MERLIN

Multi-agent systems for real-time clinical decision support in the ICU

Description: The intensive care unit is a high-pressure environment in which clinicians must combine large amounts of multimodal data under time pressure. Detecting delayed cerebral ischemia after subarachnoid hemorrhage, for example, requires neuroimaging, hemodynamic monitoring, and laboratory biomarkers to be interpreted together. Yet most clinical AI tools serve a single purpose and are poorly integrated with one another, leaving staff to work across siloed devices and dashboards. Multimodal foundation models have shown promise on medical benchmarks, but they are trained largely on retrospective, structured data, do not handle continuous data streams, and their computational and privacy requirements make hospital deployment difficult.

MERLIN takes a different route. A multi-agent system in which specialized AI models and software tools are coordinated by LLM-based planning agents that manage data flow, delegate tasks, and generate user-specific outputs. Agents can query existing systems and trigger data processing pipelines, giving clinicians a single point of interaction instead of many. The aim is not another predictive model but the orchestration layer around the models that already exist. First multi-agent pilots in the ICU — for instance in sepsis management — suggest gains in efficiency and workload reduction, but rigorous evaluation in real-time settings is still missing.

The project builds on the ICU Cockpit, our research IT platform operated by the Neurocritical Care Unit at the University Hospital Zurich since late 2016, which has processed over 1.8 trillion data points from more than 3,000 patients and hosts live algorithms for DCI risk prediction, neurologic prognostication, false alarm reduction, and seizure prediction. Together with the unit’s experience in algorithm evaluation and human-AI teaming, this provides a setting in which multi-agent systems can be evaluated prospectively on live data without affecting clinical decision-making.

Partners: UZH, ZHAW, USZ Institute of Intensive Care

Funding: DIZH

Team: Dr. Tilman Beck (UZH), Dr. Farhad Nooralazadeh (ZHAW), Dr. med. Jan Willms (USZ)