmap for advancing FCL toward robust and deployable real-world systems.

Federated Learning (FL) has emerged as a key paradigm for privacy-preserving collaborative model training across distributed and heterogeneous data sources. By keeping raw data local, FL addresses data confidentiality concerns, yet it does not resolve the opacity of modern machine learning models. In parallel, Explainable Artificial Intelligence (XAI) has gained attention for improving transparency, trust, and accountability, particularly in high-stakes domains. Their intersection has given rise to Federated Explainable Artificial Intelligence (FedXAI) paradigm, which aims to jointly satisfy privacy and explainability requirements. This survey provides a systematic review of FedXAI, highlighting the transition of explainability from a post-hoc tool to an integral component of the FL lifecycle. We show how explainability supports aggregation, personalization, robustness, coordination, and system-level decision making. To organize the literature, we introduce a taxonomy that classifies FedXAI methods by the role of explainability, model and explainer types, explanation scope, integration level, FL settings, and data heterogeneity. We review approaches ranging from model-agnostic explanations to interpretable federated models and explainability-aware aggregation mechanisms. We also examine evaluation practices and discuss the lack of standardized benchmarks and metrics for measuring explanation quality, stability, privacy leakage, and computational overhead. Finally, we identify key challenges, including explainability under non-IID data, explanation-centric security threats, communication-efficient XAI, continual FedXAI, and the integration of domain knowledge and regulatory constraints. By consolidating existing work and identifying key gaps, this survey serves as a reference framework for designing trustworthy, transparent, and privacy-preserving federated AI systems.

Brain tumor classification from Magnetic Resonance Imaging has achieved high predictive performance through deep data-driven models, which are often characterized by limited interpretability. In clinical settings, this lack of transparency is a critical limitation, as decision support systems must provide explicit and traceable reasoning processes. Most existing solutions address this issue through post-hoc explanation techniques applied to opaque models, without explicitly encoding decision logic within the model structure. This paper presents a modular framework for brain tumor detection and explainable-by-design classification from MRI images, explicitly separating tumor localization from decision making. Tumor regions are automatically detected and segmented using a YOLO-based instance segmentation model, while tumor classification is performed exclusively by a Prototype Tree-based classifier. Each prediction is produced through a single and explicit root-to-leaf decision path, ensuring intrinsic interpretability without relying on post-hoc explanation methods. The framework is evaluated end-to-end on a public brain tumor MRI dataset, considering multiple detector scales and Prototype Tree configurations. Experimental results show competitive classification performance on automatically extracted tumor regions, with a macro-averaged F1-score of approximately 0.87, while maintaining a compact and interpretable model structure with a maximum tree depth of five and thirty-seven nodes after pruning.

In recent years, fake news detection has received increasing attention in public debate and scientific research. Despite advances in detection techniques, the production and spread of false information have become more sophisticated, driven by Large Language Models (LLMs) and the amplification power of social media.We present a critical assessment of 12 representative fake news detection approaches, spanning traditional machine learning, deep learning, transformers, and specialized cross-domain architectures. We evaluate these methods on 10 publicly available datasets differing in genre, source, topic, and labeling rationale. We address text-only English fake news detection as a binary classification task by harmonizing labels into “Real” and “Fake” to ensure a consistent evaluation protocol. We acknowledge that label semantics vary across datasets and that harmonization inevitably removes such semantic nuances. Each dataset is treated as a distinct domain.We conduct in-domain, multi-domain and cross-domain experiments to simulate real-world scenarios involving domain shift and out-of-distribution data. Fine-tuned models perform well in-domain but struggle to generalize. Cross-domain architectures can reduce this gap but are data-hungry, while LLMs offer a promising alternative through zero- and few-shot learning. Given inherent dataset confounds and possible pre-training exposure, results should be interpreted as robustness evaluations within this English, text-only protocol.

The widespread adoption of Artificial Intelligence in everyday activities highlights a growing and urgent need for trustworthiness. Designing trustworthy AI systems requires addressing key technical challenges, including ensuring data privacy and model explainability. Federated Learning (FL) is a widely adopted paradigm to preserve data privacy in collaborative learning scenarios, while post-hoc methods are commonly applied to enhance the explainability of opaque AI-based models. In this paper, we propose a novel approach, called Federated SHAP, to simultaneously address privacy and explainability. Specifically, we leverage the SHapley Additive exPlanations (SHAP) method to provide post-hoc explanations of Neural Networks trained through FL. SHAP relies on a representative background dataset; however, constructing such a dataset in the FL setting is particularly challenging since raw data distributed across multiple clients cannot be shared directly due to strict privacy requirements. To address this challenge, we propose two tailored strategies depending on the data type: for tabular data, we adopt a Federated Fuzzy C-Means clustering algorithm to collaboratively summarize the distributed datasets into a suitable background dataset; for image data, we introduce a Federated Generative Adversarial Network (GAN) to synthesize representative background instances. A comprehensive experimental evaluation demonstrates the effectiveness and robustness of our proposed approaches, comparing them against several baseline and alternative strategies in terms of both representativeness and quality of generated explanations. Compared to baselines employing randomly generated representative background datasets, our approach reduces the discrepancy of SHAP explanations by up to three times on tabular data and two times on image data (depending on the test case involved), when measured against the centralized SHAP values computed using the full training set as background dataset.

The development of autonomous robotic systems offers significant potential for performing complex tasks with precision and consistency. Recent advances in Artificial Intelligence (AI) have enabled more capable intelligent automation systems, addressing increasingly complex challenges. However, this progress raises questions about human roles in such systems. Human-Centered AI (HCAI) aims to balance human control and automation, ensuring performance enhancement while maintaining creativity, mastery, and responsibility. For real-world applications, autonomous robots must balance task performance with reliability, safety, and trustworthiness. Integrating HCAI principles enhances human-robot collaboration and ensures responsible operation. This paper presents a bibliometric analysis of intelligent autonomous robotic systems, utilizing SciMAT and VOSViewer to examine data from the Scopus database. The findings highlight academic trends, emerging topics, and AI's role in self-adaptive robotic behaviour, with an emphasis on HCAI architecture. These insights are then projected onto the IBM MAPE-K architecture, with the goal of identifying how these research results map into actual robotic autonomous systems development efforts for real-world scenarios.

This paper describes three methods for analyzing electroencephalography (EEG) signals to classify users’ brain responses to exposure to art. The first two methods exploit classical machine learning approaches based on various sets of features extracted using different techniques for EEG analysis. In particular, the first method analyzes features extracted from time and frequency domains using an ensemble classifier, while the second one analyzes the Phase Locking Values of different channels using a classifier based on K-Nearest Neighbors. The third method retrains a well-known Convolutional Neural Network, namely VGG16, for image classification to analyze the scalograms obtained by applying the continuous wavelet transform to the EEG. These methods are evaluated employing a public dataset collected using mobile tools from museum visitors at an art exhibit. The dataset suffers from the problem of unbalanced classes, and this work also evaluates the impact of mitigation actions. The results reveal a significant difference between models tailored to the subject being tested, that achieve up to 95.43% of accuracy, and those trained without data from that subject (leave-one-subject-out strategy), that achieve up to 65.35% of accuracy. This suggests that, at this stage, customized approaches are more appropriate for the neuroaesthetics field, whereas models with general applicability need further development. We also conducted an analysis to assess whether age influenced the performance of the models.We split the visitors into three groups based on their age: adolescents, young adults, and adults. The analysis performed with a leave-one-subject-out strategy revealed higher accuracy for adolescents (73.21%) and adults (66.98%) than young adults (60%).

Industry 4.0 advanced technologies are increasingly used to monitor workers and reduce accident risks to ensure workplace safety. In this paper, we present an on-premise, rule-based safety management system that exploits the fusion of data from an Ultra-Wideband (UWB) Real-Time Locating System (RTLS) and AI-based video analytics to enforce context-aware safety policies. Data fusion from heterogeneous sources is exploited to broaden the set of safety rules that can be enforced and to improve resiliency. Unlike prior work that addresses PPE detection or indoor localization in isolation, the proposed system integrates an UWB-based RTLS with AI-based PPE detection through a rule-based aggregation engine, enabling context-aware safety policies that neither technology can enforce alone. In order to demonstrate the feasibility of the proposed approach and showcase its potential, a proof-of-concept implementation is developed. The implementation is exploited to validate the system, showing sufficient capabilities to process video streams on edge devices and track workers’ positions with sufficient accuracy using a commercial solution. The efficacy of the system is assessed through a set of seven safety rules implemented in a controlled laboratory scenario, showing that the proposed approach enhances situational awareness and robustness, compared with a single-source approach. An extended validation is further employed to confirm practical reliability under more challenging operational conditions, including varying camera perspectives, diverse worker clothing, and real-world outdoor conditions.

Agricultural robotics and artificial intelligence (AI) are becoming essential to building more sustainable, efficient, and resilient food systems. As climate change, food security pressures, and labour shortages intensify, the integration of intelligent technologies in agriculture has gained strategic importance. This systematic review provides a consolidated assessment of AI and robotics research in agriculture from 2000 to 2025, identifying major trends, methodological trajectories, and underexplored domains. A structured search was conducted in the Scopus database—which was selected for its broad coverage of engineering, computer science, and agricultural technology—and records were screened using predefined inclusion and exclusion criteria across title, abstract, keywords, and eligibility levels. The final dataset was analysed through descriptive statistics and science-mapping techniques (VOSviewer, SciMAT). Out of 4894 retrieved records, 3673 studies met the eligibility criteria and were included. As with all bibliometric reviews, the synthesis reflects the scope of indexed publications and available metadata, and potential selection bias was mitigated through a multi-stage screening workflow. The analysis revealed four dominant research themes: deep-learning-based perception, UAV-enabled remote sensing, data-driven decision systems, and precision agriculture. Several strategically relevant but underdeveloped areas also emerged, including soft manipulation, multimodal sensing, sim-to-real transfer, and adaptive autonomy. Geographical patterns highlight a strong concentration of research in China and India, reflecting agricultural scale and investment dynamics. Overall, the field appears technologically mature in perception and aerial sensing but remains limited in physical interaction, uncertainty-aware control, and long-term autonomous operation. These gaps indicate concrete opportunities for advancing next-generation AI-driven robotic systems in agriculture. Funding sources are reported in the full manuscript.

The use of personal protective equipment (PPE) is essential to improve workplace safety. Despite specific regulations requiring the use of PPE, workers often neglect to wear it due to factors such as inattention, urgency, or convenience. Monitoring the correct use of PPE is especially critical in high-risk tasks. Computer vision technology can automate this process, leveraging deep neural models. This study investigates the performance of modern object detection models in identifying the correct use of PPE, focusing on their accuracy and execution speed. Specifically, the YOLOv11 and RT-DERT models are employed, trained on a real-world PPE dataset. Deployment on low-cost hardware, specifically an NVIDIA Jetson Nano, is evaluated using three deployment frameworks, namely PyTorch, OpenVINO, and TensorRT. The results show that YOLOv11n, with 2.6 million parameters, provides slightly lower average accuracy than more complex models. It stands out for its speed, reaching performances of 6.6 Frames Per Second (FPS) with PyTorch, 2.3 FPS with OpenVINO, and 10.6 FPS with TensorRT. On the other hand, YOLOv11l and YOLOv11x with, respectively, 46.5 and 86.7 million parameters offer higher accuracy, especially evident in small class identification, where simpler models tend to struggle. However, they show lower throughput, with 1.2 and 0.7 FPS on PyTorch. RT-DETR has competitive accuracy but lower performance on edge devices.