This study presents a suite of AI-based services, including speech-to-text, text-to-speech, and text-to-sign translation, designed to support language learning for both deaf and hearing individuals in immersive Extended Reality (XR) environments. In particular, it explores and develops speech-to-text, text-to-speech, and text-to-sign translation technologies, to enable inclusive, multimodal communication tailored to diverse user needs. These services become available via APIs with the goal to be explored by 3D avatars within Extended Reality (XR) settings, delivering engaging, self-directed learning experiences. The paper emphasizes the importance of modular and reusable design, and proposes a framework to ensure interoperability and scalability across various XR platforms, addressing critical challenges in accessibility and multilingual interaction, opening new opportunities in the field for equitable educational solutions in immersive environments.
Large Language Models (LLMs) have emerged as auxiliary tools in clinical decision-making. However, their inherent opacity raises concerns about reliability and interpretability in medical practice. This study evaluates the reasoning structure of different LLMs in generating clinical recommendations, focusing on logical coherence and adherence to medical guidelines. Six complex clinical cases were submitted to three LLMs: OpenAI ChatGPT-4o, OpenAI O3-Mini-High, and Gemini 2 Flash Thinking Experimental. Model-generated responses were analyzed using Bardin's content analysis and statistically compared to reference guidelines. Results indicate that, while all models exhibit structured reasoning, variations exist in the depth of clinical considerations, particularly in postoperative care and diagnostic test selection. The integration of explicit reasoning frameworks enhances traceability in AI-generated medical recommendations. However, further refinements are required to ensure consistency and reliability in clinical applications.
The increasing accessibility of virtual reality (VR) systems offers new possibilities for enhancing therapeutic methods, particularly in upper limb rehabilitation. This paper presents a comparative study of a VR-based rehabilitation application operated with two types of hardware: a previously developed dedicated controller integrated with an orthosis, and a commercially available Valve Index controller. The new setup eliminated the need for external trackers and specialized orthotic equipment, simplifying the user configuration process. The VR application, featuring gamified exercises for the shoulders, elbows, wrists, and fingers, was adapted for the new controller using Unity's XR Interaction Toolkit. Key modifications included revised interaction mechanics, script updates, and enhanced physics handling. A preliminary user study involving ten healthy participants evaluated the application's usability, intuitiveness, and effectiveness through surveys. Results showed that the commercial controller was rated higher for ease of use (mean score: 4.5/5) and interaction quality (4.3/5), while the orthotic device offered greater joint-specific targeting. Participants reported high satisfaction overall, though some motion tracking inaccuracies were noted. While limited by the healthy participant sample, the findings highlight the potential of consumer-grade VR hardware for accessible and effective rehabilitation solutions. Future development will focus on further customization options, automated result tracking, and integration with clinical workflows.
This study focuses on the integration of clinical, genomic, and epigenomic data to develop advanced predictive models for risk stratification in acute myeloid leukemia (AML). AML is a hematological disease characterized by the uncontrolled proliferation of immature cells, with genetic mutations and chromosomal alterations that influence prognosis and treatment. The primary aim of the project is to enhance the prediction of survival, therapeutic response, and relapse likelihood using an innovative approach based on machine learning and the integration of multiple data sources, including RNA-seq, DNA methylation, and genetic mutations. The work explores various data fusion techniques, including early and late fusion, as well as the use of explainability methods such as SHAP and Grad-CAM to improve transparency and reliability of clinical decision-making. The data comes from five datasets within the TCGA-AML project, which include clinical, gene expression, and mutation information. The proposed multimodal approach aims to develop robust, personalized, and biologically interpretable predictive models, with the goal of optimizing treatments and reducing the risk of relapse. This approach could have implications not only for AML but also for other types of leukemia and cancers, enhancing the overall management and prognosis of oncology patients.
The application of Artificial Intelligence (AI) techniques to criminal profiling represents a significant innovation in crime analysis and management. This study employs advanced machine learning models, specifically Random Forest and multi-output classifiers such as MultiOutputClassifier, based on the RandomForestClassifier algorithm, to analyze the ISTAT dataset on reported crimes in Italy from 2019 to 2023. The primary objective is to accurately predict both the type of crimes and their geographical distribution. Compared to previous studies in the literature, this approach stands out for its greater comprehensiveness and advancement, not only by utilizing a dataset specific to crimes recorded in Italy but also by integrating multi-output models capable of simultaneously predicting crime types, such as thefts, frauds, and other offenses, alongside their locations. Additionally, the study incorporates advanced preprocessing techniques to enhance data quality and reliability, achieving a high predictive accuracy exceeding 90
Binary surgical tool segmentation is a crucial component of Computer-Assisted Intervention (CAI) applications in minimally invasive surgery (MIS). Accurate segmentation is essential for instrument tracking, surgical scene understanding, and augmented reality overlays. However, a complete CAI system consists of multiple parallel processes, necessitating fast, accurate, and robust segmentation models with low computational cost and minimal memory footprint. In this study, a comprehensive benchmarking analysis of various backbone architectures and decoder topologies for binary surgical tool segmentation is conducted. The evaluation assesses memory efficiency, computational complexity, and generalization ability across unseen surgical datasets, including robotic, laparoscopic, and endoscopic procedures. The experimental results reveal the trade-offs between segmentation performance and resource constraints, offering valuable insights for selecting efficient models suitable for real-time surgical applications.
The integration of XR technologies into healthcare has dramatically altered the landscape of medical practice, patient treatment, and medical engineering. This paper explores practical applications across three pivotal areas: XR for doctors, where anatomical visualizations help in preparation and smooth realization of complex surgical procedures and training; XR for patients, such as phobia treatment, orthopedic rehabilitation and personalized product configuration; and XR for medical engineers, showcasing the use of digital twins and virtual tryons for innovative medical solutions. Several case studies of own experience are reviewed, demonstrating the real-world impact of these technologies. The presentation also addresses technical aspects of XR application development, particularly the creation of flexible, dynamic content, ensuring adaptability and personalization in healthcare settings. This exploration illustrates how XR can transform healthcare into a more intuitive, effective, and patient-centric field.
The study outlines the creation of a groundbreaking application that implements Augmented Reality (AR) and Artificial Intelligence (AI) to identify parts of the hybrid propulsion system in the Toyota RAV4 hybrid. This research aimed primarily to develop an educational resource enabling technicians and students to grasp and explore hybrid propulsion systems without needing physical parts, utilizing interactive virtual models that can be accessed via mobile devices. Central to the system is a convolutional neural network (CNN) that has been trained using the YOLOv8 algorithm, enabling precise identification of vehicle components. A public dataset was used for model training, ensuring the results are valid and can be compared with earlier research. Furthermore, the model has been exported in ONNX format, which makes integration with platforms like Unity easier, thereby creating a smooth and engaging AR experience. The application developed illustrates how merging AR and AI can revolutionize technical education, allowing for more effective and accessible learning. With AR, users can see and learn about the components of hybrid propulsion systems in realtime, which enhances their comprehension of these intricate systems. Feedback from users indicated a high level of satisfaction regarding the interface, detection accuracy, and overall learning utility of the tool. This research shows the promise of AR and AI to enhance technical education in the automotive sector, and it holds potential for adaptation in other vehicle models or various fields, providing a scalable and effective method for training in both educational and professional settings.
This paper presents a comparative evaluation of two state-of-the-art deep-learning-based generative methods, NeuralMaterial and MaterialGAN, for material generation in 3D modeling application – Blender. For comparison, we introduce a baseline method developed using only tools available in Blender. A user study involving 20 participants, including both experts and lay users, was conducted to assess usability and effectiveness. Participants engaged in a material authoring task, evaluating all developed approaches. Usability was evaluated using the System Usability Scale (SUS) questionnaire, and material fidelity was validated by comparing user-created materials to target scene renders using Learned Perceptual Image Patch Similarity (LPIPS) metric. Users are able to achieve the desired results with a similar level of precision as with the existing tools but with higher editability. Results also show preference differences between lay and expert users on which approaches they prefer and why. Our research provides valuable insights into deep-learning-based material generation integrated into everyday workflows. It highlights the importance of considering user preferences and expertise levels when designing such workflows. By comparing performance and usability, we inform the development of more effective and accessible material generation techniques within the field.
Motion analysis plays a fundamental role in numerous fields, from sports science to rehabilitation, allowing the study and evaluation of human movement. Traditional marker-based systems, while highly accurate, present limitations such as high costs, complex setup, and restrictions in natural movement. To overcome these issues, markerless motion capture methods have gained increasing attention, exploiting advances in computer vision and machine learning.This work was born by the experimental activity conducted during the ActivE(3) project at the Human Performance Laboratory at the Lecco Campus of the Politecnico di Milano. During the project, aimed to explore new methodologies for rehabilitation and movement analysis and to promote inclusiveness and accessibility in physical activity, the Virtual Reality Nirvana (BTS Bioengineering SpA, Garbagnate Milanese, Milan, Italy) system was employed. Although the system provided highly engaging and beneficial for enhancing movement participation, it lacked the capability to quantitatively measure joint kinematics. This limitation highlighted the necessity for a reliable, markerless motion tracking solution, to bridge the gap between interactive rehabilitation technologies and precise motion tracking, ensuring a quantitative evaluation. The study focuses on the development of a markerless motion analysis system designed to be accessible and reliable for clinical and sports applications. The proposed system employs two cameras to record movement, the MediaPipe framework to extract body keypoints, and MATLAB for the computation of knee flexion-extension angles and data processing.The system was validated through experimental trials involving human subjects performing standardized motion tasks. The computed joint angles were compared against measurements obtained using an IMU-based motion capture system, which is considered with high level of accuracy in the field. Quantitative evaluation metrics such as RMSE, ICC, Spearman's rank correlation coefficient and bias were used to assess system performance. The results demonstrate that the proposed markerless system provides accurate joint angle estimations, supporting its potential application in real-world scenarios where traditional motion capture is impractical.
The paper presents a case study of the FRIA project, aimed at researching and specifying a methodology to assess the impact of Artificial Intelligence (AI) systems on fundamental rights. In this paper we present a case study on an AI-based hiring system to test the methodology and define the interactions with the final users. The research output is a prototype tool to support and automate the fundamental rights impact assessment of high-risk AI systems, which aims to comply with the requirements of the European Artificial Intelligence Act. The research methodology is interdisciplinary and based on a collaboration between legal professionals and computer scientists in the framework of the SoBigData Research Infrastructure ( www.sobigdata.eu). It starts from the study of the existing legal and ethical frameworks concerning AI and human rights at the International and European levels and the translation of the identified rules and principles into a set of parameters to measure the AI risk and provide a synthetic set of requirements to create a semi-automated risk assessment model.
Medication management is a daily and time-consuming task for hospital nurses. It includes searching for generic substitutions in online databases when a prescribed drug is not available, and also searching for specific drugs in the actual medication room to find where it is physically located. In co-creation with stakeholders an Augmented Reality-based companion was created, which resides in a holographic device that can be worn during practice and assists nurses through a holographic, mixed reality environment, supporting the task of medication management in a comfortable way. Ethical assessment of the system, conducted using the AI4EU assessment tool, highlighted key strengths of the system and also identified areas for improvement, particularly in transparency legislative clarity. A challenge for user testing in clinical environments is that hospital medication rooms are strictly regulated. To enable future user studies a Medication Room Simulator was developed.
Digital (AI-driven) working environment is reframing Indian industries to compete in the stronghold of other competing markets digitally. Its role is completely reinventing work environments and organizational culture. Organizations have well understood that they need digitalization to keep pace and be useful. It entails vision, a strategy for increasing productivity and driving innovation through technology. External entities, including customers and partners, benefit from increased effectiveness, quicker response times, and improved service delivery. Internally, Workplaces that are primarily digitalized by cloud computing, AI, and automation enable agile and collaborative environments that provide satisfaction and work-life balance for employees. Again, this digital working environment fosters a culture of reuse by digitizing learning environments, for instance, e-learning platforms, virtual classrooms, and AI-enabled training facilities, for companies to reskill and upskill employees to suit the current requirements. This study is an attempt to understand how digital working environment and digitalized decision-making process equips Indian enterprises to lead in innovation, productivity, and growth in a changing digital environment.
A systematic review of AI-enabled indoor localization and navigation is presented, focusing on how these advancements address the limitations of GPS-based systems. A PRISMA-based approach was employed to identify relevant peer-reviewed articles in English, ultimately selecting 65 papers from 54 journals involving 253 authors. Bibliometric techniques and keyword clustering reveal an evolving research landscape centered on digital indoor modeling, machine learning, deep learning, sensor fusion, and robotics-based applications. Citation trends confirm growing interest in AI-driven solutions for signal obstruction, multi-floor settings, and real-time navigation, demonstrating a shift toward integrated sensors and IoT devices for enhanced positioning accuracy. Emergent thematic clusters highlight human–machine interaction, including Augmented Reality, emergency management, and healthcare assistance, indicating diverse contexts in which AI-based indoor localization can be adopted. The findings underscore how algorithms such as fingerprinting, neural networks, and Bayesian methods refine performance and reliability. Despite considerable progress, the literature indicates a lack of standardized methodologies and infrastructures, underscoring the continued need for adaptable, cost-effective solutions. Overall, the potential of AI to elevate indoor navigation services through greater precision, robustness, and situational awareness is emphasized, while future work should focus on interdisciplinary collaboration, unified frameworks, and the ethical implications of data-driven systems.
The integration of Digital Twins (DTs) into healthcare promises transformative advancements in personalized medicine, resource optimization, and predictive analytics. However, ethical concerns about data privacy, algorithmic bias, and accountability remain critical barriers to widespread adoption. This study combines quantitative and qualitative analyses (n = 120 participants) to explore public perception and ethical implications of DTs in healthcare. Results reveal a significant correlation between prior knowledge of DTs and positive perception (78
In this study, we present a cooking activity recognition method using wearable sensors and a stacking ensemble model. Users wear a wearable device with an accelerometer and an electromyography sensor on each wrist. During cooking, the system collects acceleration, angular velocity, and EMG data. These are classified into seven activity classes (e.g., washing, peeling, cutting, frying, stirring) to train the model, which achieved an average recall of 63
Nursing homes face increasing challenges due to workforce shortages, ageing populations, and rising care demands. This paper explores the dynamics of human-robot interaction (HRI) in nursing home environments, focusing on how autonomous robots engage with nurses and residents in nursing homes. We examine the role of trust-building and emotional engagement in caregiving scenarios, where patient wellbeing depends on effective HRI. The potential for autonomous systems to support decision-making while maintaining human oversight is investigated. Using the Autonomy Levels for Unmanned Systems (ALFUS) framework, we assess care robots' autonomy in terms of mission complexity (MC), environmental complexity (EC), and external system independence (ESI). Additionally, the Autonomy and Technology Readiness Assessment (ATRA) method supports gradual capability enhancement, providing a roadmap to higher autonomy while ensuring safety and reliability. Based on this established methodology, we introduce the Nursing Home Interaction Framework (NHIF), a novel conceptual framework defining the roles of actors involved, to connect theoretical insights with real-world applications. This work highlights the significance of human-robot trust in elderly care scenarios and proposes directions for future research on collaborative decision-making in nursing home HRI.
The idea that the intelligence of Nature may be embedded in fractals has already been stated by author, since 2016, as in 2, 4. This paper is expanding on previous papers of the author 1, 3, by providing an alternative intelligence model to complete the present model of A.I. This paper also introduces two new concepts: the concept of evolutionary fractal and the concept of procedural randomness. The intelligence model presented in this paper is based on the CSN Algorithm. This CSN Algorithm can be used, on one hand, to decrypt evolutionary fractals having unknown rules, and on the other, it can be used to simulate the way in which random creative processes occur. The CSN Algorithm simulates better how brain behaves, because when making scientific discoveries, often times, existing information is not processed as in classical A.I models, which are mostly deterministic, but it is processed aleatory by brain, in random processes, as in the model based on the CSN Algorithm. Decryption of these evolutionary fractals may be the key of understanding of human brain functioning and of human intelligence. The first step, however, is the decryption of evolutionary fractals related to more simple lifeforms, and only then, after the decryption at this first step, the decryption of the human intelligence may be addressed. The model presented in this paper can either be used as an entirely independent intelligence model to simulate human intelligence random creative processes, or it can be used to create enhanced intelligence models by joining the model presented in this paper, based on the CSN Algorithm together with the present classical model of A.I. In a further section of paper, other additional possible applications of this algorithm are presented.
This research explores the development of an advanced AI-driven dialogue simulator specifically designed to assist educators in effectively practicing and navigating various pedagogical scenarios through role-playing. The system integrates state-of-the-art generative AI with virtual reality (VR), utilizing dynamic 3D avatars capable of interactive dialogue and real-time question generation based on predefined knowledge and user input. Educators can engage in realistic simulations, receiving personalized and contextually relevant questions and feedback from AI-generated virtual interlocutors. By leveraging speech-to-text technology, user interactions are seamlessly processed to generate immediate, targeted responses, facilitating an immersive and responsive learning environment. This dialogue simulator aims to enhance educators' preparedness, adaptability, and communication skills by allowing iterative practice in a safe virtual setting. Technical implementations include the utilization of open-source large language models, real-time avatar lip synchronization, customizable virtual personas, and sophisticated scenario generation methods. Initial evaluations were conducted to assess the feasibility and usability of the VR-based dialogue simulator. Specifically, exploring whether educators found the system immersive, realistic, and helpful for practicing pedagogical scenarios. Future research will further refine AI responsiveness and expand scenario complexity, contributing meaningfully to AI-assisted pedagogical training methodologies.
Precise localization of surgical instrument tips is essential for evaluating fine motor skills and enabling automation in microsurgical training. This study presents a deep learning framework based on keypoint heatmap regression to detect instrument tips in frames extracted from simulated surgical videos. A dataset of 1781 annotated frames from seven videos was used for evaluation. The framework was trained with different loss functions-root mean squared error (RMSE), weighted Kullback-Leibler divergence (WKLD), and Dice loss-and compared with direct coordinate regression and segmentation-based models. The RMSE-based model achieved the best performance (MAE = 7.54 pixels), while the WKLD-based model provided more stable predictions across thresholds for blank mask detection. Segmentation and direct regression models showed significantly higher errors. Statistical analyses confirmed the advantage of heatmap regression over baseline approaches. These results support the adoption of heatmap-based keypoint localization for robust tool tracking in simulated surgical environments and its integration into training systems for skill assessment.
