Median reading order for hemorrhage cases: 7.25 → 1.50 (P<0.001)
Hemorrhage cases read in the first quarter of the day's list: 49% → 76% (P=0.013)
Queue order vs. arrival order (Kendall's Tau): 0.61 → 0.01. The list was worked by urgency, not by timestamp.
What did not change: reader accuracy, 0.99 before and 0.99 after (P=0.343).
That is the point. Neither radiologist had fellowship training in neuroradiology, and both were already reading at 0.99. Accuracy was never the gap at a hospital with no neuroradiologist down the hall. Sequence was.
One column in the worklist. One day of training. No change to the reading environment.
Study: Kim J, et al. Diagnostic and Interventional Radiology, 7 July 2025. PurpleAI did not fund the study.
Kim_2025_Diagnostic_Interventional_Radiology_ICH_AI_Workflow.pdf