Reliable data, clear accountability and teams equipped to question AI: three key themes from Tuesday’s “Clinical AI in the ICU: how to prepare for the inevitable” session at ESICM LIVES.

Three highlights from colleagues involved in INDICATE:

• Christian Jung, INDICATE PI, and Michael Beil, member of INDICATE’s External Expert Advisory Board, presented ten dimensions of ICU data quality: completeness, accuracy, plausibility, temporal accuracy, timeliness, consistency, conformance, granularity, uniqueness, and context/provenance. Trust in AI starts with understanding and actively managing its underlying data.

• Paul Elbers, INDICATE consortium contributor, addressed clinical decision support and human–AI interaction: where model outputs meet clinical judgement and bedside decisions.

• Michel van Genderen, INDICATE PI, emphasised a crucial principle: “Accountability must be assigned before deployment, not after harm.” Clinical trust, workflow integration and evaluation of the impact on care are essential to responsible implementation.

• Stefan Schaller, INDICATE consortium contributor, together with Davy van de Sande, from the Datahub at Erasmus MC, explored how education can address “never-skilling”, “deskilling” and “misskilling”. As we increasingly rely on AI for knowledge and decisions, we must keep questioning and seeking to understand its outputs. 

For INDICATE, these discussions reinforce the importance of bringing data infrastructure, clinical expertise and education together to support responsible AI in intensive care.