Agenda Day 2: AI for Health Sciences

Abstract:
Multi-agent systems are taking the world by storm, promising high success rates in many entry-level (and not so entry-level) tasks. And yes, they sometimes work. But is “sometimes” really enough? If not, how can we evaluate the quality of a multi-agent system ? Not only in terms of accuracy, but also with respect to biases, cultural orientation, robustness, and fairness? And what is the best way to integrate them into our daily workflows, especially when human-to-agent collaboration is involved? In this talk, I will discuss how to test and govern multi-agent systems so that we can truly leverage their potential while minimizing their risks, particularly in critical domains.

Abstract:
Agentic AI systems—autonomous, goal-driven agents that act across complex workflows—are rapidly emerging as a transformative paradigm. Yet, designing such systems without considering the broader environment in which they operate risks fragility, inefficiency, and unintended consequences. This talk argues that Systems Thinking provides a powerful lens for building resilient and adaptive agentic workflows. In the first part, we will introduce the foundations of Systems Thinking: interconnectedness, feedback loops, delays, emergent behavior, and the importance of boundaries and perspectives. Illustrated with real-world examples, this section equips participants with a conceptual toolkit for understanding complexity in socio-technical systems. The second half will apply these principles to the design of agentic AI systems. We will explore how to map agent roles within larger ecosystems, design feedback and governance mechanisms, manage temporal dynamics, and anticipate emergent outcomes. By aligning Systems Thinking with agent design, we can create workflows that are not only technically effective but also context-aware, adaptive, and robust.

Abstract:
Medical Vision-Language Models (VLMs) have demonstrated substantial promise in clinical applications, yet their deployment in real-world healthcare settings remains constrained by two critical challenges: reliability concerns and limited explainability. This talk presents our recent works addressing both dimensions through uncertainty-aware modeling and grounded explanation generation. These developments advance trustworthy medical AI systems that provide reliable diagnostic support with transparent reasoning processes essential for clinical acceptance and patient safety. Key contributions include uncertainty-based reliability mechanisms for out-of-distribution detection and hallucination prevention, grounded explanation frameworks for enhanced clinical interpretability, empirical validation across multiple medical imaging domains, and performance improvements with reduced training requirements through principled reasoning approaches.

Abstract:
The era of single-modality AI models analysing only an image or a report is ending. The field is entering the age of the Medical Multimodal Mind: foundation models that jointly understand medical vision and language. This presentation explores how these models learn a unified representation of clinical data, enabling powerful capabilities like zero-shot diagnosis, where a model can recognise conditions it was never explicitly trained on by leveraging its semantic understanding. This shift is critical. It moves AI from a narrow tool to a flexible assistant, capable of generating preliminary findings, answering clinical queries about imagery, and enhancing decision-making. The future points towards a new paradigm of human-AI collaboration, where these models act as intelligent partners, scaling expert-level reasoning and improving healthcare accessibility and accuracy globally. Finally, the presentation will cover important healthcare research projects on medical multimodal foundation models.

Abstract:
The generation of "immortal time" is a frequent blunder in survival analyses for causal inference based on healthcare databases. Immortal time explains why medical treatments often look suspiciously amazing. After two centuries of warnings, immortal time still plagues causal analyses in medicine, which is fascinating because "immortal time" doesn't exist in the data. Rather, we generate immortal time when analyzing the data incorrectly. This talk summarizes why immortal time arises in causal survival analyses of healthcare databases and how to prevent it.

Abstract:
This talk introduces co-explainers, an AI systems framework that moves beyond static transparency to act as interactive partners in explanation, forming a dynamic iterative framework. Drawing on a taxonomy of harms across six domains, we argue that explainability must become a sociotechnical infrastructure, enabling designers and regulators to contest, recalibrate, and align AI system behavior. Co-explainers are framed as agents of trust calibration and institutional accountability, providing pathways for AI deployment in both high-risk domains and everyday contexts. The framework redefines XAI as a vehicle for oversight and harm mitigation, supporting auditability and trust in AI for high-risk scenarios, rather than serving as a post hoc technical add-on.