Pending QC vs Worklist Smart Routing

The worklist has no state for "not ready"

Start with the data model routing operates on. DICOM's Unified Procedure Step service is the standard's worklist object: one item holds both the step that was scheduled and the record of the step as performed. Its Procedure Step State attribute takes exactly four values — SCHEDULED, IN PROGRESS, COMPLETED, CANCELED. There is no held, blocked, or waiting-on-a-prior value, and no route back: setting an item's state to SCHEDULED fails from every other state, so a step that turns out to have been premature is canceled and the remaining work has to be scheduled as a different item.

Real worklists layer their own columns on that, and some of them model a hold. The point is what the routing layer is native to. Its vocabulary is scheduling and progress: when a step is due, who claimed it, whether it finished. Readiness is a different kind of statement — that the order, modality, body part, and laterality agree with what was performed, that the reason for exam will support this case type, that referenced priors arrived, that holds and open remediation are closed. What makes a study read-ready walks that list, and the access it takes to satisfy it — tags, statuses, source corrections — is why PACS admins own the last mile rather than the reading roster.

10:12, and distribution is not what failed

A lumbar MRI finishes at 10:12. Images are in PACS. The order reads L-spine without contrast. The reason for exam is blank. The tech note references a prior from an outside facility that never arrived. The study sits in Pending.

At 10:40 someone asks whether the worklist needs smarter auto-assignment. The study never reached distribution, so distribution is not the thing that failed. Handing it to the next available reader does not produce a reason for exam, retrieve the prior, or fill the field that both the report and the claim will need. It relocates an unfinished record one step closer to a radiologist.

What reordering actually buys

Routing is unusually well measured, because AI triage products are routing under study. The most useful result is not the headline minutes but which interval they came out of.

In that study, PE-positive exams went from a mean 59.9-minute report turnaround to 47.6 minutes, and the wait interval — exam completion to report initiation — went from 33.4 to 21.4 minutes. Read time did not move. The tool reprioritized 12.7% of exams in the post-implementation period, and during regular operational hours the wait-time improvement showed up for routine-priority exams (15.3 vs 43.7 minutes) but not for stat or urgent ones, which were already near the front. One center, one retrospective comparison, one finding type, so treat the magnitudes as local. The mechanism generalizes: reordering changed when a study got attention, not the work involved in it.

A 2025 preprint from FDA and University of Chicago authors asks when that mechanism has anything to work with. At their site, CTPA turnaround fell from a mean 68.9 to 46.7 minutes during weekday work hours, but only 44.8 to 42.0 minutes off-hours, which was not significant. Their queueing model reproduced both, and adding a fourth radiologist to the work-hours shift drove the predicted benefit toward zero. Preprint, one site, but the direction is intuitive: reordering pays where studies are genuinely queued behind each other and pays nothing where capacity absorbs arrivals as they land. That is a real thing to buy. It is also the whole of what routing does, and none of it touches a study that should not have entered the queue.

"Deprioritize the incomplete ones" is a tail-latency plan

The usual compromise is to feed incompleteness into the ranking instead of gating on it. Incomplete studies stay in the pool and sort lower. Ranking has a known failure mode when the signal misses.

A Hamburg group simulated their chest radiograph workflow over roughly 11,000 simulated days, comparing first-in-first-out reading against AI urgency ranking. Averages dropped hard — pneumothorax 80.1 to 35.6 minutes — while worst cases went the other way: maximum turnaround rose from 890 to 1,178 minutes for pneumothorax and from 916 to 2,018 minutes for congestion, because studies the model scored negative sorted to the back of the list. Retuning for a low false-negative rate made the pneumothorax worst case worse, not better. What helped was a rule from outside the ranking: promote anything that has waited past a threshold.

Different question, simulated, single site. The transferable shape is that a ranking signal's misses do not disappear, they become tail latency. Apply it to readiness and the arithmetic is worse than in the imaging case, because a deprioritized incomplete study is not merely late — it is unfixed. No one was assigned the missing prior. It ages until a reader opens it and stalls, or until it surfaces as an aging exception at the point where someone has to explain the turnaround number. A gate produces the artifact ranking cannot: a named next step attached to a named person, with the study held out of the reading path until that step closes.

Two intervals, two owners

Practices that run both well keep the sequence explicit, with a handoff that has a state on each side:

1. Pending QC clears the study as read-ready, or holds it with issue context and an owner.
2. Only read-ready studies enter the assignment pool.
3. Routing, acuity tiers, sub-specialty pools, and load balancing operate on that pool.
4. Final reader assignment stays a reading-workflow decision, downstream of the QC record.

Measure the two intervals separately: exam complete to read-ready, then read-ready to report initiation. A blended turnaround number averages a missing prior together with a busy MSK reader and tells you neither. Splitting them is also how you find out which lever your money should buy, which is the point the turnaround time work eventually forces.

One pass through last week's aging studies, sorted by cause rather than by reader, usually settles it:

• Stuck on completeness, classification, priors, or paperwork → readiness gap.
• Complete and correctly described, still waiting on a reader pool → routing or staffing gap, and the evidence above says routing helps most when this bucket is deep.
• Complete but opened and returned by readers → your readiness criteria are softer than the reading desk's.

Where Corena fits

Corena works studies in a PACS Pending queue as operational QC before reading. Checks run against versioned site and case policy; a check it cannot evaluate stays unknown rather than quietly passing; findings carry the evidence behind them; and a proposed correction or handoff waits for a human to approve it before anything is written back to a source system. The internal sequence is deliberately unforgiving about what counts as done: a recommendation is not an approval, an approval is not a source write, a write response is not a verified read-back, and a verified action is not a Pending exit. Assignment sits past the end of all of that.

So the buying question is not which product has smarter queues. It is which object the product acts on: the record, or the order of the queue. If a demo opens on auto-assignment and never shows a blocked study with an owner and a next step, you are looking at routing. Request a demo to see what the gate does instead.