brianletort.ai
← Library

Greater Houston primary stroke centers

Verified evidenceCoordination compressedExceptions only

Automated LVO detection with parallel transfer notification

Large-vessel occlusion at a non-thrombectomy primary stroke center has to be recognized and the receiving team activated through sequential imaging review and conventional communication before transfer can start.

Clinical operations · Greater Houston, Texas, United States

Collections: Human still decides · Queue eliminated · Regulated autonomy

Sequential stroke imaging and transfer handoffs contrast with parallel clinical verification, team activation, and ambulance preparation.

Executive brief

The operating-model shift, in one view.

In a hub-and-spoke transfer workflow, the scarce resource is parallel activation, not a faster read of the same image. Moving LVO detection and team notification off the sequential critical path cut DIDO by more than an hour; the downstream treatment-rate claim did not survive adjustment and should not be sold as the result.

AI value · Primary-stroke-center door-in-door-out time (univariable median)

Verified

Median transfer time: 210 → 133 minutes

Observational pre/post design. Post-implementation n=35. Corresponding author reports consulting fees from Viz.AI outside the submitted work. Pre-AI median DIDO was long relative to some published networks, so effect size may not travel to faster centers.

Before

Receives a code-stroke patient and obtains noncontrast CT plus CTA under standard care. → Reviews CTA for large-vessel occlusion without an automated LVO alert. → Notifies the care team and arranges emergent transfer to a comprehensive stroke center after LVO is recognized.

After

Receives a code-stroke patient; NCCT and CTA are acquired under the same standard-of-care protocols. → Analyzes CTA for LVO in under five minutes on average and alerts the care-team mobile app. → Verifies the mobile PACS finding, messages the care team, and transfers eligible patients.

Human boundary

Humans retain transfer and EVT-eligibility decisions. The model does not independently dispatch the patient; clinicians must confirm imaging because positive predictive value is 0.65.

Why it matters

Transfer activation does not need to wait for sequential communication.

How the work changed

Before

How the work ran before the change.

  1. Step 1 of 3

    Emergency and telestroke team

    Receives a code-stroke patient and obtains noncontrast CT plus CTA under standard care.

    ControlInstitutional stroke order sets and imaging protocols

  2. Step 2 of 3

    Radiology / interpreting clinician

    Reviews CTA for large-vessel occlusion without an automated LVO alert.

    ControlConventional PACS interpretation

  3. Step 3 of 3

    Stroke and transfer coordinators

    Notifies the care team and arranges emergent transfer to a comprehensive stroke center after LVO is recognized.

    ControlPhone and conventional messaging; transfer eligibility judged by humans

What changed

Transfer activation does not need to wait for sequential communication.

Decision rightAI alerts in parallel; clinicians verify and govern transfer

After

How the same work runs now.

  1. Step 1 of 3

    Emergency and telestroke team

    Receives a code-stroke patient; NCCT and CTA are acquired under the same standard-of-care protocols.

    ControlAutomatic image transmission from all ED stroke workups

  2. Step 2 of 3

    Viz.AI LVO algorithm

    Analyzes CTA for LVO in under five minutes on average and alerts the care-team mobile app.

    ControlPublished algorithm performance NPV 0.99, PPV 0.65; humans must verify imaging

  3. Step 3 of 3

    Stroke clinicians

    Verifies the mobile PACS finding, messages the care team, and transfers eligible patients.

    ControlHuman confirmation of imaging and treatment eligibility; AI does not independently authorize transfer

Process model built from the published workflow evidence for Greater Houston primary stroke centers. Every step, actor, and control appears in full below.
Every step, actor, and control

Exception path

False-positive alerts are discarded after human image review. Patients without LVO, in-hospital code strokes, and inbound transfers were outside this workflow. Nighttime evaluations at all seven sites used telestroke.

Work removed

  • Waiting for sequential CTA interpretation before any care-team notification can start
  • Assembling the transfer team through conventional phone and pager chains as the primary activation path

Decision authority

Humans retain transfer and EVT-eligibility decisions. The model does not independently dispatch the patient; clinicians must confirm imaging because positive predictive value is 0.65.

Before

  1. 01

    Emergency and telestroke team

    Receives a code-stroke patient and obtains noncontrast CT plus CTA under standard care.

    Control: Institutional stroke order sets and imaging protocols

  2. 02

    Radiology / interpreting clinician

    Reviews CTA for large-vessel occlusion without an automated LVO alert.

    Control: Conventional PACS interpretation

  3. 03

    Stroke and transfer coordinators

    Notifies the care team and arranges emergent transfer to a comprehensive stroke center after LVO is recognized.

    Control: Phone and conventional messaging; transfer eligibility judged by humans

After

  1. 01

    Emergency and telestroke team

    Receives a code-stroke patient; NCCT and CTA are acquired under the same standard-of-care protocols.

    Control: Automatic image transmission from all ED stroke workups

  2. 02

    Viz.AI LVO algorithm

    Analyzes CTA for LVO in under five minutes on average and alerts the care-team mobile app.

    Control: Published algorithm performance NPV 0.99, PPV 0.65; humans must verify imaging

  3. 03

    Stroke clinicians

    Verifies the mobile PACS finding, messages the care team, and transfers eligible patients.

    Control: Human confirmation of imaging and treatment eligibility; AI does not independently authorize transfer

Work that left the path

  • Waiting for sequential CTA interpretation before any care-team notification can start
  • Assembling the transfer team through conventional phone and pager chains as the primary activation path

Human role before

Identify LVO on CTA and sequentially activate the transfer pathway through conventional interpretation and communication.

Human role after

Verify the automated LVO alert, confirm imaging, and govern transfer eligibility while the care team is notified in parallel.

AI roleDetect suspected LVO from CTA, push an alert, and provide a shared mobile PACS and messaging channel so activation is no longer gated on sequential phone-tree notification.

Outcomes

Primary-stroke-center door-in-door-out time (univariable median)

Verified

210 minutes (IQR 140–328.5)→133 minutes (IQR 100–167); p<0.001

1 January 2021 to 27 February 2022 · 115 LVO patients transferred from 7 Greater Houston PSCs (80 pre-AI, 35 post-AI)

Observational pre/post design. Post-implementation n=35. Corresponding author reports consulting fees from Viz.AI outside the submitted work. Pre-AI median DIDO was long relative to some published networks, so effect size may not travel to faster centers.

Adjusted DIDO time (multivariable linear regression)

Verified

Pre-AI DIDO in the same cohort→106-minute reduction (95% CI −165 to −48); p<0.001

1 January 2021 to 27 February 2022 · 115 LVO patients at 7 PSCs; model adjusted only for sex and on/off hours

Wide confidence interval. Adjustment covariates were limited. A secular-trend test found no significant DIDO decline over calendar time (coefficient −0.05, 95% CI −0.33 to 0.23), which argues against unexplained process improvement as the whole explanation but does not prove causality. EVT after transfer rose 41.2% to 62.9% univariably (p=0.033) and was not significant after adjustment (OR 2.13, 95% CI 0.88–5.13); that secondary outcome is excluded from this catalog case.

What leaders can reuse

Anti-pattern

Publishing a vendor percentage from a bundled quality-improvement program as if it were the isolated effect of the detection model, or treating a non-significant adjusted EVT-rate change as a confirmed clinical outcome.

Questions

  1. 01Which notification still sits on the sequential critical path after the model fires?
  2. 02If the secondary clinical endpoint is not significant after adjustment, what operating metric is actually the decision-relevant outcome?
  3. 03How will false positives be handled when the model's positive predictive value is 0.65?

Portability conditions

  • CTA already obtained on all code-stroke patients as standard of care
  • A named receiving CSC and an existing transfer protocol
  • Clinicians required to verify imaging because model PPV is far from 1.0
  • Baseline DIDO long enough that a 100-minute shift is operationally available

Reputation risk

medium

Evidence and authority

What the public record supports.

Current · updated

4 peer reviewed; publication outcomes are verified.

Bundle 1.0.0 · reviewed 2026-08-23 · stable ID 8e195367ffc0725c

Related transformations

More in Clinical operations →

C.H. Robinson · Coordination compressed

AI agents for orders and appointments inside Always On Logistics Planner

That people must key routine freight orders and dock appointments, taking hours, before a carrier can be matched.

Boston Public Schools · Coordination compressed

BiRD system-level school-bus route optimization

Routes can be optimized as one district system instead of built school by school and then combined.

City of Pittsburgh and Carnegie Mellon University · Coordination compressed

SURTRAC decentralized adaptive traffic-signal control

That urban intersections should follow fixed timing plans optimized offline for predictable rush periods.

UPS · Coordination compressed

ORION algorithmic route sequencing for delivery drivers

That each driver or dispatcher must determine the daily stop sequence from experience and static route plans.

Sources

Read the evidence, freshness, caveat, and version policy.