Can hospitals study dementia-drug signals without sharing patient records? A federated method matched pooled estimates—but did not prove treatments work
FLAT-WHITE reproduced centralized estimates from four US health-data networks using one exchange of summary statistics. Twenty-one drug associations also appeared in a 339,180-person external cohort, but residual confounding means none is a treatment recommendation.
The short version
- What happened
- The method exchanged summary statistics once across four US learning-health systems and reproduced the effect estimates and standard errors of a centralized pooled analysis.
- Who it affects
- Hospitals may be able to collaborate without exporting patient-level records, but people with memory problems should not change medication on the basis of these retrospective signals.
- What remains unproven
- Retrospective electronic-health-record study
Keep in mind
Retrospective electronic-health-record study
Residual confounding and informative censoring remain possible
Drug exposure was simplified to binary initiation
Common outcome-model assumption across sites
Odds ratios do not directly express absolute risk
No randomized treatment or patient-facing outcome trial
Source analysis · 1 linked source. Primary analysed source date: 27 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
One exchange of summaries reproduced the pooled calculation
A peer-reviewed study reports a way for hospitals to run large observational treatment comparisons without sending patient-level records to a central repository. FLAT-WHITE fits propensity-score models locally, applies overlap weights, then transmits summary statistics once. A coordinating site reconstructs the combined likelihood and calibrates estimates with negative-control outcomes. In the paper's tests, the federated estimates and standard errors were nearly identical to those produced by pooling the individual records.
That is a technical result, not a new therapy. The useful advance is that research partners with legal or governance barriers to sharing records may be able to ask the same prespecified question across sites while keeping identifiable-level data behind each institution's boundary. It could make otherwise impossible comparisons feasible without treating federated analysis as anonymous or risk-free. The approach also reduces repeated communication between sites, although it still depends on common definitions, harmonised covariates and correct local implementation.[1]
The dementia case study was large—and still observational
The researchers applied the framework to 123,971 people aged 50 or older with mild cognitive impairment across INSIGHT, OneFlorida+, Penn Medicine and Yale New Haven. They emulated 42,700 single-site and 9,050 multi-site comparisons for 181 prescribed drugs, using 29 supplements as the main active-control group and following records for a dementia diagnosis, loss to follow-up or five years. Propensity models used 268 measured baseline covariates; false-discovery-rate correction addressed the many comparisons.
Twenty-one drug associations were statistically significant in the four-system analysis and appeared again in a TriNetX cohort of 339,180 eligible patients. Yet effect sizes often shrank in external data. Celecoxib, for example, had an odds ratio of 0.34 in FLAT-WHITE but 0.86 in TriNetX; ketorolac was 0.41 versus 0.94. That attenuation is a warning against reading the associations as prescriptions. People receive these drugs for different illnesses, and those illnesses, care patterns and follow-up can also alter recorded dementia risk.[1]
The distinction that matters
What the evidence pipeline actually did
The method separates the privacy result from the clinical claim that still needs testing.
Four health systems
Fit local models on 123,971 records
Patient-level data stayed at each site
Federated calculation
Exchange summary statistics once
Matched centralized estimates
External check
Repeat signals in 339,180 TriNetX records
21 associations remained observational
Privacy-preserving analysis does not remove causal uncertainty
Overlap weighting improved balance on measured factors, and negative controls plus E-values tested sensitivity to hidden bias. Even so, confidence-limit E-values ranged as low as 1.11, meaning modest unmeasured confounding could explain some signals. Informative censoring, miscoded diagnoses, treatment switching and misspecified propensity models remain possible. The method also assumes a common outcome model, handles binary rather than time-varying treatment and estimates an effect for the population where treatment groups overlap—not every patient.
For researchers, the result supports a governed route to multi-site screening of hypotheses. Data stewards would still need access controls, disclosure review and checks on whether summary outputs can leak information about small groups. For patients and clinicians, it does not support starting celecoxib, steroids, antibiotics or other flagged drugs to prevent dementia; several carry substantial harms, and earlier trials have not established such benefit. What would change the assessment is preregistered replication with transparent code, stronger designs addressing time-varying treatment and confounding, followed by randomized trials measuring cognition, dementia diagnoses and adverse events.[1]