Stroke Neurology · Int J Neuroscience · 2026

Thrombolysis After the Window Closes

The Scenario

The 4.5-hour thrombolysis window leaves many stroke patients on the wrong side of the clock. The trials that pushed past it used different doses and selection rules, and they disagree.

No single pairwise comparison could answer whether late alteplase helps and at what dose. The dosing regimens had never all met in one trial, so the comparison had to run through a network.

From Question to Answer

Clinical question

Beyond 4.5 hours, does alteplase still help, and does the dose matter?

The evidence

Late-window and wake-up stroke RCTs with inconsistent arms and doses

The design

One network, estimated twice: frequentist and Bayesian, with GRADE certainty on every comparison

The answer

Functional benefit with a quantified bleeding cost, published with certainty ratings attached

The Decisions That Mattered

A network, because the doses never met in one trial.

Network meta-analysis borrows strength through shared control arms, so regimens compared only indirectly still get an estimate.

The same network estimated in two frameworks.

The frequentist model ran in netmeta, the Bayesian model in gemtc with MCMC. When two philosophies of inference agree, the result is not an artifact of either.

GRADE certainty attached to each comparison.

A pooled estimate without a certainty rating invites overconfidence. Clinicians got both the number and how much to trust it.

Arm-level quality control before any model ran.

Scripted checks on every trial arm caught mislabeled arms and a double-counted trial in the source literature. The model is only as good as the network it sits on.

Overview

Problem

Late-window thrombolysis evidence is fragmented across doses and trial designs that never met head to head.

Approach

Dual-framework network meta-analysis (netmeta and gemtc MCMC) with GRADE and scripted arm-level QC.

Outcome

Published in The International Journal of Neuroscience, March 2026. An independent reproduction audit later re-derived the published estimates exactly.

Figures

Illustrative figure from simulated data; the real analysis stays with the journal and the clinical team.

Reproducible R Code

Two frameworks, one network

# Frequentist side
pw  <- pairwise(treat, event = events, n = total,
                studlab = study, data = arms, sm = "RR")
net <- netmeta(pw, random = TRUE, reference.group = "Control")
netgraph(net); netrank(net)

# Bayesian side: same network, different engine
mtc <- mtc.network(arm_data)
mod <- mtc.model(mtc, likelihood = "binom", link = "log")
res <- mtc.run(mod)          # MCMC
summary(res)

# Agreement between the two frameworks is the robustness check.

Publication

View on ResearchGate

Read publication ↗