跳到正文
原文
Nature 健康护理主题·· 24 天前精选AI 评分78

FLAT-WHITE:基于联邦学习的阿尔茨海默病药物重定位目标试验模拟方法

FLAT-WHITE: from target trial emulation to federated high-throughput causal machine learning using decentralized real-world data

AI 导读

在四个覆盖超40万患者的分布式医疗数据网络中,FLAT-WHITE算法通过单次通信实现了与集中分析等效的因果推断,发现21种常用药(如布美他尼、螺内酯等)与阿尔茨海默病相关认知衰退风险降低显著(OR 0.34-0.50,95%CI 0.15-0.85),并通过TriNetX外部验证(覆盖152.7万患者)确认了结果的稳健性。

推荐理由

FLAT-WHITE 算法通过联邦学习框架实现了多中心观察性数据的隐私保护性目标试验模拟,其核心创新在于单次通信即可完成与集中分析等效的因果推断,为阿尔茨海默病药物重定位提供了可扩展的方法论工具。

正文

Introduction

The growing availability of real-world data (RWD) has transformed clinical research by enabling large-scale observational studies that generate real-world evidence (RWE) to inform medical decision-making1,2,3. These data offer longitudinal, high-dimensional, and routinely collected clinical information that can be harnessed to evaluate the effectiveness and safety of medical interventions. The U.S. Food and Drug Administration (FDA) has actively supported the use of RWD and launched several initiatives, such as the Sentinel System and the Biologics Effectiveness and Safety (BEST) Initiative4, to promote methodological advancements for evaluating drug indications, with the goal of improving the quality and acceptability of RWE-based approaches5,6.

However, drawing causal inferences from observational data presents substantial challenges, particularly in the presence of unmeasured confounding. To address these issues, the framework of target trial emulation (TTE) has been proposed as a principled approach for using observational data to mimic the design and interpretation of randomized controlled trials. This framework requires careful attention to design elements such as eligibility criteria, treatment assignment, and follow-up definitions, along with appropriate adjustment for confounding7,8. Target trial emulation9 has had a growing social impact by enabling timely, equitable, and policy-relevant evidence generation in contexts where randomized trials are impractical, such as assessing treatment effectiveness during public health emergencies, studying vulnerable populations, or supporting regulatory decisions with real-world data.

Although TTE offers a rigorous framework for causal inference using observational data, the application of strict eligibility criteria often results in substantially reduced sample sizes. This sometimes precludes research on complex conditions or rare diseases because the resulting cohorts may be too small to yield reliable estimates or adequately capture relevant subpopulations. To extend the reach of TTE with scientific rigor to a broader community, there is a growing need to conduct collaborative TTE by leveraging data from multiple sites. Such approach enhances statistical power and ensures adequate representation of diverse populations, both of which are essential for improving the generalizability of clinical findings. Collaborative multi-site or multi-country studies thus play a pivotal role in strengthening statistical inference, enhancing the external validity of research findings, and addressing scientific questions that cannot be resolved using a single dataset.

Over the past decades, numerous distributed research networks (DRNs) have been established to support collaborative multi-site or multi-country analyses. A notable example is the FDA Sentinel System10,11, a national DRN developed for active medical product safety surveillance. The success of Sentinel reflects the growing recognition of DRNs as a viable and scalable model for real-world evidence generation. However, despite this progress, many emerging or independently developed DRNs still face significant barriers due to strict data sharing regulations, such as HIPAA in the United States and GDPR in Europe, which prohibit the sharing of individual patient-level data across institutions12,13,14,15,16. Despite the promise of DRNs, their practical implementation remains logistically burdensome, complex legal agreements, secure data transfer mechanisms, and extensive back-and-forth coordination, often necessitating dedicated personnel at each participating site. Although a number of federated learning algorithms have been developed, for example iterative algorithms such as GLORE17 and WebDISCO18, and non-iterative algorithms such as ODAL19,20 and ODAC21, mainly for comparative effectiveness research, and FACE22 to estimate the target treatment effect in causal inference. However, none of these algorithms have been specifically designed for collaborative TTE analysis and meanwhile accounts for confounding bias with observational clinical data.

While TTE with real-world data provides a structured approach for causal inference from observational data, it remains limited in scope and scalability. In the era of large-scale EHR data and artificial intelligence (AI), a high-throughput TTE framework is needed to systematically emulate and evaluate hundreds of candidate interventions across diverse populations and health systems. To broaden the impact and methodological integrity of conducting target trial emulation using large-scale real-world data within distributed research networks that face data sharing constraints, we propose FLAT-WHITE: a Federated Lossless Algorithm for Targeted Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication. FLAT-WHITE functions as a causal ML framework that automates high-throughput trial emulation across multiple sites without sharing patient-level data. By combining machine learning–based causal modeling with federated learning, it enables efficient and secure evidence generation at scale, transforming TTE from a manual analytic exercise into a data-driven, AI-enabled causal discovery process. Existing distributed learning methods, such as GLORE, ODAL, and ODAC, were primarily developed for association modeling or risk prediction rather than for estimating causal treatment effects in target trial emulation settings. Moreover, these are not one-shot methods: GLORE is iterative, whereas ODAL and ODAC typically require two to three rounds of communication to achieve stable estimation. FACE is also distinct from our setting, as it was developed for transfer learning and focuses on estimating the target average treatment effect. By contrast, FLAT-WHITE is designed specifically for federated causal inference in target trial emulation, enabling privacy-preserving treatment effect estimation across multiple sites without sharing individual-level data. Remarkably, FLAT-WHITE is also communication-efficient, requiring only a single round of information exchange while achieving lossless performance equivalent to centralized pooled analyses. The framework integrates a negative control outcome (NCO) calibration module, which can be executed concurrently with federated aggregation, to address unmeasured confounding and reduce systematic bias. By embedding overlap weighting23 on the propensity score, FLAT-WHITE enhances robustness in causal effect estimation, particularly in settings with poor covariate overlap and heterogeneous treatment populations.

The proposed FLAT-WHITE algorithm offers substantial real-world impact and is designed for scalability. To illustrate its utility, we conducted a large-scale case study on drug repurposing for Alzheimer’s disease and related dementia (ADRD), a critical public health challenge with limited therapeutic options and a rapidly growing patient population. This study leveraged four national health learning systems: INSIGHT ( > 5.5 million patients), OneFlorida+ ( > 24 million), the University of Pennsylvania Health System (UPHS) ( > 6.5 million), and Yale New Haven Health System (YNHHS) ( > 4 million). Within this network, we performed a high-throughput target trial emulation of 181 commonly prescribed therapeutic classes and agents, evaluating their potential to influence progression from mild cognitive impairment (MCI) to ADRD. Findings were further validated using the TriNetX global research network ( > 110 million), comprising 339,180 patients as an independent external cohort.

This real-world application underscores the power of FLAT-WHITE to strengthen causal inference with observational clinical data across decentralized research networks. The method enables secure, large-scale analysis while addressing data-sharing constraints by operating exclusively on summary-level statistics. FLAT-WHITE not only safeguards patients’ confidentiality but also integrates data from diverse healthcare systems, thereby maximizing the utility of EHRs for generating actionable clinical evidence in a transparent and methodologically rigorous manner. Unlike traditional federated approaches that rely on iterative updates or approximations, FLAT-WHITE produces lossless, reproducible results while functioning as a causal ML pipeline for population-scale discovery, making it a scalable and trustworthy framework for conducting real-world causal machine learning to accelerate therapeutic discovery and advance collaborative clinical research.

Results

In this paper, we propose a novel federated learning algorithm named FLAT-WHITE, which stands for “A Federated Lossless Algorithm for Targeted Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication”. Given the significant barriers to data sharing across decentralized distributed research networks, FLAT-WHITE offers a communication-efficient and high-throughput solution for conducting target trial emulation without sharing patient-level data. As shown in Fig. 1(a), designed as a causal ML pipeline, FLAT-WHITE automates large-scale and parallelizable target trial emulations across a large number of candidate treatments for drug repurposing, enabling rapid discovery while protecting patients’ confidentiality and methodological integrity.

Fig. 1: Overview of FLAT-WHITE pipeline, databases for internal and external validation, and drug selection pipeline.

a Schematic of FLAT-WHITE: a federated high-throughput causal machine learning framework for multi-site target trial emulation for drug repurposing. b Workflow of the FLAT-WHITE Algorithm for Target Trial Emulation. The process begins with local propensity score modeling and overlap weight calculation at each site using patient-level data. Each site then computes summary-level statistics for the outcome of interest and multiple negative control outcomes (NCOs), which are securely transmitted for aggregation. In Stage B, the global log-likelihood functions are reconstructed and maximized to estimate the treatment effect. In Stage C, NCO-based calibration is applied to correct for unmeasured confounding, yielding the final calibrated effect estimates and confidence intervals. b Demonstration of the lossless property of FLAT-WHITE: treatment effect estimates and standard errors (of partial drugs for visualization purpose) from FLAT-WHITE are nearly identical to those from pooled analyses. c Study cohort inclusion criteria and sample sizes from four large-scale learning health systems (Penn Medicine, Yale, INSIGHT, OneFlorida + ) and external validation using the US collaborative networks in TriNetX. d Drug selection pipeline: drugs prescribed to >= 100 patients for two consecutive time points (>= 30 days apart) were evaluated. For each drug, treatment and control groups were defined based on MCI diagnosis prior to drug exposure, with 29 vitamin/supplement controls used for comparative analysis.

The proposed framework deploys a robust and principled approach to addressing confounding bias inherent in observational clinical data by employing overlap weighting based on propensity scores. By assigning weights to each patient that are proportional to the probability of that patient belonging to the opposite treatment group, overlap weighting strategy mitigates the influence of patients with extreme propensity scores. Compared with the commonly used Inverse Probability of Treatment Weighting (IPTW) method, the overlap weighting method ensures improved covariate balance between treatment groups, making causal estimates more reliable. To further enhance the validity of treatment effect estimation, FLAT-WHITE incorporates a negative control outcomes (NCOs) calibration procedure, which helps mitigate residual systematic bias and improves the robustness of causal inference. The list of NCOs is available in Supplementary Appendix C.

Figure 1(b) shows the workflow of FLAT-WHITE, which includes three major stages: (a) Stage A – Local Propensity Score Modeling: At each participating site, patient-level data are used to independently fit a propensity score model (i.e., a l1 penalized logistic regression model) based on a pre-specified, harmonized, and common set of covariates. These models yield site-specific propensity scores, which are then used to compute overlap weights to balance the treatment and control groups. This step ensures covariate comparability across groups and enables the construction of a pseudo-population for valid causal inference. Further technical details on the calculation of overlap weights and the list of covariates are provided in the Methods section and Appendix B. (b) Step B – Treatment Effect Estimation at the Coordinating Center: Each site transmits only summary-level statistics to the central coordinating center. Using these summary statistics, the coordinating center estimates the overall treatment effect by reconstructing the multi-site log-likelihood of the overlap-weighted logistic regression model. (c) Step C – Negative Control Outcomes (NCOs) Calibration at the Coordinating Center: To mitigate residual biases induced by unmeasured confounding variables, the coordinating center applies a NCOs calibration procedure. This involves leveraging multiple pre-specified outcomes that are not causally affected by the treatment but may still be influenced by unmeasured confounding. By estimating an empirical “null” distribution from these NCOs, the treatment effect estimates obtained in Stage B can be calibrated, thereby enhancing the robustness and validity of causal inferences drawn from observational data.

In Table 1, we provide a detailed summary of the FLAT-WHITE algorithm, including step-by-step implementation procedures.

Table 1 The proposed FLAT-WHITE algorithm: Three major stages are involved in the process

Full size table

The proposed FLAT-WHITE has a few key properties: (1) information-Lossless: FLAT-WHITE guarantees identical results to those obtained from a centralized pooled analysis, ensuring no information loss despite operating under a federated setting. This property enables rigorous statistical inference while respecting data decentralization; (2) communication-efficient: The algorithm requires only a single round of communication between participating sites to transfer the summary-level statistics to the coordinating center, significantly reducing communication overhead and making it highly suitable for real-world distributed research networks with limited bandwidth or coordination capacity. This feature eliminates the need for synchronization, which typically requires intensive human-in-the-loop coordination in iterative algorithms, thereby improving scalability; (3) secure and trusted environment for sharing summary-level statistics: FLAT-WHITE is designed to operate without sharing patient-level data across institutions. By exchanging only summary-level statistics, the algorithm adheres to data sharing regulations and supports secure, collaborative analysis across heterogeneous data sources; (4) robust treatment effect estimation: By integrating overlap weighting to address confounding and incorporating negative control outcome calibration to mitigate residual bias, FLAT-WHITE delivers reliable and robust estimates of causal treatment effects from observational data.

To demonstrate the lossless property of FLAT-WHITE, i.e., its ability to produce results identical to those obtained from pooled individual-level data, we conducted a simulation study using data from the University of Pennsylvania Health System (UPHS). The dataset was randomly partitioned into five distinct sites to emulate a multi-site, decentralized study. Our objective was to evaluate the treatment effects of 181 drugs for patients with mild cognitive impairment (MCI) on the progression to Alzheimer’s disease and related dementias (ADRD).

In this simulation setting, we first performed a conventional pooled analysis, assuming full access to patient-level data across all five sites. In parallel, we applied FLAT-WHITE under the assumption that data could not be shared between sites, and only summary statistics were available. The comparison of treatment effect estimates from both approaches is shown in Fig. 2. As illustrated in the figure, the treatment effect estimates obtained via FLAT-WHITE exhibit a high degree of concordance with those from the pooled analysis, nearly all estimates lie on the diagonal line, indicating near-perfect agreement. The absolute difference in effect estimates was minimal, with a mean difference of 1\(6.62\times {10}^{-5}\) and a maximum difference of 0.0177. For the standard errors, the mean and maximum differences were −0.00015 and 0.001, respectively, values that can be attributed to numerical precision limits. These results demonstrate that FLAT-WHITE is lossless, achieving estimation accuracy equivalent to pooled analyses while maintaining data security by relying solely on summary-level data.

Fig. 2: Simulation results.

For illustration, we selected 10 drugs with the largest differences in effect size estimates between FLAT-WHITE and pooled analysis from the 181 candidate drugs for visualization. Left panel: comparison of estimated effect sizes between the pooled analysis and FLAT-WHITE. Right panel: comparison of estimated standard errors between the pooled analysis and FLAT-WHITE.

We have conducted an empirical target trial emulation study to investigate drug repurposing for Alzheimer’s disease and related dementias (ADRD). ADRD pose a growing public health crisis, with limited treatment options and a rapidly increasing patient population. In response, drug repurposing, identifying new therapeutic uses for existing medications, has emerged as a cost-effective and time-efficient strategy to accelerate the development of interventions for complex neurodegenerative diseases like ADRD. In this study, we collaborated closely with four large-scale health learning systems: the INSIGHT Clinical Research Network ( > 5.5 million patients), OneFlorida+ Clinical Research Network ( > 24 million patients), the University of Pennsylvania Health System (UPHS) ( > 6.5 million patients), and the Yale New Haven Health System (YNHHS) ( > 4.0 million patients). Together, these systems form a unique, large-scale decentralized research network encompassing over 40 million patients. This extensive and diverse patient population enabled us to conduct a high-throughput target trial emulation to evaluate the repurposing potential of 181 frequently prescribed therapeutic classes/agents for dementia prevention. We aimed to identify the drugs that may offer neuroprotective effects and reduce the risk of progressing from MCI to ADRD. To further validate our findings, we used the TriNetX data networks ( > 110 million patients) as an independent external validation cohort, which includes data from over 330,000 eligible patients, as illustrated in Fig. 1(c). This additional evaluation enabled us to assess the generalizability and robustness of the identified drug candidates across distinct populations and data environments.

In this study, we included a total of 123,971 patients diagnosed with mild cognitive impairment (MCI), all aged 50 or older at the time of diagnosis. Patients were required to have no prior history of Alzheimer’s disease or related dementias (ADRD) and no exposure to the trial drug prior to the index date, which is defined as the first the date the patient was first prescribed the study drug for over 30 days with all eligibility criteria met. To be eligible, patients could not have a prior diagnosis of Alzheimer’s disease or related dementias (ADRD), nor any exposure to the study drug before the index date. The index date was defined as the first date on which the patient received a prescription for the study drug lasting more than 30 days, provided all other eligibility criteria were met at that time. For each patient, follow-up starts from the baseline date until the date of first ADRD diagnosis, loss to follow-up, or five years after the baseline, whichever happens first. These patients were identified across four large-scale health learning systems. The primary outcome of interest was the onset of ADRD, defined as a binary outcome. The treatment effect was quantified using the odds ratio (OR) as the summary statistical measure. To identify potential drug candidates, we applied the proposed FLAT-WHITE algorithm to independently perform target trial emulations for each of the 181 drugs, followed by false discovery rate (FDR) correction to account for multiple hypothesis testing. The process for selecting the patient cohort and drugs, along with the methods used for validation, are detailed in Figs. 1(c) and 1(d). Detailed characteristics and target trial emulation specifications can be found in Method section and Appendix B. We also examined covariate balance between the two groups before and after overlap weighting. Before weighting, several covariates, such as history of hypertension, race, history of hyperlipidemia, and sex, showed absolute standardized mean differences greater than 0.1. After weighting, nearly all absolute standardized mean differences were close to 0, suggesting markedly improved covariate balance and better overlap between the two groups. Detailed plots for all participating sites are provided in Appendix E.

For the external validation using the TriNetX data networks, the data used was collected on April 21, 2025 from the TriNetX Global Research Collaborative Network, which provided access to electronic medical records (diagnoses, procedures, medications, laboratory values, genomic information) from approximately 152.7 million patients from 106 healthcare organizations.

We conducted an independent external validation using EHR data from the TriNetX Global Research Collaborative Network. Collected on April 21, 2025, the TriNetX Global Research Collaborative Network provided access to de-identified electronic medical records, including diagnoses, procedures, medications, laboratory values, and genomic information, from approximately 152.7 million patients across 106 healthcare organizations. The data are available from the TriNetX Analytics Network (https://trinetx.com).

This retrospective study is exempt from informed consent. The data reviewed is a secondary analysis of existing data, does not involve intervention or interaction with human subjects, and is de-identified per the de-identification standard defined in Section §164.514(a) of the HIPAA Privacy Rule. The process by which the data is de-identified is attested to through a formal determination by a qualified expert as defined in Section §164.514(b)(1) of the HIPAA Privacy Rule. This formal determination by a qualified expert refreshed on December 2020.

Drug repurposing signals via target trial emulation with FLAT-WHITE

We present the results of our high-throughput target trial emulation using real-world data from four learning health systems, along with external validation using the TriNetX database. Figure 3 highlights 21 drugs identified as having statistically significant treatment effects that were consistently validated by the external dataset. Each subpanel in Fig. 3 corresponds to one candidate drug and compares six effect size estimates: those derived from each of the four individual learning health systems, the random-effects meta-analysis combining these site-specific results, the estimate from the proposed FLAT-WHITE federated method, and the corresponding estimate from TriNetX. This comprehensive comparison demonstrates the consistency, robustness, and external validity of our federated framework in detecting promising drug repurposing signals. To enhance clarity and facilitate interpretation, we categorized the 21 identified drugs into groups based on their therapeutic class or pharmacological category. The detailed results for each group are presented and discussed in the following subsections.

Fig. 3

Drug repurposing signals identified through multiple approaches: independent analyses across four learning health systems, random-effects meta-analysis of site-specific treatment effects, the proposed FLAT-WHITE federated method, and external validation using the TriNetX database. More detailed results are available in Supplementary Appendix D.

Although early clinical trials concluded that celecoxib was ineffective for treating Alzheimer’s disease24, more recent animal studies suggest otherwise. Research has shown that celecoxib can mitigate Aβ-induced neuroinflammation, synaptic dysfunction, and cognitive impairment in preclinical models25. One reason for the discrepancy may be that earlier trials focused on late-stage AD, where COX-2 expression is already reduced, possibly limiting celecoxib’s efficacy25. Our data support a potential protective role for celecoxib, showing a lower risk of ADRD in one cohort (FLAT-WHITE: OR 0.34, 95% CI 0.15, 0.78; TriNetX: OR 0.86, 95% CI 0.79, 0.94). Further investigation is warranted, particularly in early-stage AD.

Ketorolac, a potent nonsteroidal anti-inflammatory drug (NSAID), has been studied primarily for its potential protective effects on cognition. Preclinical models have shown that NSAIDs, including ketorolac, may reduce neuroinflammation and amyloid plaque accumulation, key features of AD pathology26. While earlier clinical and observational studies have yielded inconsistent results regarding NSAIDs’ effectiveness in AD prevention, recent systems pharmacology approaches have renewed interest in specific agents27. A recent study suggested that ketorolac may modulate microglial state transitions implicated in AD progression, and that its predicted neuroprotective effects were supported by retrospective analyses of electronic health records showing reduced AD incidence among ketorolac users by applying a single-cell microglial transcriptomic transition network model combined with real-world patient data to identify ketorolac as a promising repurposable candidate for AD28. These results, while preliminary, provide a novel mechanistic rationale and epidemiologic support for further investigation of ketorolac in AD prevention or treatment. Our study found lower risks for ADRD (FLAT WHITE: aOR 0.41, 95% CI 0.24, 0.70, TriNetX: aOR 0.94, 95% CI 0.89, 0.99), supporting its potential for AD prevention.

Corticosteroids such as methylprednisolone (FLAT-WHITE: OR 0.34, 95% CI 0.19, 0.61; TriNetX: OR 0.91, 95% CI 0.85, 0.96) and dexamethasone (FLAT-WHITE: OR 0.38, 95% CI 0.26, 0.55; TriNetX: OR 0.72, 95% CI 0.67, 0.78) were associated with a reduced risk of ADRD. These findings were consistent across main analysis and external site validation. These agents suppress microglial activation and reduce inflammatory cytokines like TNF-α, IL-1β, and IL-6—key players in neuronal injury and amyloid aggregation29,30. Dexamethasone may also influence serotonin signaling, which plays a role in synaptic resilience and neuroprotection31,32. Observational studies have reported lower dementia rates among long-term steroid users, particularly in inflammatory conditions such as rheumatoid arthritis33, reinforcing the role of inflammation in ADRD pathogenesis.

Prednisone has not shown consistent benefit in AD. A randomized trial involving 138 participants found no cognitive advantage after one year and even noted behavioral decline in the treatment group34,35. Some small studies observed changes in AD biomarkers with prednisone used for non-AD indications, but chronic use has been linked to reversible cognitive impairment. Nonetheless, our study reported a lower risk of ADRD associated with prednisone in both main analysis and external validation (FLAT-WHITE: OR 0.41, 95% CI 0.28, 0.62; TriNetX: OR 0.91, 95% CI 0.86, 0.96), suggesting the need for further exploration.

Acetaminophen also showed a strong protective association (FLAT-WHITE: OR 0.51, 95% CI 0.36, 0.72; TriNetX: OR 0.75, 95% CI 0.69, 0.82), and although its mechanism is unclear, antioxidant and anti-inflammatory actions may play a role36.

Empagliflozin, a sodium-glucose cotransporter 2 inhibitor (SGLT2i) for treating type 2 diabetes, has attracted growing interest for its potential cognitive benefits37,38,39. A target trial emulation using U.S. Medicare claims data reported that older adults with type 2 diabetes who initiated SGLT2i therapy, including empagliflozin, had a significantly lower risk of developing dementia compared to those receiving other second-line glucose-lowering therapies40, and our findings provide complementary support (FLAT WHITE: aOR 0.34, 95% CI 0.15, 0.79, TriNetX: aOR 0.65, 95% CI 0.59, 0.72). Potential mechanisms underlying these neuroprotective effects include attenuation of neuroinflammation, improved insulin sensitivity, and enhanced mitochondrial function41.

Spironolactone, a mineralocorticoid receptor antagonist widely prescribed for hypertension and heart failure, has garnered growing attention for its potential neuroprotective effects and was associated with reduced AD/ADRD risk (FLAT WHITE: aOR 0.50, 95% CI 0.29, 0.85, TriNetX: aOR 0.89, 95% CI 0.81, 0.97). Preclinical and translational studies suggest that spironolactone may attenuate cognitive decline by modulating aldosterone-dependent signaling and targeting hippocampal mineralocorticoid receptors, an area that remains underexplored in Alzheimer’s disease (AD) research42. Although evidence in humans is still emerging, recent real-world analyses using TriNetX data have shown that spironolactone use in older adults with heart failure and reduced ejection fraction is associated with a significantly lower risk of incident dementia and AD compared to non-use, though no difference was observed in cognitive decline per se43. These observational findings align with emerging evidence that pharmacologic modulation of the renin-angiotensin-aldosterone system may reduce AD risk in hypertensive populations44. While randomized trials are needed to confirm causality, these data support the hypothesis that RAAS-targeting agents like spironolactone may offer dual benefits for cardiovascular and cognitive health in aging adults.

Bumetanide, a loop diuretic approved for the treatment of edema and hypertension, has recently gained attention for its potential association with reduced risk of Alzheimer’s disease (AD). Several studies, including transcriptomic analyses and electronic health record (EHR) data, support this possibility. A computational repurposing screen identified bumetanide as a top candidate for reversing APOE4-related AD gene expression signatures, and treatment in APOE4 mouse models improved cognitive performance and disease-related molecular profiles45. Observational studies using EHRs have consistently reported lower AD risk among bumetanide-exposed individuals. For example, a Stanford-based study reported an adjusted odds ratio (aOR) of 0.23 (95% CI: 0.15–0.36) for AD among bumetanide users46. Our own findings support this association (FLAT WHITE: aOR 0.44, 95% CI 0.22, 0.86, TriNetX: aOR 0.80, 95% CI 0.69, 0.92). While bumetanide is known to inhibit NKCC1, a chloride cotransporter expressed in the brain, recent pharmacologic evaluations suggest that its therapeutic effects in AD models may occur at concentrations below those required for NKCC1 inhibition, pointing to possible non-NKCC1 or peripheral mechanisms47. These findings underscore the need for further investigation into bumetanide’s potential as a repurposed therapeutic agent for AD.

Furosemide, a loop diuretic, showed a reduced risk of ADRD (FLAT-WHITE: OR 0.62, 95% CI 0.45, 0.86; TriNetX: OR 0.81, 95% CI 0.76, 0.87). Diuretics may influence cerebrospinal fluid dynamics and facilitate amyloid clearance—an emerging therapeutic avenue in AD44,54,55.

Similarly, potassium chloride was linked to reduced ADRD risk (FLAT-WHITE: OR 0.66, 95% CI 0.47, 0.93; TriNetX: OR 0.79, 95% CI 0.74, 0.85). Higher potassium intake has been associated with reduced tau phosphorylation, altered amyloid aggregation, and improved cognitive outcomes56.

Medications like gabapentin (FLAT-WHITE: OR 0.44, 95% CI 0.25, 0.78; TriNetX: OR 0.86, 95% CI 0.82, 0.92) and pregabalin (FLAT-WHITE: OR 0.44, 95% CI 0.24, 0.79; TriNetX: OR 0.79, 95% CI 0.71, 0.88) showed potential neuroprotective properties. These drugs reduce oxidative stress and neuroinflammation, both implicated in ADRD pathophysiology48. The connection between epinephrine and cognitive decline is complex. Some research indicates that epinephrine enhances cerebral blood flow and oxygenation, potentially supporting cognitive health49. Our analysis supports an inverse association with ADRD (FLAT-WHITE: OR 0.38, 95% CI 0.23, 0.61; TriNetX: OR 0.82, 95% CI 0.77, 0.88), suggesting further investigation is needed. Ondansetron, an antiemetic, was associated with lower ADRD risk (FLAT-WHITE: OR 0.60, 95% CI 0.43, 0.83; TriNetX: OR 0.73, 95% CI 0.68, 0.78). This may be due to its modulation of serotonin receptors, which are increasingly recognized in neurodegeneration pathways50.

Cefazolin, an antibiotic, was significantly associated with lower ADRD risk in one dataset (FLAT-WHITE: OR 0.47, 95% CI 0.32, 0.69; TriNetX: OR 0.88, 95% CI 0.83, 0.93). Some antibiotics exhibit anti-inflammatory, antioxidant, and anti-apoptotic effects independent of their antimicrobial role49. These properties suggest potential for repurposing in ADRD prevention though the relationship between antibiotic use and dementia remains complex and controversial51.

Proton pump inhibitors (PPIs) like pantoprazole showed a possible protective association with ADRD (FLAT WHITE: OR 0.64, 95% CI 0.47, 0.88; TriNetX: OR 0.80, 95% CI 0.75, 0.85), potentially through anti-inflammatory mechanisms52. Famotidine, an H2 receptor blocker, was also linked to a significantly lower risk (FLAT WHITE: OR 0.54, 95% CI 0.41, 0.71; TriNetX: OR 0.85, 95% CI 0.81, 0.90), possibly via histamine-related immune modulation53. Bisacodyl, a stimulant laxative, was associated with reduced ADRD risk as well (FLAT-WHITE: OR 0.75, 95% CI 0.60, 0.95; TriNetX: OR 0.86, 95% CI 0.81, 0.91), potentially through effects on gut microbiota and systemic inflammation54. Gastrointestinal health plays a vital role in ADRD development, as chronic constipation and gut dysbiosis may contribute to systemic inflammation and cognitive decline55. While the relationship between sennosides and risk of ADRD is complex56, this study found an inverse association, suggesting a potential protective effect.

Metronidazole (FLAT WHITE: aOR 0.53, 95% CI 0.36, 0.80, TriNetX: aOR 0.90, 95% CI 0.84, 0.96) was associated with a reduced risk of ADRD. Through modulation of gut microbiota, metronidazole has been shown to reduce amyloid-β plaque burden and systemic inflammation in a mouse model of Alzheimer’s disease, suggesting a potential link between microbial balance and neurodegeneration57.

Fluconazole is a triazole compound with good blood-brain barrier penetration. It showed an inverse association with AD/ADRD risk (FLAT WHITE: aOR 0.39, 95% CI 0.23, 0.67, TriNetX: aOR 0.89, 95% CI 0.82, 0.97). It has demonstrated neuroprotective properties in preclinical studies, including the promotion of neurogenesis and interaction with insulin-like growth factor 1 (IGF-1) receptors58. Despite these findings, epidemiological evidence on the association between fluconazole use and the risk of Alzheimer’s disease remains scarce. Further research is needed to elucidate its potential role in cognitive health and dementia risk.

To further assess the sensitivity of our findings to unmeasured confounding, we have additionally calculated E-values and included these results in the Supplementary Appendix F. E-values quantify the minimum strength of association that an unmeasured confounder would need to have with both the treatment and outcome, conditional on measured covariates, to explain away the observed association. As shown in Supplementary Table S1, the E-values for the observed associations ranged from 1.32 to 5.33, while the E-values for the confidence limits ranged from 1.11 to 3.04. These results provide a quantitative assessment of sensitivity to unmeasured confounding, with larger E-values indicating that stronger unmeasured confounding would be required to explain away the observed associations. However, several findings had E-values for the confidence limit relatively close to 1, indicating greater sensitivity to modest unmeasured confounding and warranting cautious interpretation.

PDA-OTA: an interactive hub for causal ML collaboration and results dissemination

To conduct this collaborative study with our data partners, we implemented the analyses through the PDA-OTA (Privacy-preserving Distributed Algorithm Over the Air) platform (https://pda-ota.pdamethods.org) (Fig. 4). PDA-OTA serves as the interactive hub for executing and disseminating results from federated causal analyses. Built upon the PDA framework, it provides a secure, web-based infrastructure that enables privacy-preserving collaboration across institutions. By restricting communication to aggregated statistics rather than patient-level data, PDA-OTA ensures compliance with data governance requirements while supporting large-scale, distributed research. Its user-friendly design allows investigators to invite collaborators, monitor project progress, upload site-specific results, and generate automated summaries, thereby streamlining the workflow of multi-site studies.

Fig. 4: PDA-OTA with Interactive Hub for Results Dissemination.

The PDA-OTA platform provides an integrated web-based environment for privacy-preserving collaboration in multi-site studies. The interface enables data partners to upload aggregated results, track study progress, and visualize outputs such as effect estimates and covariate balance plots. By embedding federated algorithms like FLAT-WHITE, PDA-OTA functions as a causal machine learning working and dissemination hub, streamlining large-scale distributed analyses and supporting transparent, collaborative real-world evidence generation.

More importantly, PDA-OTA transforms methodological advances such as FLAT-WHITE into a practical causal ML pipeline for real-world evidence generation. By embedding statistical algorithms within an interactive, cloud-based environment, the platform automates results dissemination via visualizations for effect estimates and covariate balance, which ensuring the transparent communication among partners. In this way, PDA-OTA functions not only as a software tool but as the collaborative engine of causal machine learning, bridging methodological innovation and clinical application. This integration accelerates therapeutic discovery, enhances reproducibility, and supports secure, scalable, and trustworthy data-driven healthcare research.

Discussion

In this study, we proposed and evaluated FLAT-WHITE, which is a Federated Lossless Algorithm for Targeted Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication. FLAT-WHITE is specifically designed to enable scalable, secure, and communication-efficient target trial emulation using observational clinical data across distributed clinical sites. By achieving identical results to a pooled individual-level analysis through a one-shot federated framework, FLAT-WHITE addresses the practical and regulatory challenges of multi-site observational research. Importantly, FLAT-WHITE enables high-throughput causal inference with real-world observational data, automating parallel trial emulations across decentralized datasets to evaluate many candidate interventions efficiently. By combining intelligent weighting strategies for confounding control with integration into the PDA-OTA platform, it functions as a causal ML pipeline, transforming target trial emulation statistical modeling and learning process into a scalable system for population-level evidence generation. Positioned within the emerging field of causal ML, FLAT-WHITE complements predictive AI by enabling secure, reproducible, and large-scale causal discovery that accelerates therapeutic development and supports data-driven clinical decision making.

FLAT-WHITE offers several key advantages. First, it is a secure and trusted algorithm by requiring only the exchange of summary-level statistics and not patient-level data, thus complying with data protection regulations such as HIPAA and GDPR. This feature makes the proposed algorithm a practical, readily, and cost-saving deployable end-to-end solution for real-world, multi-institutional healthcare research. Second, it is communication-efficient, requiring just a single round of communication between sites to exchange summary-level statistics. It is important to note that each site will undergo a data quality assessment and cohort diagnostics step before generating and transferring the summary-level statistics for model fitting. This one-shot design dramatically reduces logistical overhead compared to traditional iterative federated methods, which often require multiple rounds of back-and-forth updates on the model parameters. By eliminating repeated synchronization steps and minimizing data traffic, FLAT-WHITE significantly lowers the burden on computational resources, network bandwidth, and human-in-the-loop coordination. In practice, this translates to fewer institutional delays, reduced dependency on IT personnel, and easier deployment in real-world settings where cross-site coordination can be a major bottleneck. Third, it is robust to extreme propensity score estimates due to the integration of overlap weighting, which improves covariate balance and strengthens causal inference, particularly in settings with unbalanced patient populations and rare diseases. Fourth, FLAT-WHITE incorporates negative control outcomes (NCOs) to detect and calibrate residual confounding bias caused by unmeasured confounders, enhancing the reliability of effect estimates. More importantly, although this study focused on a binary outcome use case, the proposed FLAT-WHITE framework is compatible with a broad class of generalized linear models (GLMs), making it adaptable to a wide range of outcome types commonly encountered in clinical research.

We demonstrated the scalability and real-world applicability of FLAT-WHITE through a large-scale, high-throughput target trial emulation for drug repurposing in Alzheimer’s disease and related dementias (ADRD). Utilizing a decentralized research network spanning four major health systems with about 123,971 patients, we emulated over 181,000 independent trials to evaluate the neuroprotective potential of frequently prescribed therapeutic classes/agents. We identified 21 drugs that showed a statistically significant effect on reducing ADRD onset among patients with MCI. These findings were further validated using the TriNetX database with more than 330,000 patients, providing external confirmation for several clinically actionable drug candidates.

Beyond its methodological contributions, FLAT-WHITE has significant promise for broader public health applications. FLAT-WHITE moves target trial emulation from a purely statistical exercise into a scalable and intelligent system for accelerating therapeutic discovery and clinical decision support. Designed as a causal ML pipeline, it systematically scans multiple interventions across distributed networks, delivering results equivalent to centralized analyses while protecting patients’ confidentiality, thus positioning itself as a foundation for real-world causal ML at population scale. Its ability to support real-time and secure analyses across multiple sites makes it well-suited for phase IV studies of drugs and biological agents, including post-marketing vaccine surveillance. In such settings, rapid yet robust assessment of treatment safety and effectiveness is critical, especially when deployed at scale across heterogeneous populations. The communication-efficient, lossless nature of FLAT-WHITE allows stakeholders to perform continuous and near real-time monitoring of treatment outcomes without compromising data security or requiring extensive technical infrastructure. As a result, FLAT-WHITE offers a practical and generalizable solution for evidence generation in dynamic and distributed healthcare environments.

Despite its strengths, FLAT-WHITE has limitations that present valuable directions for future development. First, the current implementation assumes a common outcome model structure across all participating sites, which may reduce flexibility in highly heterogeneous settings where clinical practices or patient populations vary significantly. This limitation could be addressed by extending the framework to support site-specific outcome models, enabling better accommodation of institutional variability and enhancing model fidelity. Second, FLAT-WHITE is currently designed for binary treatment scenarios and would benefit from extensions that allow for time-varying treatments or multi-level exposures, which are common in longitudinal and comparative effectiveness research. Third, while overlap weighting improves covariate balance, its effectiveness can be sensitive to misspecification of the propensity score (PS) model at each site. To address this, future work may incorporate doubly robust estimation techniques and flexible machine learning–based PS models, which offer greater robustness to model misspecification and can adapt more effectively to complex, real-world data environments. Nevertheless, the advantages of overlap weighting should be interpreted in light of its specific target estimand. While it can improve numerical stability and covariate balance, overlap weighting targets the average treatment effect in the overlap population (ATO), and its use should therefore be guided by whether this estimand aligns with the scientific question. Moreover, in the context of relatively very rare diseases, where the cohort sample size is limited within each site, particular attention should be paid to the accuracy of PS modeling, as estimation errors may be amplified under such constraints. Additionally, this study uses the odds ratio (OR) as the primary effect measure. However, the non-collapsibility of the OR remains a potential concern59,60. Future work could extend the current framework to incorporate relative risk (RR), which is generally preferred by clinicians due to its more intuitive interpretation and its property of collapsibility. Additionally, several potential sources of bias may remain despite the use of target trial emulation and propensity score weighting. Informative censoring may arise if loss to follow-up or treatment discontinuation is related to prognosis or other factors associated with the outcome, potentially leading to biased treatment effect estimates. Residual confounding and model misspecification may also contribute to bias, especially when relevant clinical factors are incompletely measured across participating sites.

In conclusion, FLAT-WHITE offers a scalable, lossless, communication-efficient, and secure framework for target trial emulation in distributed research networks. Applied to Alzheimer’s disease drug repurposing, the framework demonstrates strong potential to accelerate real-world evidence generation while protecting patients’ confidentiality and ensuring methodological rigor. As healthcare data becomes increasingly decentralized, FLAT-WHITE illustrates how causal ML pipelines can transform observational research into a scalable, collaborative, and reproducible enterprise. By embedding statistical innovation within an ML-enabled infrastructure, it establishes a foundation for trustworthy and intelligent evidence generation that advances therapeutic discovery and supports data-driven clinical decision-making.

Methods

Data source and study cohort

INSIGHT, the INSIGHT Clinical Research Network, comprises EHR data from five major health systems across the New York City metropolitan area, covering a total of 5,532,428 patients.

OneFlorida+ is a part of the national Patient-Centered Clinical Research Network (PCORnet) and includes EHR data for over 18 million patients from Florida and more than 6 million additional patients from other states, including Alabama, Georgia, Arkansas, California, and Minnesota. The database contains rich patient-level information, including demographics, vital signs, diagnoses, procedures, medications (prescribing and dispensing), and laboratory results.

UPHS, the University of Pennsylvania Health System, operated by Penn Medicine, is a major academic health system in the Greater Philadelphia area. The Penn Data Store houses EHR data from seven hospitals, encompassing records for over 6.5 million patients.

YNHHS, Yale New Haven Health System, is a leading academic medical center and the primary teaching hospital of Yale School of Medicine, delivering comprehensive patient care, research, and education. The YNHHS EHR system includes data for approximately 4 million patients.

TriNetX is a global research collaborative network that provides access to de-identified EHR data from a large consortium of healthcare organizations worldwide. The network includes data from academic medical centers, specialty physician practices, and integrated delivery networks, representing diverse patient populations across a wide range of clinical settings. Collected on April 21, 2025, the TriNetX Global Research Collaborative Network provided access to data from approximately 152.7 million patients across 106 healthcare organizations.

Specification and emulation of target trials

Target trial emulation is a systematic approach to design and analyze observational studies by explicitly specifying and emulating hypothetical RCTs (the “target trials”) to improve the quality of real-world effectiveness studies. To identify potential new treatments (previously non-AD drugs) that may reduce the risk of progression to AD/ADRD among patients with MCI, we design and emulated 42,700 single-site target trials and 9,050 multi-site target trials by comparing 181 candidate drugs to two active control groups for AD/ADRD outcome and 24 negative control outcomes. The cohort and drug selection process, along with validation methodology, are illustrated in Fig. 1. The study variables and associated ICD-9 and 10 codes were detailed in Appendix A of Supplementary Materials.

Eligibility criteria

We included patients with at least one MCI diagnosis between October 2007 and July 2022 from INSIGHT, January 2013 and October 2023 from OneFlorida + , January 2012 and December 2020 from UPHS, and January 2013 and December 2022 from YNHHS. MCI was identified using the following diagnosis codes: ICD-9 331.83 (mild cognitive impairment, so stated) and 294.9 (unspecified persistent mental disorders due to conditions classified elsewhere); and ICD-10 G31.84 (mild cognitive impairment, so stated) and F09 (unspecified mental disorder due to known physiological condition).

Treatment strategies

Each trial drug was compared against two active control groups: (1) Supplements with minimal direct impact on AD/ADRD progression, comprising 29 commonly used supplements such as Vitamin B12, folic acid, and other micronutrients, as primary analysis; and (2) Cardiovascular health medications, including 63 drugs such as atorvastatin, angiotensin-converting enzyme (ACE) inhibitors, and beta-blockers, which have established or emerging evidence suggesting protective effects against AD/ADRD or cognitive decline, as sensitivity analysis. In the target trial emulation, valid drug initiation for each drug was defined as having at least two consecutive prescriptions covering a minimum of 30 days from the initial prescription date in the database. The index date was established as the first prescription date of a valid initiation of a trial or control drug.

Treatment assignment

To emulate the randomized allocation of treatment groups in randomized trials, we adjusted for a comprehensive set of confounding variables using inverse probability of treatment weighting (IPTW). A large-scale propensity score model was constructed for each trial at each site using lasso-penalized logistic regression with 268 baseline covariates, including age, sex, race or ethnicity, 64 indicator variables for comorbidity history, top 200 prevalent prescribed drug ingredients for medication history, and the time lag between MCI diagnosis and the index date.

Outcomes

The primary outcome of interest was a diagnosis of AD/ADRD during follow-up. AD/ADRD was defined using the following diagnosis codes. For Alzheimer’s disease, we used ICD-9 code 331 and ICD-10 codes G30, G30.0, G30.1, G30.8, and G30.9. For vascular dementia, we used ICD-9 codes 290.4, 290.40, 290.41, 290.42, and 290.43, and ICD-10 codes F01, F01.5, F01.50, and F01.51. For Lewy body dementia, we used ICD-9 code 331.82 and ICD-10 code G31.83. For frontotemporal dementia, we used ICD-9 codes 331.1, 331.11, and 331.19, and ICD-10 codes G31.0, G31.01, and G31.09.

Follow-up

Follow-up began on the index date, defined as the initiation of the medication, and continued until the earliest occurrence of one of the following: the first AD/ADRD diagnosis, administrative censoring, or five years after baseline.

Causal estimand

The observational analogy of the intention-to-treat effect of being assigned to trial drug initiation.

Identification of trial drugs and active controls

For each site, we identified candidate trial drugs based on new drug initiation among patients diagnosed with MCI at 50 years of age or older. Drugs were included in the analysis if they had at least 100 eligible patients meeting the following criteria: no prior history of an AD/ADRD diagnosis before drug initiation, no use of the drug prior to the MCI diagnosis, and ongoing interactions with the healthcare system for at least one year prior to drug initiation.

Overview of overlap weights

Propensity score methods are extensively employed to mitigate confounding in research using observational clinical data. These scores estimate the likelihood of a subject receiving a particular treatment, given their confounding variables. A common strategy to address the bias from these variables is propensity score weighting, which adjusts the weights of participants in various treatment groups according to their treatment assignment probabilities, aiming to neutralize confounding influences.

Nonetheless, challenges arise when the propensity scores between treatment groups poorly overlap. One traditional solution is to use trimming, which excludes subjects with extreme propensity scores from the analysis. This approach, however, requires setting arbitrary cutoff points. It has been shown that the results would be sensitive to the cutoff points. An alternative method, overlap weighting, has been developed to overcome these issues. This technique weighs each participant by their likelihood of being assigned to the counterfactual group, with those having a propensity score around 0.5 receiving the greatest emphasis. Overlap weighting is particularly beneficial as it reduces the variance of the causal effect estimates, providing a more stable and efficient analysis compared to traditional inverse probability weighting methods.

Specifically, in the overlap weighting method, we define A as a binary treatment group indicator (i.e., A = 0, control group; A = 1, treated group) and x as a vector of confounding variables (e.g., age, sex, history of diseases), and Y as the observed outcome. The propensity score is defined as the probability of being assigned to the treatment group given baseline confounding variables \(e(x)=P(A=1{|x})\). The overlap weight for treated group is \(w(x)=1-e(x)\) and the overlap weight for control group is \(w(x)=e(x)\). The statistical properties of overall weights are detailed in Li et al. (2020)23. In this paper, we primarily focus on employing overlap weighting within our proposed federated learning algorithm to achieve multi-site target trial emulation. This method facilitates the analysis without the need for sharing patient-level data and effectively addresses potential issues with poorly overlapping propensity scores.

Estimation of treatment effect under multi-site setting without data sharing constraints

Suppose there are K sites in total within a network, the k-th site has numbers of patients \({n}_{k}\) and the total number of patients within such network is \(N={\sum }_{k}\,{n}_{k}\). For subject i at site k, we denote \({y}_{{ki}}\) as the observed outcome, \({x}_{{ki}}\), the confounding variables, and \({a}_{{ki}}\), the treatment indicator. For the k-th site, the propensity score for the i-th subject is \({e}_{{ki}}=P({A}_{{ki}}=1|{x}_{{ki}})\), which can be estimated by a logistic regression model. The propensity score weighs the treatment and control groups to create a pseudo-population, also known as the target population, where the two groups are expected to be balanced.

If there is no data sharing constraint such that the patient-level data from the K sites can be pooled together, after the propensity score calculation within each site, the overall weighted log-likelihood of the generalized linear model (GLM) to estimate the treatment effect based on the pseudo population is written as:

$$l\left(\beta \right)={\sum }_{k=1}^{K}{\sum }_{i=1}^{{n}_{k}}{w}_{{ki}}{y}_{{ki}}h\left({z}_{{ki}}^{T}\beta \right)-{w}_{{ki}}b\left({z}_{{ki}}^{T}\beta \right),$$

(1)

where \({w}_{{ik}}\) is denoted as the weight for the i-th individual from the k-th site, and \({z}_{{ki}}={\left(1,{A}_{{ki}}\right)}^{T}\) is the design vector including an intercept and treatment indicator, which have two possibilities, \({\left(1,0\right)}^{T}\) and \({\left({\mathrm{1,1}}\right)}^{T}\). The functions h(.)depends on the link function and exponential family distribution, and b(.) is the cumulant function characterizing the exponential family distribution. The treatment effect is estimated by solving the log-likelihood shown in Eq. (1). To mitigate the influence of extreme propensity scores, we applied overlap weighting, which yields the maximum likelihood estimate \(\hat{{{\rm{\beta }}}}\). This estimate corresponds to the causal effect in the overlap-weighted target population.

Proposed method: estimation of treatment effect under multi-site setting with data sharing constraints

Suppose patient-level data cannot be pooled; in this section, we introduce our proposed federated algorithm. While the fitting of the propensity score model can be done independently at each site, it should indeed be conducted separately. Since the propensity score model may vary due to differences in study populations, data collection procedures, and clinical guidelines across sites, directly fitting Eq. (1) is not feasible given these data constraints. Under such a scenario, we propose our federated learning algorithm – FLAT-WHITE, which stands for Federated Lossless Algorithm for Targeted Trial Emulation with Overlap Weighting for Healthcare Data Integration via Trusted and Efficient Communication. Given the constraints in patient-level data sharing in practice, our method only requires summary-level statistics rather than patient-level treatment assignment information to achieve the fitting of Eq. (1) and estimate the treatment effect.

As we defined, \({z}_{{ki}}\) has two possibilities \({\left(1,0\right)}^{T}\) and \({\left({\mathrm{1,1}}\right)}^{T}\), and we denote \({z}^{(j)}\) the j-th unique individual combination z, where j={1, 2}. When implementing the proposed algorithm, within a single round of communication, the summary statistics collected from each data contributor consist of two vectors:

  1. 1.

    A q-dimensional vector \({C}_{k}=\{{c}_{k1},\,{c}_{k2}\}\), where \({c}_{{kj}}={\sum }_{i=1}^{{n}_{k}}{w}_{{ki}}{y}_{{ki}}I\left({z}_{{ki}}={z}^{(j)}\right)\,\), counting the weighted sum of outcomes for each unique combination of z.

  2. 2.

    A q-dimensional vector \({S}_{k}=\{{s}_{k1},\,{s}_{k2}\}\), where \({s}_{{kj}}={\sum }_{i=1}^{{n}_{k}}{w}_{{ki}}I\left({z}_{{ki}}={z}^{(j)}\right)\,\), representing the sum of weight values \({w}_{{ki}}\,\) for individuals who has corresponding combination of z.

The treatment effect \({\hat{\tau }}_{0}\) for the outcome y can be estimated by maximizing the likelihood function \(l(\beta )\), using the summary-level statistics \(\{{C}_{k},{S}_{k}\}\) across K sites. The variance for the treatment effect \({\hat{\sigma }}_{0}^{2}\) can also be found accordingly.

Negative control outcomes (NCOs) calibration

In RWD used for biomedical research, data are collected during routine clinical care rather than under stringent randomized conditions. Although we employed overlap weighting to approximate randomization based on measured confounding variables, the estimated treatment effect may still be subject to systematic bias, including bias from unmeasured confounding. The negative control experiment leverages negative control outcomes (NCOs), which are known as a priori to be causally unrelated to the treatment of interest, to empirically calibrate and correct for such systematic bias61,62.

To empirically calibrate such bias, we implemented the negative control experiment within the FLAT-WHITE framework. Specifically, in the first step, we estimate the treatment effect for the l-th negative control outcome (\(l=1,\ldots ,L\)) by fitting the following model:

$$l\left({\beta }_{l}\right)={\sum }_{k=1}^{K}{\sum }_{i=1}^{{n}_{k}}{w}_{{ki}}{m}_{{lki}}h\left({z}_{{ki}}^{T}\beta \right)-{w}_{{ki}}b\left({z}_{{ki}}^{T}\beta \right),$$

(2)

where \({m}_{{lki}}\) is denoted as the NCO from the k-th site for the i-th individual.

Equation (2) is based on pooled data. However, to obtain equivalent results without pooling individual-level data, our proposed method FLAT-WHITE requires the transfer of additional summary-level statistics. Specifically, we define \({C}_{{lk}}=\left\{{c}_{{lk}1},\,\ldots ,\,{c}_{{lkq}}\right\},\) where \({c}_{{lkj}}={\sum }_{i=1}^{{n}_{k}}{w}_{{ki}}\,{m}_{{lki}}I\left({z}_{{ki}}={z}^{(j)}\right)\).

In the second step, we estimate the treatment effect \({\tau }_{l}\) for the l-th NCO using the summary-level statistics \({C}_{{lk}}\). These estimated treatment effects reflect the systematic error associated with each NCO, since the true effect should be zero in the absence of bias. We assume that the systematic errors for the NCOs and the primary outcome are exchangeable and follow a normal distribution with mean b and variance \({\sigma }^{2}\). The systematic error for the primary outcome, denoted \({b}_{0}\), can then be estimated using the set \(\{{\tau }_{l}\}\). Specifically, we fit the following model:

$${\hat{\tau }}_{l}\sim N\left({b}_{l},{\sigma }_{l}^{2}\right),\,({b}_{0},{b}_{l})\sim N\left(b,{\sigma }^{2}\right).$$

The calibrated estimate can thus be estimated by \(\widetilde{\tau }=\hat{\tau }-{\hat{b}}_{0}\) and 95% confidence interval,

$$\left[\hat{\tau }-{\hat{b}}_{0}-1.96\sqrt{{{\hat{\sigma }}_{0}^{2}+\hat{\sigma }}^{2}},\hat{\tau }-{\hat{b}}_{0}+1.96\sqrt{{{\hat{\sigma }}_{0}^{2}+\hat{\sigma }}^{2}}\right].$$

Ethics statement

This study did not involve experiments on humans or animals, nor the use of human participants’ biological material; therefore, institutional ethics approval and informed consent were not required.

Data availability

The datasets used in this study are not publicly available due to privacy concerns. The individual de-identified participant data will not be shared. Data from the TriNetX Research Network are available to authorized users through their respective access procedures. Instructions for obtaining access to TriNetX data are available at https://trinetx.com.

Code availability

The R code is available at https://github.com/Penncil/pda and through our R package:“pda” at https://github.com/Penncil/pda.

References

  1. Corrao, G. & Cantarutti, A. Building reliable evidence from real-world data: needs, methods, cautiousness and recommendations. Pulm. Pharm. Ther. 53, 61–67 (2018).

    Article  CAS  Google Scholar 

  2. Sherman, R. E. et al. Real-world evidence—what is it and what can it tell us. N. Engl. J. Med. 375, 2293–2297 (2016).

    Article  PubMed  Google Scholar 

  3. Liu, F. & Panagiotakos, D. Real-world data: a brief review of the methods, applications, challenges and opportunities. BMC Med. Res. Methodol. 22, 287 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  4. BEST Initiative n.d. https://bestinitiative.org/ (accessed January 21, 2025).

  5. Advancing Real-World Evidence Program | FDA https://www.fda.gov/drugs/development-resources/advancing-real-world-evidence-program (accessed April 22, 2025).

  6. Real-World Evidence | FDA n.d. https://www.fda.gov/science-research/science-and-research-special-topics/real-world-evidence (accessed April 22, 2025).

  7. Groenwold RHHTrial emulation and real-world evidence.JAMA Netw Open 4,e213845–e213845 (2021).

  8. Gomes, M. et al. Target trial emulation for transparent and robust estimation of treatment effects for health technology assessment using real-world data: opportunities and challenges. Pharmacoeconomics 40, 577–586 (2022).

    Article  PubMed  Google Scholar 

  9. Desai R. J., et al. Process guide for inferential studies using healthcare data from routine clinical practice to evaluate causal effects of drugs (PRINCIPLED): considerations from the FDA Sentinel Innovation Center. BMJ 2024;384. https://doi.org/10.1136/BMJ-2023-076460.

  10. FDA’s Sentinel Initiative | FDA n.d. https://www.fda.gov/safety/fdas-sentinel-initiative (accessed January 21, 2025).

  11. Platt, R. et al. The FDA Sentinel Initiative—an evolving national resource. N. Engl. J. Med 379, 2091–2093 (2018).

    Article  PubMed  Google Scholar 

  12. Gruschka, N., Mavroeidis, V., Vishi, K. & Jensen, M. Privacy Issues and Data Protection in Big Data: A Case Study Analysis under GDPR. Proc. - 2018 IEEE Int. Conf. Big Data, Big Data 2018, 5027–5033, https://doi.org/10.1109/BIGDATA.2018.8622621 (2018).

    Article  Google Scholar 

  13. Voigt P., von dem Bussche A. The EU General Data Protection Regulation (GDPR) 2024. https://doi.org/10.1007/978-3-031-62328-8.

  14. The HIPAA Privacy Rule and Adolescents: Legal Questions and Clinical Challenges on JSTOR. https://www.jstor.org/stable/3181198 (accessed April 23, 2025).

  15. General Data Protection Regulation (GDPR) – Legal Text n.d. https://gdpr-info.eu/ (accessed April 23, 2025).

  16. Health Insurance Portability and Accountability Act of 1996 (HIPAA) | Public Health Law | CDC n.d. https://www.cdc.gov/phlp/php/resources/health-insurance-portability-and-accountability-act-of-1996-hipaa.html# (accessed April 23, 2025).

  17. Wu, Y., Jiang, X., Kim, J. & Ohno-Machado, L. Grid Binary LOgistic REgression (GLORE): building shared models without sharing data. J. Am. Med Inf. Assoc. 19, 758–764, https://doi.org/10.1136/amiajnl-2012-000862 (2012).

    Article  Google Scholar 

  18. Lu, C.-L. et al. WebDISCO: a web service for distributed cox model learning without patient-level data sharing. J. Am. Med Inf. Assoc. 22, 1212–1219, https://doi.org/10.1093/jamia/ocv083 (2015).

    Article  Google Scholar 

  19. Duan R., Boland M. R., Moore J. H., Chen Y. ODAL: A one-shot distributed algorithm to perform logistic regressions on electronic health records data from multiple clinical sites. Biocomputing 2019, World Scientific; 2018, p. 30–41. https://doi.org/10.1142/9789813279827_0004.

  20. Duan, R. et al. Learning from electronic health records across multiple sites: A communication-efficient and privacy-preserving distributed algorithm. J. Am. Med. Inform. Assoc. 27, 376–385 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  21. Duan, R. et al. Learning from local to global-an efficient distributed algorithm for modeling time-to-event data. J. Am. Med. Inform. Assoc. 27, 1028–1036 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  22. Han, L., Hou, J., Cho, K., Duan, R. & Cai, T. Federated Adaptive Causal Estimation (FACE) of Target Treatment Effects. J. Am. Stat. Assoc. https://doi.org/10.1080/01621459.2025.2453249 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  23. Li, F. & Thomas, L. E. Addressing Extreme Propensity Scores via the Overlap Weights. Am. J. Epidemiol. 188, 250–257, https://doi.org/10.1093/AJE/KWY201 (2019).

    Article  PubMed  Google Scholar 

  24. Martin, B. K. et al. Cognitive Function Over Time in the Alzheimer’s Disease Anti-inflammatory Prevention Trial (ADAPT): Results of a Randomized, Controlled Trial of Naproxen and Celecoxib. Arch. Neurol. 65, 896, https://doi.org/10.1001/ARCHNEUR.2008.65.7.NCT70006 (2008).

    Article  PubMed  PubMed Central  Google Scholar 

  25. Mhillaj, E. et al. Celecoxib Prevents Cognitive Impairment and Neuroinflammation in Soluble Amyloid β-treated Rats. Neuroscience 372, 58–73, https://doi.org/10.1016/J.NEUROSCIENCE.2017.12.046 (2018).

    Article  CAS  PubMed  Google Scholar 

  26. Akiyama, H. et al. Inflammation and Alzheimer’s disease. Neurobiol. Aging 21, 383–421, https://doi.org/10.1016/S0197-4580(00)00124-X (2000).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  27. Wang, J. et al. Anti-inflammatory drugs and risk of Alzheimer’s Disease: An updated systematic review and meta-analysis. J. Alzheimer’s. Dis. 44, 385–396, https://doi.org/10.3233/JAD-141506 (2015).

    Article  CAS  Google Scholar 

  28. Xu J., et al. Single-microglia transcriptomic transition network-based prediction and real-world patient data validation identifies ketorolac as a repurposable drug for Alzheimer’s disease. Alzheimer’s and Dementia 2025;21. https://doi.org/10.1002/alz.14373.

  29. Heneka, M. T. et al. Neuroinflammation in Alzheimer’s disease. Lancet Neurol. 14, 388–405, https://doi.org/10.1016/S1474-4422(15)70016-5 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Rubio-Perez, J. M. & Morillas-Ruiz, J. M. A Review: Inflammatory Process in Alzheimer′s Disease, Role of Cytokines. Sci. World J. 2012, 756357. https://doi.org/10.1100/2012/756357 (2012).

    Article  CAS  Google Scholar 

  31. Garcia-Gomara, M., Juan-Palencia, A., Alfaro, M., Cuadrado-Tejedor, M. & Garcia-Osta, A. Neuroprotective Effects of Dexamethasone in a Neuromelanin-Driven Parkinson’s Disease Model. J. Neuroimmune Pharmacol. 20, 1–16, https://doi.org/10.1007/S11481-024-10164-4/FIGURES/6 (2025).

    Article  Google Scholar 

  32. Hui, Z. et al. The combination of acyclovir and dexamethasone protects against Alzheimer’s disease-related cognitive impairments in mice. Psychopharmacol. (Berl.) 237, 1851–1860, https://doi.org/10.1007/S00213-020-05503-1/FIGURES/9 (2020).

    Article  CAS  Google Scholar 

  33. Nerius, M., Haenisch, B., Gomm, W., Doblhammer, G. & Schneider, A. Glucocorticoid Therapy is Associated with a Lower Risk of Dementia. J. Alzheimer’s. Dis. 73, 175–183, https://doi.org/10.3233/JAD-190444 (2020).

    Article  Google Scholar 

  34. Tokuda, T. et al. Prednisolone (30-60 mg/day) for diseases other than AD decreases amyloid β-peptides in CSF. Neurology 58, 1415–1418, https://doi.org/10.1212/WNL.58.9.1415 (2002).

    Article  CAS  PubMed  Google Scholar 

  35. Alisky, J. M. Intrathecal corticosteroids might slow Alzheimer’s disease progression. Neuropsychiatr. Dis. Treat. 4, 831–833, https://doi.org/10.2147/NDT.S3685 (2008).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  36. Tripathy D., Grammas P. Acetaminophen inhibits neuronal inflammation and protects neurons from oxidative stress. J Neuroinflammation 2009;6. https://doi.org/10.1186/1742-2094-6-10.

  37. Shin A., Koo B. K., Lee J. Y., Kang E. H. Risk of dementia after initiation of sodium-glucose cotransporter-2 inhibitors versus dipeptidyl peptidase-4 inhibitors in adults aged 40-69 years with type 2 diabetes: population based cohort study. BMJ 2024:e079475. https://doi.org/10.1136/bmj-2024-079475.

  38. Mone, P. et al. Empagliflozin Improves Cognitive Impairment in Frail Older Adults With Type 2 Diabetes and Heart Failure With Preserved Ejection Fraction. Diab. Care 45, 1247–1251, https://doi.org/10.2337/dc21-2434 (2022).

    Article  CAS  Google Scholar 

  39. Tang, H. et al. Newer <scp>glucose-lowering</scp> drugs and risk of dementia: A systematic review and <scp>meta-analysis</scp> of observational studies. J. Am. Geriatr. Soc. 71, 2096–2106, https://doi.org/10.1111/jgs.18306 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  40. Tang, H. et al. GLP-1RA and SGLT2i Medications for Type 2 Diabetes and Alzheimer Disease and Related Dementias. JAMA Neurol. 82, 439, https://doi.org/10.1001/jamaneurol.2025.0353 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  41. He, D. & Aleksic, S. Is it time to repurpose geroprotective diabetes medications for prevention of dementia? J. Am. Geriatr. Soc. 71, 2041–2045, https://doi.org/10.1111/jgs.18405 (2023).

    Article  PubMed  Google Scholar 

  42. Pastena, P., Campagnoli, G., Rahmani, A. R. & Kalogeropoulos, A. P. Mineralocorticoid Receptor Antagonists and Cognitive Outcomes in Cardiovascular Disease and Beyond: A Systematic Review. J. Pers. Med 15, 57, https://doi.org/10.3390/jpm15020057 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  43. Pastena, P., Skopicki, H., Halterman, M. & Kalogeropoulos, A. Spironolactone Use And Neurocognitive Outcomes Among Patients With Heart Failure And Reduced Ejection Fraction. J. Card. Fail 30, 258–259, https://doi.org/10.1016/j.cardfail.2023.10.337 (2024).

    Article  Google Scholar 

  44. Lundin S. K., et al. Association between risk of Alzheimer’s disease and related dementias and angiotensin receptor Ⅱ blockers treatment for individuals with hypertension in high-volume claims data 2024. https://doi.org/10.1016/j.ebiom.2024.105378.

  45. Taubes, A. et al. Experimental and real-world evidence supporting the computational repurposing of bumetanide for APOE4-related Alzheimer’s disease. Nat. Aging 1, 932–947, https://doi.org/10.1038/s43587-021-00122-7 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  46. Graber-Naidich A., et al. Bumetanide Exposure Association with Alzheimer’s Disease Risk. Alzheimer’s & Dementia 2023;19. https://doi.org/10.1002/alz.079378.

  47. Boyarko B., et al. Evaluation of bumetanide as a potential therapeutic agent for Alzheimer’s disease. Front Pharmacol 2023;14. https://doi.org/10.3389/fphar.2023.1190402.

  48. Aşcı, S., Demirci, S., Aşcı, H., Doğuç, D. K. & Onaran, İ Neuroprotective Effects of Pregabalin on Cerebral Ischemia and Reperfusion. Balk. Med J. 33, 221, https://doi.org/10.5152/BALKANMEDJ.2015.15742 (2016).

    Article  Google Scholar 

  49. Chandolia, P. & Agrawal, M. Antibiotics: Treatment for Alzheimer’s diseases. Alzheimer’s. Dement. 19, e061395. https://doi.org/10.1002/ALZ.061395 (2023).

    Article  Google Scholar 

  50. Claeysen, S., Bockaert, J. & Giannoni, P. Serotonin: A New Hope in Alzheimer’s Disease? ACS Chem. Neurosci. 6, 940–943, https://doi.org/10.1021/ACSCHEMNEURO.5B00135 (2015).

    Article  CAS  PubMed  Google Scholar 

  51. Underwood, B. R. et al. Data-driven discovery of associations between prescribed drugs and dementia risk: A systematic review. Alzheimer’s. Dement.: Transl. Res. Clin. Interventions 11, e70037. https://doi.org/10.1002/TRC2.70037 (2025).

    Article  Google Scholar 

  52. Kedika, R. R., Souza, R. F. & Spechler, S. J. Potential anti-inflammatory effects of proton pump inhibitors: A review and discussion of the clinical implications. Dig. Dis. Sci. 54, 2312–2317, https://doi.org/10.1007/S10620-009-0951-9/TABLES/1 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  53. Anthony, J. C. et al. Reduced prevalence of AD in users of NSAIDS and H2 receptor antagonists: The Cache County study. Neurology 54, 2066–2071, https://doi.org/10.1212/WNL.54.11.2066 (2000).

    Article  CAS  PubMed  Google Scholar 

  54. Zhou, Z. et al. Histamine and receptors in neuroinflammation: Their roles on neurodegenerative diseases. Behavioural Brain Res. 465, 114964. https://doi.org/10.1016/J.BBR.2024.114964 (2024).

    Article  CAS  Google Scholar 

  55. Escobar, Y.-N. H., O’Piela, D., Wold, L. E. & Mackos, A. R. Influence of the Microbiota-Gut-Brain Axis on Cognition in Alzheimer’s Disease. J. Alzheimer’s. Dis. 87, 17–31, https://doi.org/10.3233/JAD-215290 (2022).

    Article  CAS  Google Scholar 

  56. Yang, Z. et al. Association between Regular Laxative Use and Incident Dementia in UK Biobank Participants. Neurology 100, E1702–E1711, https://doi.org/10.1212/WNL.0000000000207081 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  57. Wasén C., et al. Bacteroidota inhibit microglia clearance of amyloid-beta and promote plaque deposition in Alzheimer’s disease mouse models. Nat Commun 2024;15. https://doi.org/10.1038/s41467-024-47683-w.

  58. Toodle, V. et al. Fluconazole Is Neuroprotective via Interactions with the IGF-1 Receptor. Neurotherapeutics 19, 1313–1328, https://doi.org/10.1007/s13311-022-01265-0 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  59. Greenland, S. Noncollapsibility, confounding, and sparse-data bias. Part 1: The oddities of odds. J. Clin. Epidemiol. 138, 178–181 (2021).

    Article  PubMed  Google Scholar 

  60. Greenland, S. Noncollapsibility, confounding, and sparse-data bias. Part 2: What should researchers make of persistent controversies about the odds ratio? J. Clin. Epidemiol. 139, 264–268 (2021).

    Article  PubMed  Google Scholar 

  61. Schuemie, M. J., Hripcsak, G., Ryan, P. B., Madigan, D. & Suchard, M. A. Empirical confidence interval calibration for population-level effect estimation studies in observational healthcare data. Proc. Natl. Acad. Sci. USA 115, 2571–2577, https://doi.org/10.1073/PNAS.1708282114 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  62. Schuemie, M. J., Ryan, P. B., Dumouchel, W., Suchard, M. A. & Madigan, D. Interpreting observational studies. Why Empir. Calibration is. Needed Correct. P-Values Stat. Med 33, 209–218, https://doi.org/10.1002/SIM.5925 (2014).

    Article  Google Scholar 

Download references

Acknowledgements

This work was supported in part by National Institutes of Health (U01TR003709, U24MH136069, U24AG098157, RF1AG077820, R01AG073435, R01LM013519, 1R01LM014344, R01DK128237, R21AI167418, R21EY034179).

Author information

Authors and Affiliations

  1. Department of Biostatistics, Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, USA

    Jiayi Tong

  2. The Center for Health AI and Synthesis of Evidence (CHASE), University of Pennsylvania, Philadelphia, PA, USA

    Lu Li, Dazheng Zhang, Yiwen Lu, Ting Zhou, Bingyu Zhang, Yuqing Lei, Huilin Tang & Yong Chen

  3. The Graduate Group in Applied Mathematics and Computational Science, School of Arts and Sciences, University of Pennsylvania, Philadelphia, PA, USA

    Lu Li, Yiwen Lu, Bingyu Zhang & Yong Chen

  4. Department of Biostatistics, Epidemiology, and Informatics, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA

    Dazheng Zhang, Ting Zhou, Yuqing Lei, Huilin Tang & Yong Chen

  5. Department of Neurology, University of Pennsylvania Perelman School of Medicine, Philadelphia, PA, USA

    Kyra S. O’Brien & David A. Wolk

  6. Leonard Davis Institute for Health Economics, University of Pennsylvania, Pennsylvania, PA, USA

    Kyra S. O’Brien

  7. Department of Biostatistics and Health Data Science, University of Pittsburgh, Pittsburgh, PA, USA

    Qiong Wu

  8. Merck & Co Inc, Rahway, NJ, USA

    Naimin Jing

  9. Weill Cornell Medicine, Cornell University, New York, NY, USA

    Haoyang Li, Zhenxing Xu & Fei Wang

  10. Department of Biostatistics and Health Data Science, Indiana University, Bloomington, IN, USA

    Yu Huang & Jiang Bian

  11. Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida, Gainesville, FL, USA

    Yu Hu

  12. Department of Biomedical Informatics and Data Science, School of Medicine, Yale University, New Haven, CT, USA

    Yujia Zhou, Fongci Lin & Hua Xu

  13. Third Affiliated Hospital of Sun Yat-sen University, Guangzhou, China

    Ying Jiang

  14. Statistical Research and Data Science Center, Pfizer Inc, New York, NY, USA

    Haitao Chu & Yong Chen

  15. College of Population Health, Thomas Jefferson University, Philadelphia, PA, USA

    Yong Chen

  16. Real-World Evidence and Epidemiology, Chief Medical Office, Pfizer, Collegeville, PA, USA

    Yong Chen

  17. Penn Institute for Biomedical Informatics, University of Pennsylvania, Philadelphia, PA, USA

    Yong Chen

  18. Penn Medicine Center for Evidence-based Practice (CEP), Philadelphia, PA, USA

    Yong Chen

Authors

  1. Jiayi Tong
  2. Lu Li
  3. Dazheng Zhang
  4. Kyra S. O’Brien
  5. Yiwen Lu
  6. Qiong Wu
  7. Ting Zhou
  8. Bingyu Zhang
  9. Yuqing Lei
  10. Huilin Tang
  11. Naimin Jing
  12. Haoyang Li
  13. Zhenxing Xu
  14. Yu Huang
  15. Yu Hu
  16. Yujia Zhou
  17. Fongci Lin
  18. Ying Jiang
  19. Haitao Chu
  20. Yong Chen
  21. Fei Wang
  22. Jiang Bian
  23. Hua Xu
  24. David A. Wolk
  25. Yong Chen

Contributions

J.T. and Y.C. designed methods and J.T., L.L., D.Z., and Y.C. designed the experiments; H.L., Z.X., Y.Huang, Y.Hu., Y.Z., F.L., F.W., J.B., and H.X. had the datasets for data analysis, and generated the aggregated data; JT and YC served as the coordinating center; all authors mentioned above, K.O., Y.W.L., Q.W., T.Z., B.Z., Y.Q.L., H.T., N.J., Y.J., H.C., Y.C., F.W., J.B., H.X., D.W., and Y.C. interpreted the results and provided instructive comments; J.T., L.L., D.Z., and Y.C. drafted the main manuscript. All coauthors provided critical edits of the early draft and approved the final version of the manuscript.

Corresponding author

Correspondence to Yong Chen.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Tong, J., Li, L., Zhang, D. et al. FLAT-WHITE: from target trial emulation to federated high-throughput causal machine learning using decentralized real-world data. npj Health Syst. 3, 102 (2026). https://doi.org/10.1038/s44401-026-00160-6

Download citation

  • Received: 19 January 2026

  • Accepted: 16 September 2026

  • Published: 27 October 2026

  • Version of record: 27 October 2026

  • DOI: https://doi.org/10.1038/s44401-026-00160-6

来源:Nature 健康护理主题 · nature.com