What AI Triage Is Actually Doing to Radiology Workflows

Will Drewes

Founder, Fern Strategy · 3 min read

The numbers coming out of radiology are worth paying attention to

Most AI coverage in healthcare leads with accuracy metrics. Sensitivity, specificity, AUC. Those matter clinically, but they don't tell a department head whether the investment changes throughput or staffing pressure. A few recent deployments and studies do report operational outcomes, and they're concrete enough to be useful.

Northwestern Medicine: productivity at scale

Northwestern Medicine reported up to 40% productivity gains for radiographs and 80% efficiency improvements for CT scans in studies cited in 2025 coverage of their AI-powered triage and workflow prioritization rollout. Those figures are unpublished, so treat them as directional rather than peer-reviewed. But Northwestern is a large academic health system with real volume, and a 40% lift on plain films is not a rounding error. If it holds under scrutiny, it means radiologists are reading meaningfully more per shift, or reading the same volume with less queue pressure.

Worklist reordering and turnaround time: what a systematic review found

A systematic review of AI triage across modalities found that algorithms which reordered the worklist (rather than just flagging studies) produced the largest turnaround reductions. Specific figures: mean 12.3 minutes faster for pulmonary embolism, 20.5 minutes for stroke, 4.3 minutes for intracranial hemorrhage, and 29.7 minutes for chest disease. For stroke and PE, those are not cosmetic improvements. Door-to-treatment time in stroke directly affects outcomes. Cutting 20 minutes off the read queue for a suspected stroke patient is a clinical outcome, not just an operational one.

Triaging normal studies: the volume play

A tertiary referral hospital deployed an AI triage system on ED and ICU follow-up films. The system triaged 40% of no-change X-rays out of the active read queue, with 88-90% accuracy for detecting changes. That's the other side of the efficiency equation. Radiologists spend a significant portion of their day on follow-up films that show no meaningful change. Pulling 40% of those out of the queue automatically, at 88-90% accuracy, changes what the remaining read time looks like. The hospital name wasn't disclosed in the published data, but the operational logic is clear.

A separate real-world chest X-ray deployment reported 89% sensitivity for normal CXRs and 93% specificity, with significantly reduced turnaround times across all patient subgroups. Again, no named institution in the abstract, but the figures come from a published real-world study, not a vendor whitepaper.

What this means for radiology and health system leaders

The pattern across these cases is consistent: AI triage earns its keep on high-volume, rule-stable tasks (normal follow-up films, worklist ordering by acuity) and on time-sensitive critical findings where queue position has direct clinical consequences. The productivity gains at Northwestern and the turnaround reductions in the systematic review are both downstream of the same mechanism: the system makes prioritization decisions that previously waited on a human.

The governance question that doesn't get enough attention is what happens when the triage algorithm is wrong. At 88-90% accuracy on a high-volume ED workflow, you're still misclassifying a meaningful number of studies. That's not an argument against deployment; it's an argument for knowing exactly which studies the system touches, what its error modes are, and who is accountable when a no-change classification is wrong.

Before any health system commits to a radiology AI build or a long-term platform contract, a two-week structured audit of the workflow, the data pipeline, and the failure modes is worth doing. The operational gains are real. So is the liability surface if the system runs without clear accountability boundaries.