Featured work
CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association
We introduce CPAgents -- an agentic framework for automatically discovering interpretable composite cardiac imaging phenotypes. It coordinates three specialised agents to generate medically and statistically motivated phenotype formulas and validate them through robustness checks. Evaluated on over 26,000 UK Biobank participants across nine diseases, CPAgents ranked first in 56 of 72 classifier-disease-metric comparisons, showing the potential of agentic systems for scalable phenotype discovery.
Published at: MICCAI 2026
🤝 MESHAgents: Multimodal Multi-Agent Reasoning in Medical AI
In this work, we introduce MESHAgents — a multi-agent framework where specialized agents reason over different clinical modalities including 3D meshes, text, and signals. Agents communicate via natural language to jointly interpret cardiac conditions, combining structure-aware perception with symbolic interaction. This approach improves transparency, modularity, and performance on clinical decision-making tasks.
Published at: MICCAI 2025
Towards Effective MLLM Jailbreaking Through Balanced On-Topicness and OOD-Intensity
We introduce Balanced Structural Decomposition (BSD) — a novel method to craft adversarial prompts that bypass safety filters in multimodal large language models (MLLMs). Unlike prior approaches, BSD decomposes malicious prompts into semantically aligned subtasks, blending relevance with subtle out-of-distribution cues. Tested on 13 MLLMs, BSD significantly outperforms existing jailbreak techniques, revealing new vulnerabilities in current safety systems.