Federated analysis reproduces a major acute stroke trial without moving patient-level data

A privacy-preserving analytical framework reproduced the non-inferiority findings of the AcT stroke trial without transferring individual patient data across jurisdictions. In a field where stroke research often spans multiple centers, provinces, and regulatory systems, the result is more than a technical exercise: it shows that trial-grade inference can be preserved even when raw data cannot leave the hospital or the country where they were collected.

The study used Starfish-FL, a general-purpose federated analysis platform, and applied it to the AcT randomized controlled trial dataset drawn from 21 stroke centers across 8 Canadian provinces. The trial cohort included 1,577 patients in total, with 806 assigned to tenecteplase and 771 to alteplase. Investigators partitioned the data by site to mimic a federated environment, then had each site compute treatment-effect estimates locally before combining those outputs in a fixed-effects federated meta-analytic framework. The clinical question remained the same one that confronts emergency stroke teams at the bedside: whether tenecteplase performs no worse than alteplase for patients with acute ischemic stroke.

The trial’s primary endpoint was excellent functional outcome, defined as modified Rankin Scale 0–1 at 90–120 days. In the reproduced analysis, the crude estimator generated a risk difference of 2.14% with a 95% confidence interval from −2.61 to 6.88%. The published centralized analysis had reported 2.1% with a 95% confidence interval from −2.6 to 6.9%. The two results lined up almost exactly, down to the estimate and the interval bounds. The site-stratified federated estimator was similarly close, yielding 2.12% with a 95% confidence interval from −2.53 to 6.77%. Non-inferiority was met under both approaches.

Precision held across all reported endpoints

The reproducibility extended beyond the primary endpoint. All four outcomes reproduced the published result at the same precision: mRS 0–2 at 0.80% against 0.8%, symptomatic intracerebral haemorrhage at 0.23% against 0.2%, and 90-day mortality at −0.11% against −0.1%. In practical terms, the federated framework did not just approximate the central analysis; it reconstructed it closely enough to preserve the trial’s inferential conclusion.

The statistical behavior of the system mattered because federated methods can fail in subtle ways. Differences in site-level case mix, local workflow, or missingness can distort treatment-effect estimates once data are split and reassembled through aggregate outputs alone. Here, the agreement between the federated and centralized approaches suggests that the underlying treatment effect was stable enough to survive decentralization, and that the platform preserved the precision required for a non-inferiority assessment. In acute stroke research, where treatment decisions are time-sensitive and networked across many hospitals, that kind of methodological durability is not an abstraction. It is the difference between a data governance barrier and a usable multicenter evidence base.

A single sentence captures the most consequential point: non-inferiority survived decentralization.

The study also tested ten sensitivity specifications, and non-inferiority held in every one of them. Across those analyses, the lower confidence bound ranged from −2.95 to −2.09%. Consistency across alternative specifications strengthens confidence that the federated framework was not leaning on one fragile modeling choice or a single favorable partition of the data. The result instead suggests a stable analytical architecture that can recover the same clinical answer even when the data remain distributed at source.

Why privacy-preserving stroke analytics could matter beyond one trial

The biological and operational logic is easy to appreciate at the bedside even if the engineering looks complex. Acute stroke trials depend on rapid enrollment, standardized endpoint assessment, and large multicenter collaboration. Yet privacy rules, institutional governance, and cross-border transfer restrictions often make central pooling of individual patient data slow or impossible. Federated analysis addresses that bottleneck by allowing standardized scripts to run locally while only aggregate outputs leave the site. In effect, the data stay where they were generated, but the scientific signal still moves.

For stroke care, that matters because treatment questions rarely stop at national boundaries. If a methodology can reproduce a large randomized trial while respecting data sovereignty, it becomes easier to validate findings across health systems that would otherwise remain isolated. It also opens a path for future global studies in which regulators, hospitals, and ethics boards may permit local computation even when they will not permit raw data transfer. The value is especially clear for comparative effectiveness studies like AcT, where the question is not whether thrombolysis works, but which thrombolytic strategy can be supported by evidence that is both robust and operationally feasible to generate.

This report does not change the clinical hierarchy of alteplase and tenecteplase on its own. It does show that a privacy-preserving framework can reproduce the statistical backbone of a multicenter stroke trial, including its non-inferiority conclusion and secondary endpoints, without moving individual records across borders. The next test is not whether the platform can match one trial under ideal conditions, but whether it can do the same across other stroke datasets, other study designs, and other jurisdictions with more complicated governance demands. That is where the method will either prove broadly reusable or remain a carefully validated solution to a single well-structured problem.


Reporting by Dr. A. Manasa, MBBS for Medicine World Foundation, with additional data provided by British Journal of Cancer.