Randomizing by site and by day, not by patient — a 40,000-patient pragmatic trial of AI-assisted liver CT reading with missed lesions as the primary measure — a clinical trial (ClinicalTrials.gov)

1You cannot switch it case by case

Turning AI on for one patient and off for the next is not something a reading room can actually do. Once a radiologist has read with the overlay in view, the next unaided read is no longer purely unaided. So this trial moves the unit of randomization from the patient to the center and the day. On one day a site reads with AI assistance, on another without it, and the whole workflow switches together. That is what a trial calling itself pragmatic looks like.

2Fixing the order in which AI appears

1. 1Initial readThe first-line radiologist assesses the study without assistance
2. 2SaveThat initial assessment is finalized and saved
3. 3RevealOnly then are the AI results shown
4. 4Final reportThe final report may be revised as needed

The order is a guard against the pull toward changing one's own judgment to match what the machine shows. See the AI first and the unaided read never enters the record at all. Fix the order, and both judgments survive: the one made alone and the one made after seeing the output, so what changed and where can be checked afterwards.

Of the 803 records this site holds as of 2026-09-04, only 5 (0.6%) describe finalizing an unaided read before the assistance appears.

3The primary measure is not benefit but misses

The primary measure is whether clinically significant malignant liver lesions are missed more often. Not a favorable metric like the number of lesions found, but an unfavorable one placed at the center. If assistance raises detection while also raising misses, the point of deploying it is gone. And the non-inferiority margin, per the record, is fixed in the statistical analysis plan before enrollment opens.

A non-inferiority trial whose margin can be chosen afterwards can be made to conclude almost anything.

4What a target of 40,000 buys

A target of 40,000 sits in the top 2% of the 797 records that state one. Missing a clinically significant malignancy is a rare event to begin with, and showing that a rare event does not become more frequent takes a large denominator. Secondary measures cover lesion detection and characterization, and then downstream clinical management and reporting efficiency.

Measuring whether reporting gets faster shows the trial has moved past the rung of accuracy and into the rung that asks whether it holds up inside daily operations.