AI in a 20-centre radiology network: latency, not the algorithm, limited usefulness

Design

Retrospective multi-site observational study with technical, turnaround and survey cohorts

Population

96,874 examinations for latency, 20,909 for turnaround, and 58 radiologists in a 20-centre outpatient network

Primary outcome

AI latency, report turnaround time and radiologist adoption

Effect

Late-result rate 7.2%; adjusted turnaround lower for trauma radiography (-26%) and knee MRI (-18%)

Researchers studied a 4.5-year, multi-vendor AI programme in a 20-centre outpatient radiology network. A technical cohort of 96,874 examinations measured PACS-to-PACS latency for 10 tools from seven vendors, a cohort of 20,909 examinations compared report turnaround with and without AI, and 58 radiologists were surveyed.

Median total latency was 2.06 minutes, 72% of it from data routing. In 7.2% of examinations the AI result arrived after the report was finalised, from 3.0% for knee MRI to 13.2% for chest CT. After adjusting for radiologist, AI availability went with shorter turnaround in trauma radiography (26% lower) and knee MRI (18% lower), but not in brain volumetry MRI.

Of 58 radiologists, 91.4% reported adopting AI and 66% of those used it regularly. Satisfaction scores fell for chest CT and aorta CT in an exploratory analysis that did not survive correction for multiple testing. The cost estimates come from the authors' own network and may not travel.

• AI results arrived after the report was finalised in 7.2% of examinations.
• Data routing accounted for 72% of the median 2.06 minutes of latency.
• Turnaround was shorter for trauma radiography and knee MRI.
• Chest CT had the highest late-result rate at 13.2%.
• Check integration and routing before judging an AI tool by its accuracy alone.

Why it matters

A useful algorithm that arrives late does nothing for the report, so plumbing matters as much as accuracy.

Don't overread it

A retrospective observational study in one network, with exploratory satisfaction data, cannot show that AI improves patient outcomes.

The statistics, in plain English

An adjusted reduction in turnaround of 26% compares AI-available with concurrent non-AI workflows, so it may partly reflect which cases were routed to AI. The satisfaction changes were nominal findings, since they lost significance after correction.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.