Existing agent-oriented conceptual modeling frameworks assume components with deterministic execution semantics. Applied to language model agents, whose outputs arise from probabilistic text generation governed by natural-language inputs, they lack three constructs that accountability requires: a binary declaration of each agent’s legitimate action space; runtime enforcement of negative obligations by a component external to the agent; and treatment of inter-agent communication as auditable evidence. To address these gaps, we introduce AI-Intent, a conceptual modeling framework organized into three pillars: Declaration encodes each agent’s action boundary, decision rights, capabilities, and risk policies in Mandates; Enforcement evaluates every Proposed Action against the applicable Mandate through a Compliance Agent before delivery; Auditability derives a structured, durable Accountability Trace from every session. A reference implementation for private investment advisory under MiFID II is evaluated across 190 evaluation sessions: boundary violation containment averaged 97.9% with zero forced-pass occurrences, while audit trace completeness was found to depend on the instructionfollowing capability of the deployed model.

Preface to special issue bridging conceptual modeling and machine learning

The substitution of a urine sample that may result in an adverse analytical finding with a previously collected, clean sample is strictly prohibited under the World Anti-Doping Agency (WADA) regulations and is referred to as sample swapping. When an athlete reuses their own clean sample, detection becomes particularly difficult through conventional analytical methods. In this paper, we propose a similarity detection framework that explicitly accounts for pattern complexity in the analysis of urinary steroid profiles. The framework is based on a convolutional network to capture more complex and subtle variations in profile pairs. Using a dataset of 67,651 steroid profiles collected between 2021 and 2023, the framework was evaluated on both synthetic and laboratory-confirmed similar samples, reflecting realistic variability in doping control processes. The results show that the proposed framework outperforms several baseline models, achieving higher accuracy compared to different baselines. These findings demonstrate the potential of machine learning to improve anti-doping workflows by enabling the automated detection of reused or identical urine samples within large-scale sample collection managed by the Athlete Biological Passport.

Early diagnosis of Alzheimer’s disease (AD) remains a major challenge due to the subtle and temporally irregular progression of structural brain changes in the prodromal stages. Existing deep learning approaches require large longitudinal datasets and often fail to model the temporal continuity and modality irregularities inherent in real-world clinical data. To address these limitations, we propose the DiffusionGuided Attention Network (DiGAN), which integrates latent diffusion modelling with attention-guided convolutional network. The diffusion model synthesizes realistic longitudinal neuroimaging trajectories from limited training data, enriching temporal context and improving robustness to unevenly spaced visits. The attention-convolutional layer then captures discriminative structural–temporal patterns that distinguish cognitively normal subjects from those with mild cognitive impairment and subjective cognitive decline. Experiments on the ADNI dataset demonstrate that DiGAN outperforms existing state-of-the-art baselines, showing its potential for early-stage AD detection.

Deploying vision models at the industrial edge requires balancing accuracy against energy and latency constraints that server-oriented models cannot meet, yet prior distilled-ViT work measures efficiency through proxy metrics (FLOPs and server-GPU throughput) rather than actual on-device energy. We close this gap on the DOES nonalloyed steel-scrap benchmark (8,131 test tiles, eight EU steel grades), distilling ViT-Large and Vision-LSTM (ViL-Base, an xLSTM-based backbone) teachers into 5.5M-parameter DeiT-Tiny students and benchmarking every model on an NVIDIA Jetson Xavier NX with onboard power monitoring and latency measurement. Our distilled students achieve 88.81% and 87.07% top-1 with 2.93–3.56% accuracy gaps to their teachers, reducing inference energy by 15.9–34× and latency by 18.1–44×. To the best of our knowledge, this is also the first cross-inductivebias comparison in knowledge distillation for vision involving a recurrent xLSTM-based teacher: same-family distillation (ViT→DeiT, 88.81%) achieves 1.74 pp higher accuracy than cross-family (ViL→DeiT, 87.07%), yet both deliver equivalent on-device efficiency.

Personalizing rehabilitation requires predicting how individual patients will experience pain yet two patients undergoing identical surgery can report starkly different outcomes. This heterogeneity is not noise: it is systematically driven by personality, beliefs, emotional state, and life circumstances. Psychology has spent decades formalizing precisely these individual differences through validated frameworks such as the Biopsychosocial model and the International Classification of Functioning. Yet existing AI approaches ignore this theoretical structure and treating patient variables as flat numerical features. We propose Psychology-Grounded Artificial Mental Models: a framework that encodes established psychological theory as structured reasoning scaffolds within large language models, enabling them to construct an artificial mental model of each patient before making predictions. To evaluate the framework, we conducted a DiscoverYourself study with 336 participants, and further validated the framework on publicly available PhysioPain dataset (N=82). Across four LLM families, theory-grounded models achieve up to +56.9 percentage points exact-accuracy improvement over the generic baseline. These results indicate that encoding psychological theory into LLM reasoning substantially improves pain prediction in rehabilitation, with the largest gains observed among older, highly heterogeneous patients where data-driven models are least reliable.

Personalizing rehabilitation requires predicting how individual patients will experience pain yet two patients undergoing identical surgery can report starkly different outcomes. This heterogeneity is not noise: it is systematically driven by personality, beliefs, emotional state, and life circumstances. Psychology has spent decades formalizing precisely these individual differences through validated frameworks such as the Biopsychosocial model and the International Classification of Functioning. Yet existing AI approaches ignore this theoretical structure and treating patient variables as flat numerical features. We propose Psychology-Grounded Artificial Mental Models: a framework that encodes established psychological theory as structured reasoning scaffolds within large language models, enabling them to construct an artificial mental model of each patient before making predictions. To evaluate the framework, we conducted a DiscoverYourself study with 336 participants, and further validated the framework on publicly available PhysioPain dataset (N=82). Across four LLM families, theory-grounded models achieve up to +56.9 percentage points exact-accuracy improvement over the generic baseline. These results indicate that encoding psychological theory into LLM reasoning substantially improves pain prediction in rehabilitation, with the largest gains observed among older, highly heterogeneous patients where data-driven models are least reliable.

Large Language Models have shown strong generalization across natural language tasks but remain underexplored for longitudinal clinical profiles. In sports anti-doping, biological profiles are analyzed to support early detection of prohibited substance use and identification of anomalous biological patterns, both of which require joint modeling of temporal dynamics and metabolic relationships. We propose STT-LLM, a structural-temporal tokenization framework that adapts LLMs to longitudinal clinical analysis without modifying their backbone architectures. STT-LLM constructs biologically grounded structural-temporal embeddings and transforms them into LLM-compatible tokens via specialized tokenizers that explicitly encode pathway structure and temporal evolution. We evaluate STT-LLM on real-world longitudinal datasets from athletes, showing consistent improvements over native LLM tokenization strategies in sequence prediction and anomaly detection. In addition, we present a case study where STT-LLM provides contextual reasoning that aligns more closely with expert assessments compared to baseline models. These results highlight tokenization as a key bottleneck and opportunity for adapting LLMs to clinical data.

Real time AI model serving workloads are characterized by bursty and variable request patterns, yet production GPUs commonly operate at their maximum frequency, wasting substantial energy whenever latency headroom exists. Reducing energy consumption without violating tail-latency SLOs is challenging because frequency scaling affects latency in highly nonlinear, workload-dependent ways. This paper presents HALO (Hierarchical Adaptive Latency-Oriented DVFS), a fully black-box power-management controller that uses probe p95 latency to classify the system into coarse safety zones, gates the set of admissible frequency actions per zone, and runs a contextual bandit within those constraints to adapt online, requiring no offline profiling, model instrumentation, or servingframework modifications. For LLM inference on the Azure LLM Inference Trace (Coding workload), HALO reduces total GPU energy by 39.4% on the primary model and up to 50.0% across model scales, while maintaining tail latency under the SLO. Applied to LLM pretraining, the same controller reduces energy with convergence unaffected, demonstrating that the approach generalizes across both inference and training workloads.

The expansion of AI workloads intensifies the energy demand of data centers, resulting in rising operational costs and environmental impacts. Although various approaches aim to enhance energy efficiency, existing solutions remain fragmented and as a consequence, organizations lack integrated transparency and actionable guidance to manage the performance–energy trade-offs of AI operations holistically. This study introduces an Information Systems perspective to enable cost-effective and energy-efficient AI operations in data centers through decision support systems. Employing a Design Science Research (DSR) methodology, a first design cycle is conducted that includes the development of a prototype consolidating energy-related information while providing decision support for cost-effective and sustainable AI inferencing. The prototype artifact is evaluated through practicing experts, indicating its usefulness in increasing transparency and supporting energy-aware operational decisions. The findings inform a set of design principles for decision support systems that foster sustainable AI operations in data centers.

Accurate prediction of dementia progression is crucial for timely intervention and care planning, yet remains challenging due to highly variable patient trajectories. Recent machine learning approaches rely on neuroimaging and genetic biomarkers. While effective, these methods are costly, invasive, and difficult to deploy at scale in primary care settings. In contrast, routinely collected longitudinal clinical data, such as questionnaire based assessments and care intervention records remain underutilized for prognosis. We propose VISTA (Velocity- Informed Sequential Transition Analysis), a framework for dementia prognosis using real-world clinical data. It explicitly encodes temporal velocity - the rate of change in clinical features between consecutive assessments together with the absolute clinical state. It addresses the limitation of existing methods that treat assessments as independent snapshots. Our framework converts longitudinal prediction as transition-level forecasting via sliding-window decomposition, which increases the number of training instances while preserving patient-specific temporal structure. We implement it using gradient boosted decision trees to capture non-linear state-velocity interactions while handling missing data natively. We evaluate the framework on DelpHi dataset an 8-year real-world primary care longitudinal dataset comprising of 459 patients (768 annual transitions). VISTA achieves R^2 = 0.74 and MAE = 2.63 MMSE points, substantially outperforming linear mixed models (R^2 = 0.25) and standard baselines in cognitive prognosis. Ablation studies and SHAP analysis also revealed several crucial patterns. Qualitative analysis shows VISTA accurately captures systematic decline patterns while still maintaining predictions on volatile trajectories. These results demonstrate that velocity-aware modeling enables accurate dementia prognosis.

To this day, the requirements of data warehouses, user visualizations and ML projects have been tackled in an independent manner, ignoring the possible cross-requirements, collective constraints and dependencies between the outputs of the different systems that should be taken into account to ensure a successful analytical project. In this work, we take a holistic approach and propose a methodology that supports modeling and subsequent analysis while taking into account these three aspects. This methodology has several advantages, mainly that (i) it enables us to identify possible conflicts between actors on different tasks that are overlooked if the systems are treated in an isolated manner and (ii) this holistic view enables modeling multi-company systems, where the information or even the analytical results can be provided by third-parties, identifying key participants in federated environments. After presenting the required formalism to carry out this kind of analysis, we showcase it on a real-world running example of the tourism sector.

Individuals' increasing generation of digital data and evolving data ownership rights reinforce monetization opportunities in personal data markets. Understanding how personal data valuation and pricing is construed becomes essential. So far, psychological factors such as data sensitivity or proximity are under-researched. This study examines how construal level theory contributes to construing and explaining individual data value through an exploratory quantitative survey (n = 633). The results show that different construal levels influence data pricing, psychological ownership and willingness-to-disclose.

Mobility in rural regions is critical for ensuring access to essential services such as education, healthcare, and employment. However, rural mobility remains largely individualized, with heavy reliance on private cars—leading to environmental, social, and infrastructural challenges. Despite increasing data availability from GPS, mobile devices, and public transport applications, data-driven approaches for generating integrated mobility insights tailored to specific rural areas remain underdeveloped. This paper addresses this gap by investigating how decision support systems (DSS) can be designed to provide actionable, regionspecific mobility insights for policymakers, businesses, and public transport providers. Following a Design Science Research (DSR) methodology, we identified factors of mobility mode choice. From this, we derived meta-requirements and proposed a set of design principles for data-driven DSS in rural mobility contexts. These principles were implemented in a DSS for the Saarland region of Germany and evaluated with policymakers and public transport planners. Our findings contribute both conceptual design knowledge and practical guidance, demonstrating how DSS can foster more efficient, inclusive, and sustainable rural mobility planning.

Machine learning enables the extraction of useful information from large, diverse datasets. However, despite many successful applications, machine learning continues to suffer from performance and transparency issues. These challenges can be partially attributed to the limited use of domain knowledge by machine learning models. This research proposes using the domain knowledge represented in conceptual models to improve the preparation of the data used to train machine learning models. We develop and demonstrate a method, called the Conceptual Modeling for Machine Learning (CMML), which is comprised of guidelines for data preparation in machine learning and based on conceptual modeling constructs and principles. To assess the impact of CMML on machine learning outcomes, we first applied it to two real-world problems to evaluate its impact on model performance. We then solicited an assessment by data scientists on the applicability of the method. These results demonstrate the value of CMML for improving machine learning outcomes.