AI in the emergency department: promising, powerful but still unproven

A new study in Science reveals that AI can now outperform doctors at diagnosing patients in emergency departments.
The AI was compared against two doctors’ diagnosis rates at the triage stage, achieving 67% accuracy compared to doctors’ 50% and 55%.
The study’s relevance in medical practice stems from its use of genuine clinical text from real emergency department records.
The AI’s performance raises questions about its potential to aid doctors in considering a wide range of possible diagnoses to prevent serious condition misses.
There are cautions about relying too heavily on AI; it works from written text and doesn’t engage directly with patients.
The gap between possible diagnoses and improved patient outcomes is a concern, along with issues regarding benchmarks used in AI training.
The critical question is how AI should be correctly tested and governed in real clinical settings, especially given that many doctors in the UK are already using AI tools daily.
The study suggests a careful approach to implementing AI: trialing in clinical settings, using it as supplementary rather than substitutive, and measuring it against actual patient care improvements.