Cross-Talks Workshop - Program (Session II)

29 junho 2025, 20:56 • Ana Almeida Matos

30 of June, 12:00-15:30 - Tagus @ room 0.25, or Zoom (course link)

Title: Enhancing Patient-Clinician communication throughout a surgical journey

By: Luis Gordete

Abstract: 
The patient journey faces several persistent challenges, namely fragmented communication, inadequate preparation, difficulties in collecting relevant data, and associated environmental impact. This proposes an innovative mobile application that fosters connection between patients and healthcare professionals through real-time notifications, personalized educational content, and a 24/7 chatbot. The solution ensures secure data sharing supported by blockchain, collects structured indicators—namely patient-reported outcome measures (PROMs) and patient-reported experience measures (PREMs)—and integrates gamification and artificial intelligence mechanisms to personalize the experience. Preliminary results from professionals and patients indicate benefits in terms of adherence to medical guidance, improved health literacy, and reduced cancellations. The solution demonstrates strong potential to transform the patient journey by promoting greater engagement, personalization, and sustainability. User feedback is particularly emphasized in evaluating application usability and in comparisons with groups who did not have access to the tool.

Title : From Play to Plot: Recognising Narrative Value in Emergent Games

By : Miguel Belbute

Abstract:
Emergent games are grounded on complex and unpredictable simulations, being incompatible with traditional narrative structures. This work explores how story recognition – the open design challenge of recognising relevant gameplay event sequences as stories – can be implemented as a bottom-up approach of creating narrative value in emergent games. Log data clustering is explored to identify gameplay event sequences that are sufficiently relevant to formulate narratives. Multiple clustering heuristics were considered (e.g., temporal/spatial distance, shared participants and verb importance attribution), with two major overarching architectures being developed: clustering pipelines and voting ensembles. After a relevant enough sequence is identified, the underlying logs are summarised through an Large Language Model (LLM), whose prompts were also iterated upon. The heuristics and LLM’s prompts were evaluated separately. It was found that, when compared to clustering pipelines, voting ensembles present significantly better F1-score in identifying hunting, building and social stories from game logs. To evaluate the story recognition system as a whole, two preliminary studies were performed with distinct in-game tasks and session durations. Both players familiar and new to emergent games feel like their experiences are improved when in contact with a narrative originated by story recognition, especially considering shorter gameplay sequences. Inherent noise from longer play sessions heavily impact the system’s accuracy, suggesting story recognition should be applied in shorter periods of gameplay. Study participants also mentioned the concept of story support, underlying their desire for the recognised stories to be naturally incorporated into gameplay.

Title: Leveraging machine learning to enhance pipelines' resilience: insights from failure analysis

By: Ana Silva

Abstract:
Pipelines are critical to the transportation of energy, but their vulnerability to failure presents ongoing challenges for operators and engineers. Traditional approaches to pipeline maintenance often rely on periodic inspections and corrosion-focused models, which can miss subtle or emerging risks. This study explores how machine learning can transform the way pipeline infrastructure is managed. By analyzing over 12,000 historical pipeline incident records from 1970 to 2023, the study develops a predictive model that identifies patterns in pipeline degradation and forecasts potential failure points. The model provides a more comprehensive understanding of risk, allowing for proactive, data-driven maintenance strategies. This study integrates diverse engineering perspectives, with the collaboration of civil, geo-resources, mechanical, and aerospace engineers, thus offering a holistic approach to improve pipelines’ resilience, increasing safety, and minimizing downtime across the industry.

Title: Data Consistency in Microservices: Challenges and Tool for Analyzing Applications
By: Mafalda Ferreira
Abstract:
The microservice architecture is a modern software paradigm that decomposes applications into small, loosely coupled components. In contrast to a monolith system, this approach provides benefits such as independent deployment, increased scalability, and improved fault tolerance. However, as microservices are decomposed, each service becomes accountable for its own business functionality, data schemas, and underlying datastores. As a result, data decentralization causes applications' schemas to be partitioned across datastores employing heterogeneous consistency models and distinct replication mechanisms, leading to potential violation of integrity constraints. This work studies these challenges, categorizes different integrity constraints, and introduces a static analysis tool to automatically detect problematic executions that can result in inconsistencies at runtime.

Title: H3-Based Geographic Peer Discovery for Privacy-Preserving Decentralized Storage
By: Hugo Duarte

Abstract:
Decentralized storage networks distribute encrypted data chunks across multiple independent nodes, improving fault tolerance, censorship resistance, and user control. Efficient peer discovery is crucial for locating storage providers, and structured overlays like Kademlia are widely used due to their scalable, logarithmic routing based on XOR distance. However, XOR distance does not account for physical network latency, often causing inefficient lookups and data transfers. To address this, we incorporate Uber’s H3 spatial indexing to organize peers geographically within the Kademlia routing table, enabling locality-aware peer selection without revealing precise user coordinates. Additionally, zero-knowledge proofs for location membership are explored as a privacy-preserving means to enforce geographic data residency policies, adding interdisciplinary depth to the system. Despite these improvements, geographic proximity does not guarantee low latency due to complex network routing conditions, highlighting the need for future work integrating real-time latency measurements and latency-aware routing overlays. This research lays foundational steps towards more efficient, privacy-conscious, and regulation-compliant decentralized storage architectures.

Title: Enhancing ChatSpecies: Designing AI-Integrated Chatbots to Enhance Species Knowledge and Foster Ecological Literacy

By: Ying Xu

Abstract:
The growing disconnect between humans and the natural world has led to a limited understanding of biodiversity, deterring environmental conservation efforts. Digital technologies have shown promise in supporting ecological education through playful, interactive approaches that engage users in learning about local ecosystems and species. However, existing approaches often fail to create meaningful, memorable learning experiences that foster emotional connections with biodiversity. Here we show that AI-integrated chatbots designed with playful elements and species-first perspectives can significantly enhance biodiversity learning engagement while building empathetic connections with local flora and fauna. Our exploratory study with 21 participants testing the ChatSpecies chatbot revealed three key design insights: fact-checking mechanisms foster trust and information transparency, suitable playful elements enhance learning effectiveness, and sympathetic mentality promotes emotional ties with species. This work provides immediate pathways for museums and educational institutions to leverage AI technologies that bridge the gap between urban societies and the natural world, ultimately supporting greater ecological literacy and environmental stewardship.

Title: The way it looks: How the gaze of a robotic teammate impacts multiparty dynamics

By: Sandra Andrade

Abstract:
As human-robot teams become more common, optimizing team dynamics is crucial to improve collaboration. More anthropomorphic robots are mainly used in decision-making tasks, while less anthropomorphic are assigned physical roles. Since both types increasingly collaborate with humans, designing intelligent, adaptive robots that integrate smoothly into diverse team environments, regardless of their embodiment, is essential.
Additionally, robotic gaze has been widely studied in dyadic settings, showing humans respond to it similarly to human gaze, interpreting it as a cue for attention, intention, and persuasion. These responses enhance likability and agency. In teamwork, robotic gaze also improves perceptions of adaptability and intelligence. Despite this, research on robotic gaze in multiparty teamwork (i.e., more than two members) is limited due to technical complexity.
This proposal addresses these two gaps by studying robotic gaze in multiparty teamwork settings across different embodiments, using the Input-Process-Output model and user-centered evaluations to develop novel gaze models.

Title: Thinking Through the Interface: Using XR and LLMs to Teach Algorithms Without Losing Critical Thought
By: Inês Alves
Abstract: This research approaches an emerging issue in engineering education: how can we use immersive technologies and AI to make abstract concepts, such as algorithms, more accessible while still preserving the critical and reflective thinking essential to engineering practice? We answer this through an ongoing systematic review and a practical case study, both of which explore how Extended Reality (Virtual Reality/ Augmented Reality/ Mixed Reality) environments have been applied to algorithm education. Many systems visualize algorithm processes or simulate engineering contexts, but few actively support deep engagement or metacognitive reflection. At the same time, new tools such as Large Language Models (LLMs) offer exciting opportunities to guide learning in personalized and interactive ways. Yet, there is growing concern that such tools may weaken cognitive effort by offering quick answers over meaningful struggle. In this talk, we present early design principles aimed at promoting thoughtful engagement and critical reasoning. We invite interdisciplinary discussion on how to design systems that are not only engaging and adaptive but also stimulate thinking, drawing insights from cognitive science, AI ethics, HCI, and education research.

Title: Machine Learning applied to Forensic Age Estimation
By: Mohamed Elbawab

Abstract:
Machine learning is a subtype of artificial intelligence where the machine helps to predict outcomes. One of the medical applications is to use machine and deep learning technologies in forensic odontology. The research goal is to help with human identification, where human identification can be applied in disaster victim identification scenarios. This help could be done through predicting the age of people using x-rays. The aim of this research is to develop a machine learning model that predicts the human age. The model will also need to identify the gender of the participant. Our focus is on developing a model that predicts the age while still using the ethical guidelines of the European Union and Portugal.

Title: Multivariate Analysis of Metric Spaces Defined by Embedded TCGA Pathology Reports
By: Inês Duarte

Abstract:
Recently, a batch of 9,523 pathology reports from The Cancer Genome Atlas (TCGA) was processed as a Machine-Readable public resource for Benchmarking Text-Based AI Models. This invaluable public reference resource provides a unique opportunity to investigate the mathematical properties of embedded spaces created by operating generative AI models with real-world medical data. Recognizing pathology reports as vital sources of clinical data often not captured in structured datasets and their central role as an integrative representation for myriad analytical trajectories elucidating disease progression, this study leverages Google's Gemini API to generate 768-dimensional text embeddings from these reports. This also presents an opportunity to explore the computational statistics foundations of multimodal generative AI models to project diverse data types into a shared multimodal embedded space. These embeddings convert complex textual information into numerical vectors, hypothesized to reveal underlying differences between cancer types.
The methodology involved preprocessing the reports, and applying dimensionality reduction techniques such as PCA, UMAP, and t-SNE for 3D visualization. The analysis, performed entirely in the browser to ensure data privacy, revealed distinct clusters corresponding to different cancer types. Notably, some clusters contained multiple cancer types, suggesting potential common origins or metastatic patterns. The findings indicate that Gemini's embeddings effectively capture the unique characteristics of various cancer types.

Title: Spiking Neural Networks and the future of AI inference on the edge.

By: Rishi Tripathi

Abstract:

Spiking Neural Networks (SNNs) are a class of brain-inspired models known for their energy efficiency and event driven computation. These networks show immense potential for low power, real time applications, especially in neuromorphic and edge computing. In my talk, I’ll share how my research addresses key challenges in SNNs through advanced weight quantization techniques that significantly reduce model size and power consumption. I’ll also delve into memory and caching strategies designed to streamline spike-based data flow and minimize latency. Together we will look towards a path toward scalable, high-performance SNNs through efficient design and hardware-aware optimization.