Most rodent stroke drugs that work in one laboratory fail in the clinic. One recurring reason is that each lab tests young, healthy, mostly single-sex animals under its own surgical habits, and no single lab has enough animals to see whether an effect holds in the older, comorbid, or female cohorts that resemble patients. We propose a way for several stroke labs to jointly estimate how a treatment changes the time course of recovery in one chosen target cohort, for example aged or female mice at one lab. Each lab shares only small summaries of its data, and the estimator borrows from a partner lab only when that lab's animals look compatible with the target. The recovery signal comes from automated 24/7 home-cage tracking, which records movement without handling the animals. From it we define a time-to-event outcome: the day an animal's night-time activity returns to its own pre-stroke baseline.
The methodological core adapts a recent federated causal survival estimator. That method has each site send per-animal correction terms (efficient influence functions, i.e. the first-order corrections that make an estimate robust to errors in either the outcome model or the treatment model) and learns sparse weights over sites. Our change is to have sites send these terms projected onto a smooth basis in time, rather than evaluated at every distinct event time. This addresses the jagged curves and communication costs the original authors flag. We are honest that rodent data carry no patient-privacy obligation. The federated design is therefore justified by unpublished compound data and the size of raw imaging and sensor archives, and the experiments explicitly test whether it costs anything relative to pooling.
Core Claim
A doubly robust, site-selective estimator of the treatment effect on recovery-time curves can be computed across labs from basis-projected summaries. It will (a) match a fully pooled version of the same estimator, (b) beat target-lab-only estimation and classical meta-analysis in the small-sample, heterogeneous regime typical of preclinical stroke, and (c) give a better-calibrated prediction of the effect actually observed in a held-out comorbid cohort than an average across all labs. Claim (c) is the translational one. Claims (a) and (b) are what make the method usable.
Where this sits in prior work
Federated learning in stroke so far is clinical and predictive. Federated Learning for MRI-based BrainAGE trains models across 16 thrombectomy centers with federated averaging (FedAvg, which averages model weights from local training) to predict brain age and 3-month outcome. Federated Prediction Model for Stroke Prevention predicts stroke risk from electronic health records across five hospitals, and federated EEG and rehabilitation-robot work follows the same pattern (Federated GNNs for EEG-Based Stroke Assessment, Federated Joint Learning of Robot Networks in Stroke Rehabilitation). These papers motivate the proposal by showing that multi-site training helps weak sites. All of them predict fixed-timepoint outcomes, though. None estimates a treatment effect, none uses time-to-event outcomes, and none touches preclinical data.
The method comes from federated survival statistics. Privacy-Protected Causal Survival Analysis Under Distribution Shift shares influence functions with penalized site weighting. Its authors report RMSE reductions "by up to 59%" at a target site on flchain and HIV trial data, and they name its limitation themselves: communication scales with the number of time points, and the resulting curves are jagged. CAFTA obtains centralized-quality accelerated failure time estimates from center-level summaries. It is confined to parametric failure distributions, and it showed that fixed- and random-effects meta-analysis miss effects that pooled analysis detects. That finding constrains our design: plain meta-analysis is not an adequate default.
The application side supplies the outcome. Automated Analysis of Stroke Mouse Trajectory Data With Traja records position every 250 ms for 35 days after transient middle cerebral artery occlusion (a filament briefly blocks the main forebrain artery and is then withdrawn). It finds a Fortasyn diet effect on velocity in days 1–3 and declining daytime activity through day 33, but analyzes these as group curves rather than per-animal recovery events. The translational-failure literature, for example Sex-Specific Adaptations in Alzheimer's Disease and Ischemic Stroke, motivates choosing comorbid or sex-specific target cohorts.
Research Program
Outcome construction. For each animal we fit its pre-stroke nocturnal activity and velocity distribution over the baseline week. We then define recovery time $T$ as the first post-stroke day on which a three-night rolling mean re-enters the animal's own baseline band. Animals are censored at end of follow-up or at death. We pre-specify the band width and run sensitivity analyses across it. Nocturnal windows are used because Traja documents that daytime testing disrupts activity.
Estimand. The estimand is the target-cohort treatment effect on the recovery curve, $\Delta(t) = S_1(t) - S_0(t)$. Here $S_a(t)$ is the probability of not yet having recovered by day $t$ under treatment $a$, evaluated over the covariate distribution (age, sex, comorbidity model, occlusion duration, baseline activity) of the target cohort.
Basis-projected influence sharing. Each site $k$ fits local outcome and censoring models and computes influence terms for the target estimand. It then sends only their projections onto a fixed $J$-dimensional B-spline basis $\{b_j(t)\}$ over days 0–35, together with a $J\times J$ covariance summary. The combined estimate is
$$\hat\Delta(t) = \sum_{j=1}^{J} b_j(t)\Big(\hat\theta_{0j} + \sum_{k} w_k\,\hat\delta_{kj}\Big),$$
where $\hat\theta_{0j}$ are the target lab's own coefficients, $\hat\delta_{kj}$ is site $k$'s coefficient-level discrepancy-corrected contribution, and $w_k \ge 0$ are site weights chosen by an $\ell_1$ penalty on the discrepancy between site $k$ and the target. This follows the selection logic of the Privacy-Protected paper, moved from time points to basis coefficients. Communication per site is $O(J^2)$ numbers regardless of how many event days occur, and the curve is smooth by construction. The cost is a bias from projection that shrinks as $J$ grows. We will characterize this trade-off and choose $J$ by cross-validation within the target site.
Disclosure control. Summaries are passed through DataSHIELD-style minimum-cell filters (DataSHIELD). This matters more for protecting unpublished compound results than for animal privacy.
Baselines
The naive reference point is target-only estimation: the same doubly robust estimator run on the target cohort's animals alone. This is what a single lab does today. Beating it means lower RMSE of $\Delta(t)$ integrated over days 0–35 in semi-synthetic settings, and narrower confidence bands with nominal coverage on real data.
The upper reference is the pooled oracle: the identical site-selective estimator computed on centrally pooled per-animal data. It is the right ceiling because rodent data can legally be pooled. Our claim (a) is satisfied if the federated estimate lies within Monte Carlo error of it.
The strongest existing federated competitor is the per-time-point influence-function estimator of Privacy-Protected Causal Survival Analysis Under Distribution Shift. It is run with identical nuisance models, the same site-selection tuning grid, and the same folds. Its reported benchmark is up to 59% RMSE reduction over target-only estimation on flchain/AMP. We will first reproduce its setting on flchain before moving to stroke data. We count our estimator as beating it if it matches its RMSE while using fewer transmitted numbers and producing curves with lower total variation, or if it achieves lower RMSE at equal communication.
We also run CAFTA (Collaborative Inference for Accelerated Failure Time Model) with Weibull and log-normal fits, plus random-effects meta-analysis of per-lab restricted mean recovery time. There is no reported result on Lacuna for any of these methods on rodent home-cage recovery outcomes, so all stroke-domain numbers will be new.
Two ablations isolate the key ingredients. One sets all $w_k$ equal, removing site selection. The other replaces the spline basis with the per-time-point grid.
Why not just…?
Why not just pool the data, since mouse data are not regulated? In many consortia this is exactly right, and the pooled oracle is in our design for that reason. Pooling does not make the core statistical problem go away: unweighted pooling estimates an average over young, healthy cohorts rather than the comorbid target. The site-selective estimator is therefore the contribution whether or not data are pooled. Federation adds one claim: labs that will not release per-animal data on proprietary compounds can still participate at no accuracy cost. If the oracle comparison showed a meaningful gap, we would report the federated route as a compromise rather than a free lunch.
Why not just run a random-effects meta-analysis or a mixed Cox model with lab effects? The CAFTA paper reports that fixed- and random-effects meta-analysis showed bias, under-coverage, and a missed covariate effect that pooled analysis detected. A pooled mixed Cox model with lab-by-covariate interactions is a serious threat on pooled data. We include it as an additional pooled baseline. If it matches the site-selective estimator on transported-effect accuracy, we will say so and narrow our claim to the federated protocol and smooth-curve communication.
Why not just run one large, well-powered single-lab study with comorbid animals? Single-site training was worse than federation in both clinical stroke papers: brain-age MAE was 7.04 years single-site versus 6.46 years federated, and AUC rose from 0.701 to 0.813 for the small hospital. More importantly, one lab's surgical idiosyncrasies are exactly what fails to translate. The held-out-cohort test below measures whether multi-lab borrowing predicts a new cohort better than any single lab's estimate does.
Why not just use CAFTA, which already works from summaries? CAFTA requires a parametric failure distribution and does not target a shifted covariate population. Recovery times in mice are likely multimodal, because some animals never recover within 35 days. The semi-synthetic arm includes such mixtures to show where the parametric form breaks.
Proposed Experiments
The first stage is a semi-synthetic test on the Traja/DVC tMCAO trajectories. We derive real per-animal recovery times and covariates from them. We then create 4–8 virtual labs by resampling with controlled shifts: age and sex mix, occlusion duration, site-specific treatment effects, and some deliberately incompatible labs. We inject a known treatment effect into the recovery-time model. This lets us report RMSE, coverage, curve total variation, and transmitted numbers for every baseline, with known truth.
The second stage is a method sanity check on flchain. We reproduce the Privacy-Protected paper's target-site gains before claiming improvements.
The third stage is a prospective consortium of four to six labs running DVC cages after transient occlusion with one shared candidate. We would use Fortasyn, given the prior signal. One lab contributes an aged or female cohort that is never used in estimation.
The decisive check is pre-registered. Using only the other labs' summaries plus the held-out cohort's covariates, we predict that cohort's $\Delta(t)$ and compare it with the effect directly observed there. The proposal succeeds if the site-selective transported estimate is closer to the observed effect, with valid coverage, than both the unweighted multi-lab average and the best single-lab estimate. It fails if site selection makes no difference to that error. In that case the honest conclusion is that preclinical heterogeneity is not captured by measured covariates, which would itself be a useful negative finding about translation.
References
- Federated Learning for MRI-based BrainAGE: a multicenter study on post-stroke functional outcome prediction
- Privacy-Preserving Technology to Help Millions of People: Federated Prediction Model for Stroke Prevention
- Federated GNNs for EEG-Based Stroke Assessment
- Federated Joint Learning of Robot Networks in Stroke Rehabilitation
- Privacy-Protected Causal Survival Analysis Under Distribution Shift
- Collaborative Inference for Accelerated Failure Time Model Using Clinical Center‐Level Summary Statistics
- Automated Analysis of Stroke Mouse Trajectory Data With Traja
- Sex-Specific Adaptations in Alzheimer’s Disease and Ischemic Stroke: A Longitudinal Study in Male and Female APPswe/PS1dE9 Mice
- DataSHIELD: mitigating disclosure risk in a multi-site federated analysis platform