UAB AI triage system for intracranial hemorrhage detection shows no improvement in radiologist accuracy or turnaround times

Intracranial hemorrhage requires fast, accurate detection to guide life-saving treatment decisions. UAB researchers sought to rigorously evaluate whether a commercial AI triage tool could meaningfully improve radiologist diagnostic performance and report turnaround times in a real-world, prospective setting — addressing gaps left by prior retrospective studies.

The UAB Heersink School of Medicine deployed Aidoc's commercial AI triage system across nearly 10,000 noncontrast head CT scans in 2021. The tool processed CT exams and notified radiologists of positive intracranial hemorrhage findings via a floating widget pop-up outside the standard worklist. Emergency and neuro radiologists interpreted images across two phases: pre-AI (May–June 2021) and post-AI (September–December 2021).

Radiologist accuracy for ICH detection slightly decreased with AI (99.2%) compared to without it (99.5%), and specificity also dropped from 99.8% to 99.3%. Average report turnaround time for positive exams increased from 147.1 minutes without AI to 149.9 minutes with AI. Radiologists alone outperformed the AI alone for ICH detection.

• High-performing specialist radiologists may not benefit from AI triage tools designed primarily as safety nets for less specialized workflows.
• Widget-based AI notification systems that sit outside the radiologist's worklist may disrupt workflow rather than accelerate it.
• Prospective real-world evaluations frequently contradict positive findings from retrospective studies, highlighting the importance of rigorous study design before broad AI adoption.