ABSTRACT. Within this lecture, mental health is conceptualized not merely as the absence of mental health disorder, but as a set of human operational capabilities, related to mental readiness, stress resilience and cognitive agility, enabling individuals to perform their professional tasks effectively in highly competitive, unpredictable and potentially stressful environments. In this context, stress resilience is defined as the capacity of individuals exposed to extreme stress or trauma to maintain stable psychological and physiological functioning, while avoiding long-term mental health consequences. It encompasses mental and cognitive readiness, emotional stability, and decision-making effectiveness under pressure. In the era of artificial intelligence, new opportunities are emerging to enhance human cognitive, emotional and behavioral performance through virtual assistants, adaptive human machine interfaces, large language models, state-of- the-art immersive AR/VR technologies, and wearable multimodal neuro-psycho-physiological sensors. Special emphasis is placed on multimodal emotion elicitation, estimation and regulation, implemented through AI based cognitive behavioral training and emotionally based strategic communications, enabling real-time optimization of human performance in highly stressful jobs and professions. Application of these concepts are illustrated in learning, training and performance optimization of professionals operating in potentially highly stressful environments, such as fighter pilots, air traffic controllers, and other cognitively demanding roles.

ABSTRACT. The research area of image processing and computer vision AI based object detection and classification approaches have massively improved their performance in the last recent years. These performance gains will be addressed during the keynote speech and demonstrated at a few sample applications. Although there are a lot of new AI technologies under development (e.g. pre-trained models as well as training procedures which work fine also with smaller training sets) the availability of reliable training samples remains still key to enable high performance and reliable models. So, one crucial challenge of AI based solutions remains valid, how to get “enough” and trusted training samples to train the models for particular applications and tasks. Therefore, in the second part of the talk the focus will be put on methodologies how to generate or synthesize training images under controlled conditions. The solutions presented have reduced significantly the amount of data to be manually labelled and annotated. Moreover, the approaches presented also support the generalization capabilities of the models, work fine with classical augmentation techniques and provide full control over the variations to be considered as well as inclusion and exclusion criteria.

ABSTRACT. The Implantable Cardioverter-Defibrillator (ICD) represents one of the most significant advancements in modern cardiac electrophysiology. Designed to prevent Sudden Cardiac Death (SCD), the device serves as a sophisticated monitoring and intervention system for patients at high risk of lethal ventricular arrhythmias. However, current ICD implantation criteria have limited precision, resulting in both a high number of unprotected patients and a high percentage of devices that never deliver the therapy. This paper evaluates the ability of simple neural networks to predict patients at higher risk of an appropriate ICD activation among a cohort who received the device at the Croatian Tertiary Hospital Centre following the current criteria. Various architectures and training strategies were employed, with SHAP analysis used to interpret the influence of different clinical, device-derived, and demographic features on the model predictions. The two best-performing neural network architectures and results are reported. While the first network achieved an AUC-ROC of 0.72, the other one achieved an ideal recall of 1.00, with the AUC-ROC remaining relatively stable at 0.69. SHAP analysis identified preimplantation ventricular tachycardia and its frequency, secondary prevention, device type, and follow-up time as some of the key predictors, consistently aligning with findings from our previous studies.

ABSTRACT. Endovascular treatment of cerebral aneurysms and other neurovascular pathologies requires navigating a microcatheter through narrow, tortuous intracranial vessels under fluoroscopic guidance. Current practice relies on dye injection and either biplane fluoroscopy, which doubles radiation exposure, or monoplane imaging, which sacrifices depth perception. We present a complete pipeline for monoplane 3D catheter localisation in brain endovascular procedures trained entirely on synthetic data. First, we develop a synthetic data generation pipeline that produces anatomically plausible catheter trajectories through the cerebral vasculature derived from publicly available CTA volumes with per-vessel annotations, and renders them as physics-based fluoroscopic images using DeepDRR. We assess the visual fidelity of the synthetic data by training nnU-Net for 2D catheter and guidewire segmentation, achieving Dice scores of 0.976 and 0.988, respectively. Second, we propose a graph neural network (GNN) that formulates 3D catheter detection as binary node classification on the pre-operative vessel graph, fusing independently encoded 3D vessel and 2D catheter graph representations through cross-modal attention conditioned on the C-arm projection geometry. The model achieves an F1 score of 0.828 on the cranial 30° view with a median tip localisation error of 1.50 mm after postprocessing. To the best of our knowledge, this is the first work to address learning-based monoplane 3D catheter localisation in the cerebrovascular domain.

ABSTRACT. A practical, reproducible pipeline for olive yield estimation is presented, combining close-range UAV still imagery, automated fruit detection, and a simple harvest-based calibration step. For a chosen set of six test trees, one high-resolution canopy image was captured at multiple dates, a fine-tuned object detector was applied to obtain digital fruit counts, and the median of per-tree detections was used as a robust representative digital count. Harvest weights collected at the end of the season were converted to estimated fruit counts using per-cultivar mean fruit mass, while calibration coefficients were computed as the ratio of harvest-derived counts to the representative digital counts. Calibration coefficients are proposed for two observed olive cultivars, enabling estimation of total per-cultivar yield in the observed orchard. The workflow is evaluated and the primary sources of uncertainty -- viewpoint-dependent coverage, occlusion, illumination variation, and detector errors -- are analysed. The results demonstrate the feasibility of translating single-view digital counts into calibrated yield estimates, while also highlighting substantial variability that motivates pragmatic mitigations: further in-domain labelling and retraining, multispectral foliage masking, scale-aware architectures, and multi-view sampling. The proposed approach is lightweight and suitable for operational orchard monitoring; recommendations and a roadmap to reduce uncertainty and improve robustness are provided.

ABSTRACT. Input sequence length is a critical factor in long-term time series forecasting, as it determines the amount of historical context available to the model. Despite its importance, most studies in the literature adopt a fixed input length of 96 as a standard across different models and datasets. In this work, we systematically investigate the impact of input sequence length by evaluating six benchmark forecasting models (TimeXer, TimeMixer, TimesNet, iTransformer, SOFTS, and DLinear) on six widely used benchmark datasets. We consider a range of input lengths, including both shorter and longer contexts (48, 72, 96, 168, and 336). Through extensive experiments, we analyze the effects of input sequence length on both predictive accuracy and model stability/robustness across multiple runs. The results show that the impact of input length is significant but highly dependent on the model and dataset, with no single configuration consistently achieving the best performance. Overall, the goal of this study is to understand how input sequence length affects different models under varying conditions, in order to provide deeper insights into its role in long-term forecasting. Such analysis may also help identify more appropriate configurations and potentially establish improved baseline settings for future research.

ABSTRACT. Extreme precipitation events represent high-impact hazards, motivating the development of lightweight early-warning models that can be trained directly from station observations. This paper investigates whether rolling Benford deviation features, widely used in forensic analytics and anomaly detection, provide additional predictive value for short-horizon forecasting of precipitation extremes.

Using daily NOAA station data from five locations representing arid, semi-arid, and humid precipitation regimes, we formulate the task as predicting whether a station-specific extreme threshold ($P95$, estimated on the training period) will be exceeded within the next three days. A strong temporal baseline is constructed using lagged predictors, rolling statistics, seasonal encoding, and temperature variables when available. Benford features are computed in a past-only rolling manner across multiple window sizes (90, 180, and 365 days), including first-digit deviation (MAD), divergence measures, rounding ratios, and valid-sample counts.

Evaluation using leakage-aware rolling-origin cross-validation shows that the baseline achieves strong performance (PR-AUC $\approx 0.76$--$0.84$), while Benford features provide no improvement: MAD-only features are nearly neutral, whereas richer Benford feature sets slightly reduce performance. The results further indicate that Benford reliability strongly depends on wet-day coverage, becoming unstable in arid climates and largely redundant in humid regimes. These findings suggest that Benford analysis is more suitable for meteorological data quality control and regime characterization than for short-term extreme-event prediction.

ABSTRACT. Education is a continuously evolving process with new teaching methodologies and tools emerging each year. A common constraint in any course is time. Every lecture or practical exercise would benefit from having more time to show more examples or clarify complicated concepts. One such course that would benefit from more time is digital logic (DL). During laboratory exercises, students are expected to create and implement logic gate circuits, and this has to be monitored by a teacher. As group sizes increase, teachers have to divide their attention more and more, reducing the amount of time that can be spent helping individual students. This paper aims to introduce a system that can help students and teachers in and out of classes with their education in DL. A proof of concept for one of the main building blocks of the proposed system is explained, along with results of its implementation, extracting the logic function from an image of a handwritten logic gate circuit. This is achieved using YOLO combined with additional image processing algorithms. The workflow used to create a dataset of logic gates and the methods and algorithms used to extract the logic function are explained in this paper. Results obtained with YOLO-v11-small are comparable to the literature (99.04\%), confirming its potential to detect small objects such as logic gates in the hand-written circuit diagrams.

ABSTRACT. This study investigated autonomic fatigue phenotypes in competitive athletes using multidomain heart rate variability (HRV) features and unsupervised AI-based analysis. Twenty-two athletes (18–23 years) underwent 10-min Holter ECG recordings under standardized conditions of rest (pre-exercise), stress (immediately post-exercise), and recovery (two hours later). Time-domain (MeanNN, SDNN, RMSSD, pNN50, CVNN) and frequency-domain (Total Power, LFn, HFn, LF/HF) features were extracted to characterize autonomic regulation. Statistical analysis demonstrated significant differences across all physiological states (ANOVA p < 0.0001), with RMSSD, pNN50, Mean RR, and LF/HF showing the strongest discriminative ability. Principal component analysis (PCA) revealed that 60.1% of the total variance was explained by the first two components, while three components accounted for 71%, indicating a structured low-dimensional autonomous feature space. Hierarchical clustering (Ward linkage) applied to the PCA-transformed domain identified three distinct clusters corresponding to states of rest, stress, and recovery, with recovery occupying an intermediate adaptive position. The main contribution of this study is the demonstration that fatigue can be represented as a latent multidimensional autonomous phenotype rather than a single-parameter deviation. The findings support AI-based unsupervised modeling as a robust framework for fatigue recognition and provide a methodological foundation for future HRV-based digital twin systems in athlete monitoring.

ABSTRACT. The paper discusses reliable perception, control, and system integration in terms of autonomous UAV landing in a constrained environment. It offers a modular simulation pipeline for vision-based UAV landing, which is developed through PX4 SITL, Gazebo Harmonic, ROS 2 Humble, and a ROS 2 offboard state machine controller. This simulation pipeline offers a unified workflow to conduct image acquisition, target detection, middleware com-munication, trajectory setpoint generation, and landing execution. Through 40 independent trials, it achieves a high probability of 87.5% in terms of autonomous landing, a mean final horizontal landing error of 0.29 m, and a mean mission duration of 29.8 s. This study has shown that autonomous landing is greatly influenced by continuous target visual detection in the final phase of landing. The contribution of this paper is a transparent simulation pipeline for UAV landing, which is practical, as opposed to a new con-trol algorithm.

ABSTRACT. This paper presents the theoretical foundation and conceptual architecture for the application of Multi-Agent Systems (MAS) in automated generation of Scalable Vector Graphics (SVG) from natural language text descriptions. The paper is conceived as the initial phase of a broader research program to be developed within a doctoral dissertation, establishing theoretical foundations, identifying research gaps in the literature, and proposing a system architecture whose implementation and formal evaluation will be the subject of future scientific work. The proposed multi-agent architecture consists of five specialized agents collaborating on the decomposition of the complex creative task of SVG generation: Natural Language Understanding Agent (NLU Agent), Composition Planning Agent (Layout Agent), SVG Primitive Generation Agent (SVG Generator Agent), Validation and Optimization Agent (Validator Agent), and Coordination Agent (Orchestrator). The paper includes a preliminary benchmarking experiment that empirically documents the limitations of single-agent models and motivates the multiagent approach.

ABSTRACT. The integration of extended reality (XR) technologies into educational practice requires structured professional development that supports both technical and pedagogical competencies of educators. This study presents the design, implementation, and evaluation of a modular XR training framework developed through a design-based research approach combining needs assessment, iterative module design, and post-implementation evaluation. The framework organizes training into competency levels that gradually introduce XR concepts, practical tool usage, and pedagogical application. It was implemented through a series of professional development modules conducted nationally across multiple locations, and its effectiveness was evaluated using structured participant feedback. The results indicate that practice-oriented activities, particularly those involving hands-on work with XR tools and content creation, are perceived as the most valuable components of training, while more theoretical elements require further adaptation to better support applied learning. In addition, participant background influenced perceived difficulty and learning experience, especially in more advanced modules. The findings suggest that modular and practice-oriented training frameworks provide an effective approach for supporting the gradual adoption of immersive technologies in educational contexts and offer guidance for the design of future teacher professional development programs.

ABSTRACT. Despite all technological advancements and the modernisation of processes, a business system still cannot function without the presence of the human, who is the key factor in every stage of its life cycle. Humans are its input resource, an integral part of its business process, and the objective towards which all its output components are directed. The human remains an irreplaceable segment of the entire system, regardless of the constant progress in the automation of business processes driven by intensive advancements in the development of artificial intelligence (AI). To put it simply, everything that a business system produces or achieves as an output is directed towards the human as its end user.

The use of artificial intelligence (AI) significantly influences traditional management and decision-making methods by introducing new, enhanced forms of automation that simplify established processes, while systematically expanding into other areas of human labour and activity. Consequently, a challenging question inevitably arises: where lies the boundary between the human and the machine, or rather, between the human and a hybrid of human and machine? Every day, artificial intelligence increasingly shapes our contemporary reality, while simultaneously guiding the scientific community towards researching and testing the role of the human in the smart systems of the postmodern era.

ABSTRACT. This study explores trainers’ perceptions of the digital transformation of EU-related training in public administration. Training plays a key role in strengthening administrative capacity for European integration process. However, it is still predominantly delivered through traditional face-to-face formats. As public administration increasingly adopts digital technologies, there is a growing need to explore the potential of digital learning formats in the field of EU-related training. This study examines the views of trainers involved in EU-related training regarding the perceived effectiveness of digital training, their readiness to adapt to digital formats, and the perceived contribution of digital training to strengthening administrative capacity for EU integration. The research is based on a survey conducted among certified trainers of Directorate for European Integration involved in EU-related training. The findings provide insights into trainers’ attitudes towards the digital transformation of EU-related training, providing for key opportunities and challenges in relation to future development and introduction of digital learning in this field.

ABSTRACT. This study investigates the development of audio description (AD) for a UNESCO-protected cultural heritage site (the Diocletian’s Palace complex) using a bilingual tourist guide as the primary source text. An interdisciplinary framework integrates perspectives from translation studies, audio description theory, accessibility research, and artificial intelligence (AI). The methodological design combines several complementary stages: AI-assisted summarization of the English tourist-guide content; the creation and adaptation of narration scripts through a hybrid human–AI workflow, translation and terminological alignment into Croatian; both automatic and expert-based evaluation of AI-generated scripts; the production of AI-generated English AD tracks alongside human-authored Croatian AD tracks and end-user assessment of the resulting AD tracks. The research addressed the following questions: What are the main linguistic, cognitive, and stylistic adaptations required when designing adult-oriented AD? What role can AI tools play in supporting the creation of accessible and culturally meaningful AD, and what are their limitations? How do visually impaired adults perceive and evaluate the adapted AD in terms of clarity, informativeness, and engagement? By situating AD at the intersection of intersemiotic translation, human–AI collaboration, and user-centred accessibility practices, the study seeks to contribute to emerging scholarship on AI-supported AD generation as well as to broader discussions on inclusive cultural-heritage communication and accessible design methodologies.

ABSTRACT. This study examined the effects of simulation-supported instruction in elec-tronics on students’ affective learning outcomes within the school subject Technical Culture. A quasi-experimental research design was used, with ex-perimental and control groups comprising students from five primary schools (N = 204). During the instructional intervention, teachers in the experi-mental group used the Tinkercad simulator to demonstrate and support the teaching of electronic circuits, while traditional teaching methods were used in the control group. After the intervention, data were collected using a questionnaire measuring three affective dimensions: satisfaction with in-struction, interest in electronics, and career preferences related to electronics. Data were analysed using descriptive and inferential statistical methods, in-cluding Pearson’s correlation coefficient, independent samples t-test, and the Mann–Whitney U test. The results showed a statistically significant differ-ence between the experimental and control groups in students’ satisfaction with instruction, with the experimental group reporting higher satisfaction. However, no statistically significant differences were found in students’ in-terest in electronics or their career preferences related to this field. Statisti-cally significant positive correlations were identified between all analysed af-fective dimensions, with the strongest relationship observed between interest in electronics and career preferences. The findings highlight the pedagogical potential of simulation-supported instruction to improve students’ affective engagement in electronics education.

ABSTRACT. Schools differ substantially in readiness for AI integration, yet systematic frameworks for differentiated implementation remain limited. The present paper examines how digital maturity foundations enable the development of AI literacy in Croatian schools. Four data sources were analyzed: e-Schools digital maturity evidence from the pilot phase (151 schools), complemented by findings from the final Phase II evaluation (313 schools and 10 educational centres), BrAIn curriculum design, recent empirical research (3,913 students, 382 teachers), and international literature. The integrated framework connects digital maturity, knowledge management, stakeholder readiness, and AI literacy outcomes. Results reveal differentiated and evolving levels of digital maturity: 50.4\% of schools were in the initial maturity phase, 43.0\% digitally enabled, and 6.6\% advanced, while subsequent Phase II evidence suggests that these readiness levels developed over time. While 67\% of students use AI tools, only 45\% of teachers feel prepared, indicating critical implementation gaps. The framework provides differentiated pathways tailored to school maturity levels, offering actionable guidance for policy, leadership, and professional development. This research addresses the urgent need for context-sensitive AI integration strategies in diverse educational settings.

ABSTRACT. In this paper, we explore the possibilities and limitations of applying artificial intelligence (AI) in the process of student revision in mathematics courses in engineering studies. Students starting their studies in technical and natural sciences often need to have a good knowledge of high school mathematics in order to successfully follow courses such as Mathematics 1. In practice, however, it often turns out that some students come to such courses with insufficient prior knowledge and are unable to adequately follow the lessons. The development of generative artificial intelligence opens up the possibility for students to independently use AI tools to review and practice the necessary mathematical content. The aim of this study was to examine whether students in higher education can effectively use AI tools, with minimal teacher guidance, to review secondary school mathematics content. The research was conducted using an experimental design with two groups of students. The experimental group used AI tools to review the material with guided instructions, while the control group studied the same material using traditional methods and textbooks. Before the intervention, a pre-test was administered, and during and after the learning process short knowledge checks and midterm exam results were analyzed in order to compare performance between the groups. The results show that students who used AI tools did not achieve better results than students who used traditional methods of reviewing the material. The obtained results are interpreted in the context of recent research on the phenomenon of the “illusion of competence,” according to which AI-generated explanations can create a feeling of understanding without actual acquisition of mathematical knowledge.

ABSTRACT. This paper analyzes the application of modern information and com-munication technologies in improving the operational capabilities of fire service commanders and proposes an educational model for man-aging crisis situations. Increasingly complex fire and flood scenarios re-quire data-driven decision-making using technologies such as drones, satellite systems, real-time meteorological data, simulation models, IoT systems, and artificial intelligence. Despite the availability of these technologies, a gap exists between their capabilities and user competencies, along with a lack of standard-ized training programs. The paper proposes a modular educational model based on simula-tions, scenario-based learning, and an interdisciplinary approach. Ex-amples of wildfires in Dalmatia and floods in Slavonia confirm the im-portance of integrating technology into command processes. In conclusion, systematic education is key to the successful digital transformation of firefighting and crisis management systems.

ABSTRACT. This paper presents the design and educational applications of four 3D-printed robots developed for undergraduates, students, doctoral students and students from specialized and technological schools. A detailed description of humanoid hand, delta robot, walking robot “Big Foot” and a modular planar robot is pro-vided. The main goals and tasks for which the robots can be used are systema-tized in a table. For each of the robots, a case study from the education system is given, with links to illustrative videos. The generally valid advantages and disad-vantages of using 3D printed robot models in education are discussed. Specific advantages of the four presented robots are: the ability to print already assembled mechanical components; the creation and use of parts with a complex shape, in-cluding internal cavities, and the use of a minimalist approach for designing mo-bile robots. The minimalistic approach to the design of a robot makes it applicable to students of different age and to users with specific educational needs. Exam-ples of the various applications can be seen in the links provided in the article.

ABSTRACT. Moodle offers a wide range of complex functionalities and options, which may become a barrier to effective use for some teachers. Several chatbots are designed to assist students using Moodle, but there is limited evidence on chatbots designed to support teachers. This work addresses that gap by de-signing, implementing, and evaluating a retrieval-augmented generation (RAG) system as a technical foundation for a future chatbot meant to sup-port teachers in selecting and using Moodle tools. The RAG system devel-oped in this research comprises a vector database, a retriever and a generator. For the experiment, a validation set of 50 questions was created from the knowledge base, and three open-source LLMs (Mistral 7B, LLaMA 8B, Qwen2.5 7B Instruct) were used to generate answers in two chunk-size set-tings (500 vs 1000 characters). Automated, reference-free evaluation was conducted using RAGAS, with a single LLM-based evaluator (GPT-4o-mini) and the following metrics: faithfulness, answer relevance, and context utilisation. The results suggest that the system provides an appropriate technical basis for a chatbot to support teachers using Moodle, as faithful-ness and context utilisation were generally high. However, answer rele-vance was consistently lower. The most robust difference in generator selec-tion was observed in context utilisation, where Mistral scored significantly higher than LLaMA for both chunk sizes with large effects. Chunk size did not consistently improve any metric.

ABSTRACT. Contemporary financial and accounting reporting systems are characterized by fragmentation and insufficiently integrated information systems, which limit comprehensive, transparent and timely insight into business transactions for all relevant stakeholders. The research gap lies in the absence of a unified multi-entry accounting model that would enable the synchronized recording of transactions in real time. This paper proposes a conceptual model of automated financial reporting based on the integration of blockchain technology and artificial intelligence. Blockchain enables decentralized and secure, auditable recording of financial transactions, while artificial intelligence supports the automation of accounting entries, data analysis and compliance with accounting and regulatory standards. The proposed model has the potential to enhance transparency, reduce operational risks and strengthen trust among stakeholders in financial reporting.

ABSTRACT. The project is entitled “Investigation, Mathematical Analysis, and Evaluation of the Impact of Stress on Cardiological Data.” Its primary objective is to conduct fundamental research through the application of a novel mathematical approach for analysing and evaluating the effects of mental and physiological stress on heart rate. Mental stress has been investigated through the development of an extreme virtual reality game using various stereoscopic technologies, and its impact on heart rate variability has been analysed. The influence of physiological stress has been studied during sports training sessions in order to assess training load based on heart rate variability.