The Time Is Brain Review

GOLDEN BRIDGE II — Clinical Decision Support for Acute Ischaemic Stroke

· DOI: 10.1136/bmj-2025-085810

stroke deep-dive clinical-decision-support quality-improvement

Reconstruction note. Reconstructed on October 14, 2026 for the missed October 13 publication slot. This is a replacement article, not a recovered original. The paper was selected in verified PMC look-back mode; it was not recovered as that day's original briefing article.

Clinical Question (PICO)

In adults hospitalized within seven days of MRI-confirmed acute ischaemic stroke at eligible Chinese hospitals, does a hospital-level clinical decision support system (CDSS)—combining imaging analysis, stroke-cause classification, reminders, and treatment recommendations—compared with usual care reduce new vascular events by three months and improve care quality?1 The primary comparison was cluster-randomized by hospital; later hospital-region, hospital-grade, and stroke-unit subgroups were post hoc.1

Bottom Line

Among 21,603 patients in 77 hospitals, three-month new vascular events occurred in 2.9% (320/11,054) with CDSS and 3.9% (416/10,549) with usual care (adjusted HR 0.74, 95% CI 0.58–0.93; P=.01); a weighted hospital-level analysis also favored CDSS (effect −0.01, 95% CI −0.02 to −0.004; P=.003).1 Care-process scores improved, but disability and mortality did not differ significantly.1 This cluster trial tests one integrated, supported implementation in China—not an AI component in isolation or every decision-support product.1

Design

  • Trial type: Multicentre, open-label, pragmatic, cluster-randomized clinical trial; hospitals, not individual patients, were allocated.1
  • Clusters and participants: 77 hospitals (38 CDSS, 39 usual care), 1:1 allocation; 21,603 analyzed (11,054 and 10,549).1 Three of 80 approached hospitals declined; 86 patients who withdrew within one day were excluded from the final analysis.1
  • Setting: Secondary and tertiary hospitals in 23 Chinese provinces, with emergency and neurology services and 1.5-T or 3.0-T MRI capability.1
  • Enrollment and follow-up: January 13, 2021–June 25, 2023; outcomes assessed at 3, 6, and 12 months.1 Follow-up was 99.8% at 3 months and 93.6% at 12 months.1
  • Analysis: Intention-to-treat as reported; mixed-effects Cox models included a hospital random effect, with a weighted hospital-level generalized linear analysis for the primary outcome.1 The adjusted outcome models accounted for the baseline prestroke mRS imbalance.1
  • Primary outcome: New vascular event—ischemic stroke, haemorrhagic stroke, myocardial infarction, or vascular death—within 3 months after symptom onset.1

Population

Inclusion criteria. Patients aged 18 years or older with acute ischaemic stroke confirmed by brain MRI within seven days of symptom onset, admitted to an eligible secondary or tertiary hospital with emergency/neurology services and MRI capability.1

Exclusion criteria. Patients with other cerebrovascular diagnoses (including TIA, haemorrhagic stroke, or cerebral venous sinus thrombosis) or non-cerebrovascular disease were excluded.1 Primary and specialised hospitals were not eligible clusters.1

Baseline characteristics. Median age was 67 years in the intervention group and 66 in control; 64.5% and 64.4% were men.1 Admission NIHSS median was 3 (IQR 1–5) in both groups.1 Prestroke mRS ≥3 was more common with CDSS (6.5% vs 3.8%; absolute difference 12.2%), the baseline characteristic with an absolute imbalance greater than 10% and the adjustment variable reported for the outcome table.1 About 89.6% of hospitals had stroke units.1

Interventions

  • CDSS hospitals: An integrated system extracted clinical information, reminded clinicians about pending examinations, used deep learning for MRI lesion segmentation/pattern features, classified stroke cause, and offered guideline- and evidence-based diagnostic, acute-management, secondary-prevention, and performance-measure recommendations.1 The workflow included antithrombotic and risk-factor recommendations, dysphagia screening, and DVT prophylaxis.1 Clinicians retained discretion to accept or reject recommendations.1 Intervention clinicians received two weeks of training and technical support.1
  • Usual-care hospitals: Routine diagnosis and treatment based on clinicians' experience and guideline knowledge; no assigned CDSS workflow.1

Outcomes

Estimates below are adjusted comparisons unless stated; hazard ratios below 1 favor CDSS for adverse outcomes, while odds ratios above 1 favor CDSS for care-process outcomes.1 The composite-care denominator counts eligible measure opportunities, not unique patients.1

Outcome and time CDSS Usual care Adjusted effect (95% CI); P
Primary: new vascular event, 3 mo 320/11,054 (2.9%) 416/10,549 (3.9%) HR 0.74 (0.58–0.93); .01
Primary, hospital-level weighted analysis Incidence 0.03 (SD 0.02) 0.04 (SD 0.02) Effect −0.01 (−0.02 to −0.004); .003
Ischaemic stroke, 3 mo 2.4% 3.4% HR 0.70 (0.54–0.90); .006
Composite evidence-based care measure 77,049/84,276 (91.4%) 70,794/78,834 (89.8%) OR 1.21 (1.17–1.26); <.001
All-or-none care measure 6,504/11,054 (58.8%) 5,619/10,549 (53.3%) OR 1.18 (0.87–1.61); .29
New vascular event, 6 mo 379/11,054 (3.4%) 503/10,549 (4.8%) HR 0.71 (0.56–0.91); .006
New vascular event, 12 mo 440/11,054 (4.0%) 576/10,549 (5.5%) HR 0.73 (0.56–0.95); .02
Disability (mRS 3–6), 3 mo 1,297/11,040 (11.8%) 996/10,519 (9.5%) OR 1.14 (0.90–1.44); .29
Disability, 6 mo 1,015/10,807 (9.4%) 740/10,147 (7.3%) OR 1.18 (0.89–1.56); .26
Disability, 12 mo 841/10,557 (8.0%) 663/9,652 (6.9%) OR 1.11 (0.81–1.52); .51
All-cause mortality, 3 mo 141/11,054 (1.3%) 138/10,549 (1.3%) HR 0.91 (0.60–1.40); .68
All-cause mortality, 6 mo 218/11,054 (2.0%) 243/10,549 (2.3%) HR 0.79 (0.53–1.17); .24
All-cause mortality, 12 mo 335/11,054 (3.0%) 372/10,549 (3.5%) HR 0.77 (0.52–1.13); .18
Moderate/severe bleeding, 3 / 6 / 12 mo 32/11,054 (0.3%) / 39/11,054 (0.4%) / 49/11,054 (0.4%) 32/10,549 (0.3%) / 46/10,549 (0.4%) / 51/10,549 (0.5%) HR 0.89 (0.44–1.79), .74 / 0.79 (0.43–1.46), .44 / 0.89 (0.49–1.65), .72
All bleeding, 3 / 6 / 12 mo 88/11,054 (0.8%) / 96/11,054 (0.9%) / 106/11,054 (1.0%) 129/10,549 (1.2%) / 143/10,549 (1.4%) / 151/10,549 (1.4%) HR 0.68 (0.41–1.12), .13 / 0.67 (0.42–1.07), .09 / 0.66 (0.41–1.07), .09

The all-or-none care measure did not reach statistical significance.1 Several process measures improved, including dysphagia screening (98.5% vs 91.2%), dual antiplatelet treatment in eligible non-disabling cases (76.2% vs 69.6%), and in-hospital anticoagulation for atrial fibrillation (77.0% vs 69.3%).1 DVT prophylaxis was 831/2,240 (37.1%) vs 486/1,622 (30.0%; rate difference 7.1 percentage points, 95% CI 4.1–10.1); discharge anticoagulation for atrial fibrillation was 471/609 (77.3%) vs 357/529 (67.5%; difference 9.9 points, 95% CI 4.7–15.0).1 No statistically significant differences were found in disability or all-cause mortality at measured time points; nonsignificant bleeding comparisons do not establish absence of harm.1

Figures

Workflow diagram for the stroke clinical decision support system, including information extraction, image analysis, stroke-cause classification, and care recommendations.
Figure 1. The CDSS workflow combines clinical-data extraction, MRI analysis, etiologic classification, and treatment/performance prompts.[1](#fn-source-1)

Source: GOLDEN BRIDGE II authors, The BMJ, CC BY-NC 4.0; PMC13003563 — zhax085810.f1.jpg. Click image to expand.

Trial flowchart showing hospital allocation, patient enrollment, exclusions, and analyzed groups.
Figure 2. The flow diagram traces cluster participation and patient inclusion through the final analysis.[1](#fn-source-1)

Source: GOLDEN BRIDGE II authors, The BMJ, CC BY-NC 4.0; PMC13003563 — zhax085810.f2.jpg. Click image to expand.

Twelve-month cumulative-probability panels for vascular events, ischemic stroke, all-cause mortality, and moderate or severe bleeding by trial group.
Figure 3. Kaplan–Meier panels display 12-month vascular events, ischaemic stroke, all-cause mortality, and moderate/severe bleeding by group; displayed log-rank tests are unadjusted.[1](#fn-source-1)

Source: GOLDEN BRIDGE II authors, The BMJ, CC BY-NC 4.0; PMC13003563 — zhax085810.f3.jpg. Click image to expand.

Criticisms

  • Cluster design and analysis: The hospital is the randomization unit, so patient counts alone overstate independent information.1 The paper appropriately reports mixed-effects models with hospital random effects and a weighted cluster-level analysis; readers should not substitute unclustered comparisons for these estimates.1
  • Implementation package: This was an open-label, multicomponent intervention with two weeks of training and technical support.1 The causal estimate belongs to the whole supported system; it cannot isolate AI image analysis from workflow reminders, physician review, performance measurement, or implementation effort.1
  • Applicability: The trial enrolled mostly mild strokes (median NIHSS 3), excluded primary/specialised hospitals, and did not include thrombectomy decision support.1 Costs and hospital length of stay were not evaluated; generalization beyond these Chinese secondary/tertiary settings remains an implementation question.1
  • Outcome interpretation: The primary event was uncommon and the observed event-rate difference was about one percentage point as described by the authors.1 Disability and mortality were not significantly different.1 The trial reported no multiplicity adjustment for the primary outcome, and secondary analyses were exploratory.1
  • External review: No independent editorial or letter was retrieved for this local reconstruction; this appraisal is limited to the trial report.

Funding

Funding was provided by the National Key Research and Development Program of China (2022YFC2504902), National Natural Science Foundation of China (92046016), Beijing Municipal Administration of Hospitals’ Mission Plan (SML20150502), Ministry of Industry and Information Technology (2020-0103-3-1), CAMS Innovation Fund for Medical Sciences (2019-I2M-5-029), and Beijing Ande Yizhi Technology Co.; the authors state that funders had no role in design, conduct, analysis, interpretation, manuscript preparation, or submission.1 Disclosures include consulting relationships reported by GCF and LHS and consulting/research support reported by YX.1

The paper

  • Authors. Xinmiao Zhang, Lingling Ding, Jing Jing, Chunjuan Wang, et al.; GOLDEN BRIDGE II Investigators (the first four authors are reported as joint first authors).
  • Title. Effect of a clinical decision support system on stroke care quality and outcomes in patients with acute ischaemic stroke (GOLDEN BRIDGE II): cluster randomised clinical trial.
  • Journal. The BMJ. 2026;392:e085810.
  • DOI. 10.1136/bmj-2025-085810
  • PMCID. PMC13003563
  • Registration. ClinicalTrials.gov NCT04524624.
  • License. Creative Commons Attribution Non-Commercial 4.0 International (CC BY-NC 4.0); figures are attributed to the original authors and publication.1
  • Source [1]. Europe PMC full-text XML.
Deep Dive — click to expand

What this is

GOLDEN BRIDGE II is a large cluster-randomized test of an end-to-end stroke-care support package, not a head-to-head contest between an AI model and a clinician.1 The intervention linked patient data and MRI-derived lesion features to cause classification, reminders, treatment prompts, and care-quality monitoring; clinicians could decline recommendations.1 The primary outcome favored intervention at three months, while disability and mortality estimates remained inconclusive.1

1. Shadow Audit

The headline HR of 0.74 is clinically notable but applies to a composite of ischaemic stroke, haemorrhagic stroke, myocardial infarction, or vascular death, not to disability.1 Ischaemic stroke was lower (2.4% vs 3.4%; HR 0.70), but the paper found no significant difference in disability or mortality in its adjusted outcome table.1 Process adherence improved on several measures, but the all-or-none measure did not (OR 1.18, 95% CI 0.87–1.61; P=.29).1 These are distinct outcomes and should not be collapsed into a claim that patients were less disabled.1

2. Inversion Engine

If the event reduction is smaller or less transportable elsewhere, the result could reflect this package's training, integration, and local implementation support rather than a portable software effect.1 Confidence would change with independent cluster-randomized replication using prespecified cluster-aware analysis, transparent uptake/fidelity reporting, and patient-centered outcomes.1 The primary mixed-effects HR and weighted hospital-level estimate both favored the system, but neither identifies which component produced the benefit or establishes cost-effectiveness.1

3. Second-Order Catalyst

Editorial inference, not a tested implementation prescription: hospitals considering a comparable system would first need reliable MRI/data integration, a governance process for clinician overrides, and capacity for sustained training and technical support.1 The trial's 38 intervention hospitals received two weeks of training; reproducing those conditions may be part of the intervention effect.1 These findings do not establish that a generic chatbot, stand-alone imaging model, or unintegrated alert will produce the same result.1

4. Asymmetric Leverage

The event-rate difference was modest in absolute scale but observed across more than 21,000 patients; the authors describe an approximately one-percentage-point difference and a 25.6% relative decrease.1 The process gains were largest where baseline adherence was lower—such as dysphagia screening and atrial-fibrillation anticoagulation—while already-high performance measures changed little.1 The asymmetry is an implementation hypothesis: target bottlenecks may matter more than adding alerts to already reliable workflows.1

5. Paradigm Destroyer

The trial counters the idea that decision support in stroke care is only an imaging classifier: the tested product combined imaging with cause classification, reminders, recommendations, and performance tracking.1 But it does not justify replacing clinician judgment; recommendations remained discretionary.1 It also does not generalize to severe stroke or thrombectomy triage, which the study says were not adequately addressed by this enrolled population and system.1

MVP — Minimum Viable Proof

The minimum clinical signal is 320/11,054 versus 416/10,549 new vascular events at three months, with a hospital-randomized adjusted HR 0.74 (95% CI 0.58–0.93; P=.01) and a supportive weighted cluster-level estimate of −0.01 (−0.02 to −0.004; P=.003).1 The lack of a corresponding significant disability or mortality difference limits the claim to fewer observed vascular events and better process performance, not proven functional recovery.1

Best Combination

Interpret the primary result alongside the process table, follow-up completeness, cluster-aware analyses, and 12-month event estimate.1 The mixed-effects model accounts for hospital clustering, while the cluster-level analysis offers a second scale of inference; both point in the same direction.1 The primary paper's discussion proposes improved prevention and workflow as explanations, but mediation by any one component was not established.1

Overvalue Warning

Do not call this proof that AI alone prevents stroke, that every hospital will reproduce the result, or that better performance scores necessarily mean less disability.1 The intervention bundled software, local integration, training, and support; severe-stroke and thrombectomy use, costs, and hospital days were not resolved.1 Also, the table's adjusted 12-month mortality estimate was not significant (HR 0.77, 95% CI 0.52–1.13; P=.18); the Figure 3 mortality log-rank display is a different, unadjusted analysis and should not replace it.1

Educational use notice. The Time Is Brain Review is dedicated to medical education. Figures from published articles are embedded for teaching purposes, with attribution to the original publication. We do not claim copyright over any embedded figure; all rights remain with the original publisher and authors under the article's published license. If you are a rights holder and would like a figure removed, please contact the editorial team. See our editorial notice for the full policy.