Domain Atlas / Clinical decision support & deterioration alerting
Viz.ai LVO Stroke Triage
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 3 assumed · 6 published baseline.
On 13 February 2018 the FDA granted De Novo request DEN170073 for Viz.AI's ContaCT, creating the radiological computer-aided triage and notification device class (21 CFR 892.2080, product code QAS) that later stroke-AI entrants reach by 510(k). The FDA decision summary describes the device as a notification-only, parallel workflow tool: it analyses CT angiograms and alerts a neurovascular specialist 'in parallel to standard of care image interpretation', identification of suspected findings is not for diagnostic use beyond notification, and the benefit-risk section concludes 'there are no major risks for the device because the device operates in parallel to the current usual standard of care'. The three risks the order does list are triage risks: deprioritization of other patients' images, inappropriate use for primary interpretation, and delayed management if prioritization fails. In the pivotal standalone study of 300 CT angiograms from two US sites the algorithm measured 87.8% sensitivity (95% CI 81.2-92.5), 89.6% specificity (83.7-93.9) and an area under the curve of 0.91; in the same record notification arrived a mean 51.4 minutes earlier than the standard radiology pathway (7.3 against 58.7 minutes; medians 5.6 against 51.5; earlier in 42 of 44 cases) — a figure the FDA's own summary flags as potentially biased because it was computed on the 44 true-positive cases that happened to have documented standard-of-care notification times, and which this atlas therefore carries as that study's measurement rather than a field-wide constant.[2]
What happened
On 13 February 2018 the FDA granted De Novo request DEN170073 for Viz.AI's ContaCT, and in doing so created a device category that had not existed: radiological computer-aided triage and notification software, 21 CFR 892.2080, which later stroke-AI entrants reach by 510(k). The marketed system is three parts — Viz LVO, the detection algorithm; Viz Hub, the messaging platform; Viz View, the mobile image viewer — and the FDA's decision summary describes what they do together in a phrase that governs everything else in this file: a notification-only, parallel workflow tool. The detector analyses CT angiograms acquired under stroke-code protocols and alerts a neurovascular specialist "in parallel to standard of care image interpretation." Identification of suspected findings is not for diagnostic use beyond notification. The benefit-risk section concludes that "there are no major risks for the device because the device operates in parallel to the current usual standard of care," and the risks it does list are triage risks: deprioritization of other patients' images, inappropriate use for primary interpretation, and delayed management if prioritization fails. Nothing is blocked, queued or denied. The device can only add an earlier edge into the treatment team.
The pivotal evidence is standalone and retrospective: 300 CT angiograms from two US sites, roughly balanced positive and negative, on which the algorithm measured 87.8% sensitivity (95% CI 81.2-92.5), 89.6% specificity (83.7-93.9) — the share of scans with no occlusion that it correctly left alone — and an area under the curve of 0.91, a single number for how well its scores separate the scans that carry an occlusion from the ones that do not. The headline timing finding in the same record is that notification arrived a mean 51.4 minutes earlier than the standard radiology pathway — 7.3 minutes against 58.7, medians 5.6 against 51.5, with the software earlier in 42 of 44 cases and a per-study difference ranging from 12.7 minutes later to 206.4 minutes earlier. That comparison was computed on the 44 true-positive cases that happened to have documented standard-of-care notification times, and the FDA's own summary flags that the subset selection could carry bias. This atlas cites that figure to the FDA record and to nothing else, always with that caveat attached.
What independent evaluation later measured is materially lower, and lowest exactly where the anatomy is hardest. In a large integrated hub-and-spoke network handling more than 6,000 code strokes a year, 3,851 patients were screened and 220 had a neuroradiologist-confirmed internal-carotid or M1 occlusion: sensitivity 78.2% (95% CI 72-83), specificity 97%, positive predictive value 61%, negative predictive value 99% — and sensitivity fell to 60.6% once M2 occlusions were counted, with false negatives driven by non-main-vessel occlusions, segmentation errors and difficult anatomy. A prospective study of 1,822 consecutive stroke-code angiograms at a three-tiered network measured 93.8% sensitivity for internal-carotid-terminus and M1 occlusions but 74.6% once M2 was included, with per-site detection of 100% at the carotid terminus, 93% at M1 and 49% at M2. A positive predictive value of 61% means roughly two proximal alerts in five are false. The Atrium authors state the other side of it plainly: the software "does not allow the clinician to confidently rule out" an occlusion on a negative result. Both groups publish inside networks with substantial vendor-adjacent authorship, and the record should be read with that attached.
Pathway compression was measured in two small before-and-after studies, and this atlas states them with their designs rather than as outcomes. In a hub-and-spoke network (n=43: 28 before, 15 after) the median time from angiography at the primary centre to door-in at the comprehensive centre fell from 132.5 to 110 minutes, a 22.5-minute reduction (p=0.047), with reductions in overall and neurological intensive-care-unit (ICU) length of stay. In a transferred thrombectomy cohort at one academic centre (n=55) the median door-to-team-notification interval fell from 40.0 to 25.0 minutes (p=0.01) with less variance, while the 25-minute door-to-puncture improvement did not reach statistical significance (p=0.15). No randomised trial of this deployment exists, several of these authors have vendor-affiliated work, and no outcome or mortality benefit is asserted anywhere in this file.
The governance channel that actually scaled the deployment was not a clinical one. In the FY2021 inpatient prospective payment final rule (85 FR 58432, 18 September 2020) CMS approved a new-technology add-on payment for ContaCT of up to $1,040 per case — 65% of an applicant-estimated $1,600 cost — projecting roughly 12,700 cases and about $20.6 million in FY2021, billed through a dedicated inpatient procedure code, with the newness period anchored to 1 October 2018. CMS renewed it for FY2022. The characterization of this as the first add-on payment ever granted for AI software belongs to the applicant, to peer-reviewed commentary and to the press: the rule text approves the technology without declaring a first, and in the same rule CMS questioned whether AI-based workflow streamlining is a unique mechanism of action at all, asked what "new" means when an algorithm updates or a competitor's algorithm is better, and reserved the category question for future rulemaking. What CMS ultimately accepted was the applicant's asymmetry argument: a false positive costs only an earlier image review, while a false negative leaves the patient in the unchanged standard-of-care pathway. The add-on is temporary by construction, and this atlas does not assert that it is still being paid.
The payment left a per-use administrative trace, and that trace is the only system-wide measurement of this deployment anyone holds. A 2026 study from the Harvey L. Neiman Health Policy Institute, published in the American Journal of Neuroradiology, read 2,116 Medicare inpatient acute-ischemic-stroke episodes at 1,076 facilities from October 2020 to December 2023: add-on-billed use of the tool peaked at 21% of eligible episodes in 2022 and then declined, 14.8% across the whole window. Use concentrated at comprehensive stroke centres and facilities of 1,000 or more beds, with roughly twice the odds in the Stroke Belt, and showed no differences by patient demographics or stroke severity. The lead author's summary of it is that "access to these technologies depends more on where a patient is treated than on their clinical needs." Adoption rode the payment up and rode it down; what predicted whether a patient's hospital had the tool was the hospital's resources.
No lawsuit, enforcement action, recall or adverse regulator finding against the vendor or this product appears anywhere in this record as of 28 August 2026. The contested ground is intra-rulemaking — CMS's own mechanism-of-action and evidence questions, resolved in the applicant's favour — and peer-reviewed accuracy critique. Every adoption count in the record is vendor-tier and time-stamped: "over 800 U.S. hospitals" at the August 2021 renewal release, and 1,400-plus hospitals and health systems across the US and Europe on the vendor's current, undated page.
The sociotechnical reading
This is the atlas's inverse case, and the inversion is structural rather than rhetorical. Almost every other network in this catalogue is denial machinery: a model that blocks, queues, denies or actions something, with a human channel worn thin beside it. Here the model holds nothing back. The FDA order that authorized it forbids diagnostic use beyond notification and reaches a no-major-risks conclusion precisely because the device runs in parallel to an unchanged standard of care. So there is no enforcement node on this board, no gate, no queue a case can be held in, and the manual channel is not starved — the radiologist's read runs on every study, flagged or not, holds sole diagnostic authority, and is the reference standard against which every published error of this detector was counted. That last fact is worth sitting with: the accuracy numbers exist because the human check always fires. The map draws the second read of the angiogram at full strength for that reason, and it is the strongest check in the file.
What is left, once nothing can be denied, is a different failure surface, and the record names all of it. Alarm burden: at 61% positive predictive value roughly two proximal alerts in five are false, and each one spends a specialist team's attention. Queue displacement: the device's own regulatory class is computer-aided TRIAGE and notification, moving a flagged study up the reading list moves an unflagged one down, and deprioritization of other patients' images is one of the three minor risks the authorizing order lists — with nothing published measuring how often it happens or what waits behind it. And false reassurance, which is the inverted governance question this deployment poses: not whether a human can override the machine, but whether a team quietly starts treating machine silence as an all-clear, on distal occlusions the detector finds between 49% and 61% of the time. The independent authors say in terms that a negative result cannot be read that way. Between an order that confines the tool to notification and a warning that its silence carries no information sits the drift the map draws as a pressure.
The second reading concerns where governance actually sat. The channel that moved this deployment was a payment. An annual rule granted an add-on, reviewed it, renewed it for a second year and let it lapse on a fixed newness clock, and billed use rose and fell with it. That rule read billed use and wrote hospital purchasing; it never read the detector. The accuracy record that does exist was built by academic groups on the clinical side, and played no part in the payment lifecycle — the rule's own deliberation turned on mechanism of action and an argument about asymmetric harm. The map draws that as its one absent edge: use aggregates from the working record into the payment ledger, and nothing reads the ledger back against what the detector found and missed in those same episodes. That reconciliation is genuinely holdable — a hospital holds both records for its own cases, and an independent team proved the computation works by doing it across 2,116 Medicare episodes — which is what makes its absence a finding rather than an impossibility.
The boundary discipline is the usual one and it matters more here than usual, because this is a benefit case and benefit cases invite overclaim. Patients are not modelled. No clinical outcome, treatment decision or occlusion status for any person is computed anywhere on this diagram. Whether an earlier notification changed what happened to a patient is a question this Lab never answers, and the published record does not settle it either: no randomised trial exists and the compression findings come from small single-network studies with vendor-adjacent authorship. The one deployment-level equity measurement is double-edged and is carried exactly as measured — within the facilities that had the tool, billed use showed no difference by patient demographics or severity, while access to a facility that had it tracked comprehensive-centre status, bed count and region. That is a finding about which hospitals bought a product, recorded as an external observation of coverage. It is not derived from this network, and no node here computes it.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.