Comprehensive Summary
This study tested an AI system to help neurologists decide which emergency department patients need hospital admission and predict death risk. The researchers built a system combining a large language model (Gemini 1.5-pro) with traditional machine learning methods, then tested it on 1,368 patients who had neurological consultations from January to February 2024. The AI achieved strong performance with area under the curve scores of 0.88 for predicting general hospital admissions, 0.86 for neurological department admissions, and 0.91-1.0 for predicting death at different time points from 48 hours to 230 days. When three senior neurologists reviewed 100 cases, the AI's predictions matched their combined judgment well (correlation of 0.79), even though the doctors often disagreed with each other about admission decisions. The authors noted that such AI tools could help reduce cognitive biases in clinical decision-making and improve patient outcomes.
Outcomes and Implications
Emergency departments are overcrowded, and admission decisions are often inconsistent, leading to higher costs and potential safety problems, especially in neurology, where specialists are scarce. This AI tool could provide consistent, objective support for critical decisions about patient admission and mortality risk in neurological emergencies. The system processes both structured data and physician notes to give transparent recommendations, making it clinically useful for real-world implementation. However, the study was done at only one hospital over two months, so larger studies across multiple centers are needed before this technology can be safely used in clinical practice.