What Clinicians Want AI to Measure, and How They Want It Done

This qualitative study explores healthcare professionals' perspectives on AI tools for patient assessment through interviews with 28 clinicians across 14 countries. We identified two fundamental components of clinical assessment: baseline health status (a patient's pre-illness state) and acute illness severity. While existing tools often measure these separately, clinicians integrate both when making decisions. Participants demonstrated evolving attitudes toward AI, prioritising reliable performance and workflow integration over understanding internal algorithms. They emphasised tools must function with incomplete data, support clinical judgement rather than replace it, and integrate seamlessly with existing systems. Key requirements included minimal data entry, customisable alerts based on experience level, and evidence of improved outcomes. Our findings reveal important implementation trade-offs across five domains: data selection, model architecture, integration approach, alert system design, and transparency level. Each decision balances benefits like wider applicability against drawbacks like potential lower precision. Successful AI tools must bridge the gap between technical capability and clinical utility—measuring what clinicians need measured while fitting naturally into healthcare workflows. As healthcare systems face growing challenges from ageing populations with multiple conditions, AI tools that assess overall patient health rather than specific diseases could provide valuable support, if designed to meet clinicians' practical requirements.