AI-driven emergency radiology triage — continuous worklist scoring, critical-finding detection, and instant on-call notification
Every CT and radiograph acquired in a busy emergency department lands in the PACS (Picture Archiving and Communication System) worklist the instant the scanner finishes reconstruction. Without triage intelligence, that worklist is ordered by arrival time alone — meaning a catastrophic intracranial bleed acquired during a busy overnight shift can sit, unread, behind ankle films and routine chest X-rays for well over an hour.
A standard PACS worklist is, by default, agnostic to clinical severity. A subdural hematoma expanding in real time and a stable ankle fracture are functionally identical line items — both are "unread studies" competing for the same radiologist attention, ordered only by when the scanner produced them.
In a Level I trauma center running 24/7, a single overnight radiologist (or an on-call teleradiology group) may face 40–80 unread studies at any given moment. Because reading happens roughly in queue order, a critical study acquired between two batches of low-acuity exams inherits whatever wait time the queue happens to have at that moment — regardless of its actual urgency.
This is the structural gap that AI triage exists to close: not by replacing the radiologist's interpretation, but by re-ranking which study reaches their screen first.
Emergency imaging streams are heterogeneous. A typical feed interleaves non-contrast head CT (stroke/trauma rule-out), CT angiography of the chest (PE, aortic dissection rule-out), portable chest radiographs (pneumothorax, line placement), and CT abdomen/pelvis (trauma, acute abdomen) — each with different acquisition times, slice counts, and clinical stakes.
At high-volume centers, image volume itself is a bottleneck: a single non-contrast head CT generates 150–300 axial images; a CT angiogram can exceed 1,000 images across reconstructions. AI triage systems ingest the DICOM series the moment it is pushed to PACS, well before a human has opened the study, so scoring latency is bound only by inference time, not by radiologist availability.
As each study reaches PACS, an AI triage engine pulls it automatically and runs inference in parallel with every other study in the queue — no manual submission required. Modern deep learning triage tools return a probability score in under a minute, turning a passive archive into an actively monitored stream.
Most deployed emergency-imaging triage tools use convolutional neural networks (CNNs), often 3D architectures adapted from 2D backbones like ResNet or DenseNet, trained on tens of thousands of labeled studies. Some newer systems layer transformer-based attention over CT slice sequences to better capture cross-sectional context (e.g., tracking a hyperdense clot sign across contiguous slices).
The model does not attempt a full diagnostic read. It solves a much narrower binary or multi-label classification problem: "does this study contain finding X, yes or no, with what probability?" That narrow scope is precisely what allows sub-minute inference and very high sensitivity — the task is orders of magnitude simpler than generating a full radiology report.
Because inference runs on GPU-accelerated infrastructure independent of any single radiologist's screen, the AI can score every study in the queue simultaneously rather than sequentially. A queue of 18 studies does not take 18× longer to triage than one study — it takes roughly the same wall-clock time as a single inference pass, batched across available compute.
This is the structural advantage over human review: a radiologist reads one study at a time, in order; the AI evaluates the entire queue continuously and re-evaluates it every time a new study arrives, so urgency ranking never goes stale.
Triage algorithms are deliberately tuned toward high sensitivity (catching nearly every true critical finding) at some cost to specificity (accepting a higher false-positive rate). This is a clinically sound trade-off for a triage tool, whose job is only to reorder review priority — a false positive costs a radiologist a few extra seconds of a normal read moved up the queue; a false negative costs a delayed diagnosis on a genuine emergency.
Regulatory clearance reflects this framing precisely: FDA has cleared these tools as "computer-aided triage and notification" (CADt) devices, explicitly not as autonomous diagnostic devices. The radiologist remains the sole decision-maker; the AI only changes the order and speed at which studies reach them.
Behind the single triage score sits a bank of pathology-specific detectors, each trained and validated against one target condition: intracranial hemorrhage, pneumothorax, aortic dissection, pulmonary embolism, and others. Each detector fires independently, and a study can trigger more than one flag simultaneously.
Rather than a single generalized "abnormal vs. normal" classifier, production triage suites typically bundle several narrow detectors: an ICH detector trained on non-contrast head CT, a pneumothorax detector trained on chest radiographs, a large-vessel-occlusion (LVO) detector trained on CT angiography, and a PE detector trained on CT pulmonary angiograms. Aortic dissection detectors look for the intimal flap and true/false lumen pattern on contrast-enhanced CT chest.
Each detector is validated on its own labeled dataset with its own sensitivity/specificity operating point, because the imaging appearance, prevalence, and consequence of a missed case differ enormously between, say, a small pneumothorax and a ruptured aortic dissection.
"Time is brain" is not a slogan — it is a measured rate. In large vessel occlusion stroke, an estimated 1.9 million neurons and 14 billion synapses are lost for every minute that a proximal cerebral artery remains occluded (Saver, 2006). For acute type A aortic dissection, registry data (IRAD) puts untreated mortality at roughly 1–2% per hour during the first 24–48 hours — meaning a diagnostic delay measured in hours, not days, materially changes survival odds. Tension pneumothorax can progress from stable vital signs to obstructive cardiac arrest within minutes as intrathoracic pressure rises.
These pathologies do not wait for a radiologist to reach the bottom of a chronological queue. Detection has to be simultaneous with acquisition, not sequential with reading order.
The Alert Threshold control in this simulation reflects a real operating decision every deployment makes: a lower threshold flags more studies (higher sensitivity, more false alarms competing for attention); a higher threshold flags fewer studies (higher precision, but some true positives fall back into the ordinary queue). Most FDA-cleared triage tools report sensitivities in the 90–96% range for their target finding at their chosen operating point, with corresponding false-positive rates that hospitals tune against their own alert-fatigue tolerance.
Once a study is flagged, the PACS worklist re-ranks it above studies that arrived earlier but carry no critical-finding flag. The queue transforms from a strict first-in-first-out list into a priority queue, where clinical urgency — not scanner timestamp — determines what the radiologist sees next.
Technically, reordering is implemented as a priority-queue insert: the PACS worklist server maintains studies in a data structure ranked by a composite key of (flag status, AI confidence score, wait time), rather than raw accession timestamp. When a new critical flag arrives, the study is re-inserted near the head of the queue, and the radiologist's worklist view — typically refreshed every few seconds — reflects the new order automatically, without any manual queue management.
Multiple flagged studies are themselves ranked relative to one another, often by a combination of AI confidence and estimated clinical severity, so that (for example) a high-confidence tension pneumothorax is not buried behind three lower-confidence flags.
A frequently cited 2018 study (Arbabshirani et al., npj Digital Medicine) deployed a deep learning ICH-detection and worklist-reprioritization system across a real hospital system's CT stream. The result: mean time from image acquisition to radiologist notification for a positive ICH case dropped from roughly 512 minutes to about 19 minutes — a reduction of approximately 96% — simply by moving flagged studies to the front of the queue rather than changing how the radiologist read them.
This kind of result illustrates that most of the achievable time savings in emergency imaging triage comes from queue position, not from any change to interpretation accuracy or speed at the individual-study level.
Reordering the worklist helps only if someone is looking at it. The final step pushes an active notification — a mobile app alert, secure text, or pager message — directly to the on-call radiologist's device, interrupting their current workflow for studies that cannot wait for the next natural break in reading.
A well-designed critical-finding alert is deliberately minimal: patient identifier, exam type, suspected finding, AI confidence, and a direct deep link into the PACS viewer at the relevant series/slice. It is not a diagnosis — it is a summons. The radiologist still performs the definitive interpretation; the alert simply ensures that interpretation happens in minutes rather than potentially an hour or more.
Many deployments route the same alert, or an abbreviated version, to the ordering ER physician in parallel, so that clinical teams can begin acting on a strong suspicion (e.g., activating a stroke or trauma protocol) even before the radiologist has formally signed the report.
Tools in this category are cleared by the FDA under the "computer-aided triage and notification" (CADt) framework, a distinct regulatory lane from diagnostic AI. Cleared examples include Viz.ai's LVO and ICH triage modules, Aidoc's intracranial hemorrhage and pulmonary embolism triage products, and RapidAI's stroke imaging suite — each validated against retrospective and, increasingly, prospective clinical datasets against radiologist-adjudicated ground truth, with reported sensitivities generally in the low-to-mid 90% range for their target finding.
Clearance requires demonstrating that the notification reduces time-to-notification without an unacceptable increase in false alerts, since the clinical value proposition rests entirely on speed, not on replacing the radiologist's read.
Every additional automated alert carries an attentional cost. If false-positive flags are too frequent, on-call radiologists begin to habitually deprioritize or silence notifications — the exact failure mode the system was built to prevent. Sites tune the Alert Threshold specifically to balance this: too sensitive, and true emergencies drown in noise; too specific, and genuine critical findings slip back into the ordinary queue.
Liability questions also shape deployment: because the AI never issues a diagnosis, documentation of the alert (timestamp, confidence score, delivery receipt) becomes part of the medico-legal record, and institutions generally require an explicit human acknowledgment step to close the alert loop.