Skip to content
ParamergeParamerge
Stephen Lieberman
Through Paramerge
Contact

PAN Lab example

Viz.ai LVO Stroke Triage

The alert that gates nothing: notification beside the standard read

A stroke code starts, a CT angiogram is acquired, and two things happen at once. A detector scores the study for a large-vessel occlusion and, on a hit, pushes an alert with a compressed preview to the neurointerventional team's phones. Meanwhile the pathway that existed before the detector did carries on untouched: the radiologist reads the study, reports onward, the team is paged. Modeled on the imaging-triage system a United States regulator authorized in 2018, creating the computer-aided triage and notification device class. This is the atlas's inverse case. Almost every other board here is denial machinery — a model that blocks, queues, denies or actions something, and a human channel worn thin beside it. This model holds nothing back. The order that authorized it fixes the output as triage and notification, forbids its use for diagnosis beyond that, and reaches a no-major-risks conclusion precisely because the device runs in parallel to the usual standard of care. So the manual channel is not starved on this board; it is stronger than the demand on it, and the second read of the angiogram runs at full strength on every case. The failure surface is what remains once nothing can be denied. Roughly two proximal alerts in five are false. Distal occlusions are detected between 49% and 61% of the time, and the evaluators warn that a quiet detector does not let a clinician rule an occlusion out. Moving a flagged study up the reading list moves someone else's study down, which the authorizing order lists as a risk in its own words. And the deployment's governance channel is a payment: an add-on of up to $1,040 a case, granted for one fiscal year, renewed for a second, temporary by construction. Billed use rose with it to 21% of eligible episodes in 2022 and fell afterwards, concentrated at comprehensive stroke centres and thousand-bed hospitals, with no difference by patient demographics or severity — access tracked where a patient was treated rather than what they needed. Nothing reads the payment ledger back against what the detector found and missed. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets, and the obstacle is money. Those two target levels also ask you to close every open pathway. That is possible on this board and it is not out of reach in principle, but the cheapest arrangement that does it costs more than you are given, on both. Nothing in this deployment's record ever paid for governing this tool. Money arrived to buy it, on a two-year clock, from outside the hospital, and none of it was attached to watching it. So the budget on this board is what a hospital genuinely holds, and it is short of what clearing this board costs. There is a second thing worth knowing before you start, because it is the shape of the deployment rather than a trick. The pathway that is hardest to close is the alert to the specialist team, which is the single edge this system adds to the network and the whole of what it is for. Escalated checking closes it. Widening the coverage dial opens it again, because more of the stroke stream reaching the detector is more of that same edge. Benefit and closure pull against each other here in a way they do not on a board where the model denies something. Explore and Service Targets Only can be won, and cheaply.

Stylized model of a documented deploymentClinical decision support & deterioration alerting

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the LVO-triage-class parallel stroke notification network: 11 components and 23 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 3 assumed · 6 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

  • assumed

    D48-derived new org (Phase 6, clinical-decision-support). TOPOLOGY. Eleven nodes, all documented, none decorative. The shape is the deployment's defining claim drawn as structure: an ADDITIVE, non-gating alert beside an unchanged standard pathway. One model (the authorized detector), one inputSource (the stroke-code angiography feed the model scores as it is acquired), three recordStores doing genuinely different work (the imaging-and-alert record, the per-use add-on-payment billing record under ICD-10-PCS 4A03X5D, and the hospital-purchased stroke-network routing record), three operatorClasses on documented separate roles (the radiology reading service holding diagnostic authority, the neurointerventional team the alert is built to reach, and the transfer coordinators both independent network evaluations were run around), one reviewer wired in by an inbound read of the billing record (the annual payment-rulemaking channel), one worklist (the radiology reading queue the device's own class name refers to), and one externalBoundary with LIVE egress (the vendor's hosted platform). FOUR absences are equally derived and each is load-bearing: NO enforcement node and no gating surface anywhere, because nothing on this board is blocked, queued, denied or actioned by the model and the authorizing order's benefit-risk conclusion rests on exactly that; NO storeToModel loop from the working imaging-and-alert record back into the detector, because the detector scores each study off the acquisition feed as it leaves the scanner, and nothing in the record documents it reading back the alert events, the timestamps or either read that the working record then accumulates — so the contaminated-record-repeated-as-fresh-output loop does not close here, which is what separates this org from the domain's record-reading bedside-alert siblings; NO guardrail and no automated output screen between the detector and the people who act, because the regulator's controls here are pre-market performance testing and labeling limits rather than a screen; and NO operatorToExternal egress, because no unsanctioned-tool pathway appears anywhere in the record.

  • baseline

    DEMAND 2 / CAPACITY 3 — and this org is where that mapping inverts. Demand 2: the load is real, time-critical and bounded. The detector runs on the whole stroke-code stream (3,851 patients screened in one integrated network's evaluation window, 1,822 consecutive stroke-code studies in a prospective one), but the alerting load on the specialist team is a minority of it — 220 of 3,851 screened patients had a confirmed proximal occlusion, and at 61% positive predictive value roughly two proximal alerts in five are false. That is an attention cost, not a backlog. Capacity 3: the manual counterfactual here is intact and working, which is the deployment's defining fact rather than a generous reading. The standard interpretation continues in parallel with unchanged diagnostic authority; the authorizing order's benefit-risk conclusion is that there are no major risks because the device operates in parallel to the current usual standard of care; and the neuroradiologist read is the reference standard every published error count of this detector was scored against, which is the proof that the unassisted channel still works. Held at 3 rather than higher because the same order lists deprioritization of other patients' images among its three minor risks, and nothing published measures how often a flagged study jumps the reading queue.

  • baseline

    BASELINES, derived from documented coverage and measured rates rather than from a template. The angiogram feed is 3 because the detector scores the whole stroke-code protocol population by design. The alert to the specialist team is 3 because it is the one edge the device adds and the entire product is that edge, arriving at a pivotal median of 5.6 minutes after acquisition. The radiologist's read of the study and the radiologist's write back into it are both 3 because they happen on every case, flagged or not, which is exactly why they could serve as the reference standard of every published evaluation. The alert into the reading room is 1 because it reorders a queue rather than causing the read. Billing is 1 on both its operator write and its record aggregation because the measurement says so: billed use reached 21% of eligible Medicare episodes at its 2022 peak and 14.8% across the full window, so the trace covers a minority of the eligible stream. The peer hops are 2 for the standard notification pathway and the transfer conversation and 1 for the specialist's back-call, following the record's own emphasis. The routing read by the detector is 1 because configuration shapes an alert's destination and not its content.

  • baseline

    THE STRONGEST CHECK ON THIS BOARD IS 3, AND THAT IS EVIDENCE, NOT GENEROSITY. The second read of the angiogram is drawn at full strength because the authorizing order requires the standard interpretation to continue in parallel with unchanged diagnostic authority, forbids the detector's output from being used for diagnosis beyond notification, and reaches its no-major-risks conclusion on that structure; and because both independent multi-site evaluations were built by scoring this detector's alerts against neuroradiologist reads, which is a direct demonstration that the check fires on every case. The specialist team's own confirmation of the imaging before treating is a second, real check, drawn at 2 because it runs on the flagged subset and because two documented pressures bound it: roughly two proximal alerts in five are false at 61% positive predictive value, and the independent authors warn that a quiet detector does not let a clinician rule an occlusion out, which measured distal detection of 49% to 61% supports. No automated output screen exists between the detector and the people who act, so no modelToOperatorCheck is drawn.

  • baseline

    THE ONE LATENT EDGE IS A DOCUMENTED ABSENCE, AND IT IS THE CASE'S GOVERNANCE FINDING. Billed use aggregates from the working record into the payment ledger, and the reconciliation back — reading what was billed against what the detector found and missed in those same episodes — is drawn at zero. The payment ledger holds the fact of use and nothing that measures the detector; the annual rulemaking that granted, renewed and lapsed the add-on payment turned on whether workflow streamlining by software is a unique mechanism of action and on an accepted argument that a false alert costs an earlier image review while a missed occlusion leaves the patient in the unchanged pathway; and the two evaluations that did score this detector were built by academic groups on the clinical side, not by the payment channel. That is why the reviewer node emits no check of its own: the governance channel that scaled and unscaled this deployment reads adoption and writes purchasing, and the accuracy record sits outside its loop.

  • baseline

    THE EGRESS IS DRAWN PRESENT, WHICH IS UNUSUAL AND DELIBERATE. Most orgs in this catalogue draw their boundary edges at zero because egress is the failure they document. Here two of the three egress kinds are the operating mode: whole angiographic studies cross from the hospital estate to a vendor's hosted platform for analysis, and the detector, the messaging and the mobile image viewer are the vendor's three marketed components running on that platform, which is how a compressed preview reaches a specialist's phone. Both are drawn at 2 — contracted, device-authorized, with no breach, enforcement action or onward disclosure anywhere in the record — rather than at 3, and the third egress kind is not drawn at all because no unsanctioned-tool pathway appears in the record. The privacy consequence carried on the diagram is the minimization one: the studies that cross belong to the whole stroke-code population, and in the largest independent evaluation 3,631 of 3,851 screened patients had no confirmed proximal occlusion.

  • assumed

    WHERE THE RECORD IS SILENT, THE CONSERVATIVE VALUE — and the silences here are specific. No source publishes an override, correction or acceptance rate for any class on this board, because there is nothing to override: the alert is advisory by construction and the standard pathway is untouched. No source measures how often a flagged study jumps the reading queue, or what waits behind it, although deprioritization is one of the three risks the authorizing order lists. No source measures whether a quiet detector changes what a team does next, although the independent authors name that hazard explicitly. Those magnitudes are drawn within the qualitative rungs the documented figures support, matching the PAN org's own estimated-on-every-parameter discipline, and each is named here rather than smoothed over.

  • baseline

    EVIDENCE TIERING IS CARRIED ONTO THE DIAGRAM, because this deployment's record mixes tiers unusually sharply. Regulator-primary and asserted as fact: the device class creation, the notification-only parallel-workflow indication, the pivotal standalone performance, the three listed minor risks, and the payment rule's terms and stated reasoning. Independent peer-reviewed and asserted as fact with its design attached: the multi-site accuracy figures, including the distal weakness, and the health-policy measurement of adoption. Attributed, never asserted: the characterization of the add-on payment as the first for software of this kind, which belongs to the applicant, to commentary and to the press — the rule text approves the technology without declaring a first and reserved the category question for later rulemaking. Vendor-tier and dated: every adoption count. NOT asserted anywhere on this diagram or in its case file: any outcome or mortality benefit, any current payment status, and any framing in which this system diagnoses a stroke, which the authorizing order forbids.

  • assumed

    SERVED PEOPLE ARE NOT IN THE DYNAMICS. Stroke patients, the patients behind the studies that cross to a vendor platform, the patients at spoke hospitals whose transfer depends on a hub being notified, and the patients whose access tracks their hospital's resources are all boundary populations recorded in the case file. No node, edge, baseline or lever here computes a clinical outcome for any patient, and none may. The one deployment-level equity measurement in the record is double-edged and is carried as measured: within the facilities that had the tool, billed use showed no difference by patient demographics or stroke severity, while access to a facility that had it tracked comprehensive-centre status, bed count and region. That is a property of the routing record's coverage, not a property of anyone's treatment, and it is recorded as an external observation rather than derived from this network.

What this example does not show

  • Patients are not modeled. No clinical outcome, treatment decision or occlusion status for any person is computed anywhere on this diagram; the network propagates record and workflow error through operators and stores. 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 randomized trial of this deployment exists, and the pathway-compression findings that do exist come from small single-network studies, several with vendor-affiliated authorship.
  • The notification-advantage figure this case carries comes from the regulator's own pivotal record and from nowhere else. In the pivotal study the system reached notification a mean 51.4 minutes earlier than the standard pathway (7.3 minutes against 58.7), and the decision summary that reports it flags the reason to be careful: the comparison was computed on the 44 true-positive cases that happened to have documented standard-of-care notification times, a subset the regulator itself says could carry bias. It is that study's measurement, not a field-wide constant, and this atlas cites no other study for it.
  • The accuracy figures on this board are two different things and the difference matters. The pivotal 87.8% sensitivity and 89.6% specificity are a standalone retrospective measurement on 300 studies from two sites. The 78.2% proximal sensitivity, 61% positive predictive value and 49% to 61% distal detection are independent multi-site measurements against neuroradiologist reads in real networks. Both are cited here with their designs attached, and the second set is the one this network's dynamics are ordered by.
  • The characterization of the add-on payment as the first of its kind for software of this sort is the applicant's, commentators' and press framing. The rule text approves the technology without declaring a first, and in the same rule the payer questioned whether workflow streamlining by software is a unique mechanism of action and reserved the category question for later rulemaking. The payment is also temporary by construction — granted for one fiscal year, renewed for a second, tied to a newness period that began in October 2018 — and this atlas does not assert that it is still being paid. Every adoption count in this case is vendor-tier and dated.
  • Litigation and regulatory posture, carried verbatim from the evidence record: "None documented as of 2026-08-28: no lawsuits, enforcement actions, recalls, or adverse regulator findings located against Viz.ai or the LVO product; the contested terrain is intra-rulemaking (CMS's mechanism-of-action and evidence concerns, resolved in the applicant's favor) and peer-reviewed accuracy critique."
  • The equity measurement here is double-edged and is carried as measured, not resolved. Within the facilities that used the tool, billed use showed no difference by patient demographics or stroke severity. Access to a facility that used it tracked comprehensive-centre status, bed count and region instead. 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.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

  • 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.

    empirical
    • Government U.S. Food and Drug Administration, Center for Devices and Radiological Health (2018). Evaluation of Automatic Class III Designation (De Novo) for ContaCT: Decision Summary (DEN170073) https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN170073.pdf
    • Government U.S. Food and Drug Administration (2018, February 13). FDA permits marketing of clinical decision support software for alerting providers of a potential stroke in patients (press announcement) https://www.fda.gov/news-events/press-announcements/fda-permits-marketing-clinical-decision-support-software-alerting-providers-potential-stroke

  • Independent multi-site evaluation of this detector measured accuracy materially below the pivotal figures, and lowest on the distal occlusions. In a large integrated hub-and-spoke stroke network handling more than 6,000 code strokes a year, 3,851 patients were screened and 220 (5.7%) had a neuroradiologist-confirmed internal-carotid or MCA-M1 occlusion: sensitivity 78.2% (95% CI 72-83), specificity 97% (96-98), positive predictive value 61% (55-67), negative predictive value 99% (98-99), with sensitivity falling to 60.6% once MCA-M2 occlusions were included and false negatives driven by non-main-vessel occlusions, segmentation errors, and difficult anatomy; those authors state that the software 'does not allow the clinician to confidently rule out' an occlusion on a negative result. A prospective study of 1,822 consecutive stroke-code CT angiograms (May 2019 to October 2020) at a three-tiered network found 190 occlusions and measured 93.8% sensitivity with 99.7% negative predictive value for internal-carotid-terminus and M1 occlusions, dropping to 74.6% sensitivity and 97.6% negative predictive value 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. Both groups publish inside networks with substantial vendor-adjacent authorship, and the prospective study's site is a frequent vendor co-publisher.

    empirical
    • Academic Karamchandani, R. R., et al. (2023). Automated detection of intracranial large vessel occlusions using Viz.ai software: Experience in a large, integrated stroke network. Brain and Behavior, 13(1), e2808 https://pmc.ncbi.nlm.nih.gov/articles/PMC9847593/
    • Academic Matsoukas, S., et al. (2022). AI software detection of large vessel occlusion stroke on CT angiography: a real-world prospective diagnostic test accuracy study. Journal of NeuroInterventional Surgery, 15(1), 52-56 https://doi.org/10.1136/neurintsurg-2021-018391
    • Government U.S. Food and Drug Administration, Center for Devices and Radiological Health (2018). Evaluation of Automatic Class III Designation (De Novo) for ContaCT: Decision Summary (DEN170073) https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN170073.pdf

  • In the FY2021 hospital 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 ICD-10-PCS 4A03X5D, with the newness period anchored to 1 October 2018; CMS renewed it for FY2022. The characterization of this as the first add-on payment granted for AI software is the applicant's, peer-reviewed commentary's, and the press's framing: 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, raised what 'new' means when an algorithm updates or a competitor's algorithm performs better, and reserved the category question for future rulemaking. What CMS accepted was the applicant's asymmetry argument — that a false positive causes 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 its post-FY2022 status was not verified in rule text; this atlas does not assert that it is still being paid. Adoption counts for this deployment are vendor-tier and point-in-time: '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.

    empirical
    • Government Centers for Medicare & Medicaid Services, FY 2021 Hospital Inpatient Prospective Payment Systems final rule (85 FR 58432, 18 September 2020; ContaCT NTAP at 85 FR 58619), read via the Healthcare Financial Management Association final-rule summary https://www.federalregister.gov/documents/2020/09/18/2020-19637/medicare-program-hospital-inpatient-prospective-payment-systems-for-acute-care-hospitals-and-the
    • Academic Hassan, A. E. (2021). New Technology Add-On Payment (NTAP) for Viz LVO: a win for stroke care. Journal of NeuroInterventional Surgery, 13(5), 406-408 https://pmc.ncbi.nlm.nih.gov/articles/PMC8053346/
    • Vendor Viz.ai (2020, September 3). Viz.ai Granted Medicare New Technology Add-on Payment; and Viz.ai (2021, August 4). Viz.ai Receives New Technology Add-on Payment (NTAP) Renewal for Stroke AI Software from CMS https://www.viz.ai/news/viz-granted-medicare-ntap

  • The add-on payment left a per-use administrative trace, and that trace is the only system-wide measurement of this deployment anyone holds. A 2026 Harvey L. Neiman Health Policy Institute study in the American Journal of Neuroradiology read 2,116 Medicare inpatient acute-ischemic-stroke episodes at 1,076 facilities between October 2020 and 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 — facility resources, not patient factors, predicted use, which the lead author summarised as access depending 'more on where a patient is treated than on their clinical needs'. Two smaller before-and-after studies measured pathway compression with their designs attached and are not outcome evidence: in a hub-and-spoke network (n=43, 28 before and 15 after) the median interval 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 neuro-ICU length of stay; and 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) while the 25-minute door-to-puncture improvement did not reach significance (p=0.15). No randomised trial of this deployment exists, several of these authors hold vendor-affiliated work, and no outcome or mortality benefit is asserted.

    empirical
    • Academic Pelzl, C. E., et al. (2026). Patient and Facility Factors Associated with Uptake of NTAP-Billed AI to Identify Suspected LVO in Ischemic Stroke. American Journal of Neuroradiology (published online 24 June 2026) https://www.ajnr.org/content/early/2026/06/24/ajnr.A9494.abstract
    • Trade press Harvey L. Neiman Health Policy Institute (2026). Hospital Resources, Not Patient Need, Drive Stroke AI Tool Adoption Under Medicare (press release); and AuntMinnie.com (2026). Hospital resources tied to AI adoption for stroke detection https://www.neimanhpi.org/press-releases/hospital-resources-not-patient-factors-linked-to-adoption-of-ai-for-stroke-detection-under-medicare-new-technology-add-on-payment/
    • Academic Hassan, A. E., et al. (2020). Early experience utilizing artificial intelligence shows significant reduction in transfer times and length of stay in a hub and spoke model. Interventional Neuroradiology, 27(1); and Morey, J. R., et al. (2021). Real-World Experience with Artificial Intelligence-Based Triage in Transferred Large Vessel Occlusion Stroke Patients. Cerebrovascular Diseases, 50(4), 450-455 https://scholarworks.utrgv.edu/som_pub/1894/

Where this connects

Institutional pressures in this domain

  • Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
  • Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.
  • Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
  • Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
  • Deadline pressure — Statutory or managerial timeliness rules reward fast approval of machine output over slow disagreement.

All of them in context on the Clinical decision support & deterioration alerting domain page.

Levers available here and the patterns behind them

Documented case histories