ENFIELD Human-Centric AI Webinar #2: From Smart Environments to Wearable Health Monitoring

This webinar brings together two research talks at the intersection of human-centric AI and real-world applications. The first explores how communities of AI agents can collaborate with humans to accomplish goals in smart environments, making their reasoning transparent and adaptable. The second presents a novel wearable-derived biomarker for stress detection based on sleep dynamics, combining transformer models with physiological signals and explainability techniques. Together, the talks highlight how explainability and human involvement can make AI systems more trustworthy and actionable in everyday life.

Date: 22 April 2026, 14:00-15:00 CET
Location: Online (Teams)
Secure your spot here (registration closed).

Participation is free.

The webinar is organized in the framework of the Horizon Project ENFIELD - European Lighthouse to Manifest Trustworthy and Green AI – and it will take place on the Teams platform.

Registration Closed

Target audience

AI researchers, practitioners, and students with an interest in human-centric AI, explainable AI, or smart environments. Also relevant for professionals working on wearable health technology, multi-agent systems, or AI-assisted wellbeing and stress monitoring.

Speakers

Andrei Olaru, Associate Professor, University POLITEHNICA of Bucharest

Ander Cejudo, PhD Researcher, Vicomtech, Basque Research and Technology Alliance, Spain

Topics & Abstracts

Title:	Goal-Based Interaction of Humans and Agent Communities in Smart Environments
Speaker:	Andrei Olaru
Bio:	Andrei Olaru is a researcher at the University Politehnica of Bucharest and a member of the Artificial Intelligence and Multi-Agent Systems (AI-MAS) group at the University POLITEHNICA of Bucharest. He is actively involved in the Human-Centric AI task in the ENFIELD project. His current work focuses on goal-based interaction between humans and agent communities in smart environments.
Abstract:	Our recent work explores hypermedia-driven, goal-based interaction in smart environments using communities of agents built on the paradigm of Hypermedia Multi-Agent Systems (HMAS). In this approach, agents receive high-level goals from users and construct plans by following hypermedia links and affordances exposed by devices and services in the environment. Rather than acting independently, agents operate within communities that share interaction experience, allowing them to reuse previously successful solutions and adapt to changing conditions. Humans are an essential part in this process. Human users provide feedback, clarify goals, or validate intermediate steps, and this input can be incorporated into the shared knowledge of the agent community. This supports gradual improvement of behavior without requiring fully predefined workflows. Explainability of the final plans devised by agents comes both from the interaction with humans, and from the participation of the community. Different agents in the community may propose alternative actions, as they rely on different contextual information, and reference different past experiences. By making these contributions explicit, the system can provide simple explanations of how a decision was reached, including which options were considered and why one was selected.

Title:	Explainable Wearable-Derived Sleep Dynamics Biomarker for Day-to-Day Stress Detection
Speaker:	Ander Cejudo
Bio:	PhD researcher working on AI and health applications in the department of digital health and biomedical technologies at Vicomtech.
Abstract:	Sleep dynamics, defined as the temporal patterns of stage transitions during the night, are essential for understanding sleep regulation, but most wearable-derived features do not capture this information. Artificial intelligence models have shown promise for modeling these dynamics, and their integration with night-time physiological signals could further improve their representation and interpretation. This work advances sleep dynamics modeling by combining transition prediction with physiological data to compute a novel digital biomarker from wearable recordings. Experiments are conducted on the LifeSnaps dataset, a multimodal longitudinal resource with more than four months of continuous Fitbit data from 71 participants, including sleep information and physiological signals such as heart rate, oxygen variation, and body temperature. The pipeline includes three steps: (i) training transformer-based models for minute-level next-stage prediction; (ii) using these predictions as input to a dynamics model that enhances the representation of sleep processes and enables the computation of an individual-specific biomarker; and (iii) evaluating this biomarker for stress prediction based on State-Trait Anxiety Inventory (STAI) scores, where it shows the strongest association with stress among wearable features and significantly improves predictive performance. Interpretability is ensured through SHAP, Integrated Gradients, and perturbation-based quality analysis. These results highlight the potential of wearable-derived digital biomarkers of sleep dynamics to improve stress assessment and support research on how sleep alterations may precede disease onset or progression.