Federation, not centralization: a new paradigm for electronic health record–based critical care research

Introduction

Critical care research has evolved with shifts in study design, data sources, and clinical focus.¹ Early single-center investigations generated insights about organ-level pathophysiology. Subsequently, large registries and multicenter randomized controlled trials helped define syndromes, evaluate interventions, and standardize practices. The current era’s focus on precision medicine and real-world applicability has been driven by advances in trial design and methodology, analytics, and high-volume electronic health record (EHR) data.^2,3 However, these advances have also highlighted the need for efficient and accessible approaches to multicenter data collaboration. Available models differ principally in where patient-level data resides and how it is governed: Project-specific data sharing transfers records to study-initiated repositories; centralized enclaves aggregate data within durable, multiuse infrastructure; and federated networks retain data locally while exchanging only aggregate results (Figure 1). Each presents distinct tradeoffs (Table 1). Here, we present a novel federated approach to critical care research and propose that federation within a community of physician-informaticists offers the greatest promise to accelerate modern collaborative investigation.

Figure 1.

Two models for multicenter observational intensive care unit research using EHR data. (A) Traditionally, data have been centrally stored, prepared, and analyzed. (B) In federated approaches, such as CLIF, each site maintains its own data repository in adherence to a harmonized data model suitable for investigator-initiated projects that use publicly available shared analysis code. Abbreviations: CLIF, Common Longitudinal ICU data Format; DUA, data use agreement; EHR, electronic health record; ETL, extract-transform-load (common parlance for the earliest stages of data preparation and cleaning).

Table 1.

Approaches to multicenter observational research.

	Project-specific data sharing	Centralized enclave	Federation
Examples	Most NIH-funded multisite studies and investigator-initiated collaborations	National Clinical Cohort Collaborative (NC3, https://n3c.ncats.nih.gov/clinical-cohort) NIH All of Us Research Program (https://allofus.nih.gov/) National Patient-Centered Clinical Research Network (PCORNet, https://pcor-net.org/)	CLIF Health Care Systems Research Network (https://www.hcsrn.org/)
Data use agreements	Requires multi-institutional DUAs for raw data transfer. Timeline depends on institutional legal review capacity.	Requires DUAs but often simpler since data are de-identified and in standardized format. May require tiered access agreements for different data elements.	Individual institutions control access decisions. No cross-institutional DUAs for raw data. Sites agree to data model and query protocols rather than data transfer terms.
Infrastructure and cost	Moderate upfront investment in central storage and security. Low technical coordination burden—one team handles harmonization. Most cost is in initial data transfer and cleaning. Inefficient if data practices differ substantially across sites.	High upfront investment in secure infrastructure, governance framework, and common data model development. Requires sustained funding for maintenance, security updates, and access brokering. Economies of scale emerge with multiple studies. Expertise in common data model required for sites to prepare contributions	Distributed infrastructure costs—each site maintains extract-transform-load pipelines and harmonization to common data model. Shared development of harmonization tools amortizes costs. Lower central infrastructure needs. Requires coordination mechanisms and technical support network.
Governance and security	Single governing body controls access and usage. Clear accountability but requires trust in central authority. Must balance competing interests of contributing sites. Raw data transferred and stored centrally creates single point of vulnerability. Breach exposes all contributed data. Requires robust security infrastructure and monitoring.	Formal governance structure with defined roles for data stewards, access committees, and oversight boards. Established precedents and frameworks. Can be bureaucratic. De-identified data in secure environment with access controls and audit logs; lower risk than raw data centralization but still concentrated. Established security protocols.	Distributed governance; each site retains control over its own data. Consortium agreements define participation terms. Requires consensus-building for protocol changes. Potentially more democratic, with decision-making and execution efficiency dependent upon consortium size, member engagement. Patient-level data never leaves local institutions. Distributed risk—breach at one site doesn’t compromise others. Each site responsible for its own security posture, creating variable protection levels.
Patient privacy	Cross-institutional transfer of identifiable data requires rigorous consent and de-identification. Re-identification risk if multiple datasets linked.	Strong de-identification needed before entry to enclave. Access controls limit who can query data. Aggregation across large populations provides privacy protection. Small cell suppression policies required.	Individual patient data remains local. Only aggregate statistics or models leave institutions. Lowest re-identification risk. Local IRBs maintain oversight. Requires technical safeguards against inference attacks on aggregated results.
Scalability	Scales well once centralized infrastructure is built—each new site requires new DUAs, data transfer, and integration into harmonized dataset. Data heterogeneity compounds with size. Error-prone data cleaning pipelines difficult to address without knowledge of local practices at source site.	Scales well for users once built, as many investigators can access the same data. Scalability for contributors is more challenging. Adding new data sources requires substantial work to map to common data models. Best for established, stable networks. Challenging to scale software/computational requirements.	Highly scalable for adding sites—new institutions adopt existing data model and tools. Network effects increase value. Analytical scalability more limited—some methods don’t federate well, and data quality is not inherently compared/controlled across sites.
Analytics	Full access to patient-level data enables any statistical or machine learning method. Exploratory analyses straightforward. Can iteratively refine models and check assumptions. Maximum analytical flexibility.	Complete analytical flexibility with pooled data. Supports complex machine learning, causal inference, and exploratory work. Can calculate any statistic. However, access may be restricted or mediated, limiting iteration speed.	Limited to analyses that can be conducted on summary statistics or that federate across sites. Means, counts, regression coefficients work well. Medians, complex machine learning models, and highly iterative analyses challenging. Requires prespecification. Growing technical solutions (federated learning, secure multiparty computation). Rare conditions may not be appropriate for federation if small n categorical data could potentially identify individual patients.
Collaboration	Collaboration can be limited to initial data contribution; once data transferred, sites may have reduced engagement. External investigators can access but contributing sites may not directly benefit from ongoing work.	Enables broad investigator access to shared resource. However, sites that contributed data may not maintain active role, as there is no incentive to maintain local data in the common data model format unless contributions are recurring. Democratic access possible but doesn’t inherently build community.	Investment in local data maintenance directly enables participation. Creates sustained engagement incentives. Shared tool development builds community. Sites remain active partners rather than data sources. Regular interaction through consortium activities.
Scientific transparency and reproducibility	Reproducibility depends on documentation of central harmonization decisions. Often opaque to external investigators. Difficult to trace data lineage. Reuse for future questions possible but depends on central team’s choices during initial harmonization.	High reproducibility within enclave environment—same data, same code produces same results. However, harmonization decisions made during common data model creation may not be transparent. Versioning of datasets supports reproducibility over time.	Harmonization code shared across sites provides transparency into data preparation. Reproducible pipelines from raw data to analysis. Local practices visible to consortium. However, site-to-site variation in implementation can reduce reproducibility. Regular validation checks required.
Enables deployment of tools developed with the data	Tools or models developed centrally cannot easily be deployed back to source institutions without separate implementation work. No direct path from research to practice at contributing sites.	Models developed in enclave trained on de-identified data from multiple sites. Deployment to any single site requires re-implementation and validation on local data. Enclave environment differs from clinical production systems.	Research conducted through network immediately compatible with local systems since data models are already aligned. Prediction models, clinical decision support tools, and quality metrics can deploy directly to all participating sites. Shortest path from discovery to implementation.

Abbreviations: CLIF, Common Longitudinal ICU data Format; DUA, data use agreement; IRB, institutional review board; NIH, National Institutes of Health

Challenges in multicenter collaboration

Multicenter intensive care unit (ICU) research has traditionally relied on transferring patient-level data from individual institutions to project-specific repositories housed at one site. This model is conceptually straightforward with clear lines of authority and responsibility, and allows maximal analytical flexibility for the coordinating center. However, this model can also hinder the collaboration it was meant to support. Administrative barriers frequently delay data sharing; new studies often require negotiating separate data use agreements, each of which can take months. Even when the same institutions collaborate on multiple investigations, overlapping data elements might be re-extracted and re-cleaned based on study-specific data dictionaries and variable definitions.⁴ This duplicated effort slows progress, especially for collaborations with limited data management support.

Centralized enclaves address some of these inefficiencies through reusable data infrastructure where a single harmonization effort supports multiple investigations. Established enclaves afford broad analytical flexibility on de-identified patient-level data and can facilitate durable access for external investigators without site-by-site agreements for new projects. However, enclaves face their own barriers: substantial upfront and maintenance infrastructural investment, strict anonymization requirements, concentrated security vulnerabilities, and complex multi-institutional data contribution agreements.

Both centralized approaches also share a critical limitation: Preparing data for off-site migration often degrades the contextual richness that provides meaning—clinical subtleties, institutional protocols, and idiosyncratic local practices.⁵ These losses are especially problematic when research questions depend on understanding how providers interact with the system. Wholesale off-site data migration can thus yield simplified extracts that misrepresent complex clinical realities, risking flawed conclusions. Although the overall result is more data, there is paradoxically less confidence in its integrity.⁶ Moreover, these issues may disproportionately affect institutions and investigators with less research infrastructure and support.

Federated research enables secure, scalable, collaborative ICU research

Technological advancements allow reconsideration of the centralized data paradigm. In federated analyses, patient-level data are queried and compiled locally, with only summary statistics, models, or parameters exchanged between sites.⁷ This approach, well-established in genomics and finance, has gained favor in healthcare for its strong privacy protections.^7,8 For instance, the International Severe Acute Respiratory and Emerging Infection Consortium (ISARIC) aggregates results (rather than individual patient data) from standardized local analyses built on manually extracted case report forms.⁹ Federated analyses are also possible with EHR data. The 20-system Health Care Systems Research Network maintains locally controlled, EHR-based harmonized datasets focused on population health innovation.¹⁰ While these examples demonstrate the power of distributed data networks, their structure and content are not optimized for the temporal resolution, physiologic detail, or content heterogeneity required for critical care research. Here, we present an ICU-specific federated approach to collaboration (Figure 1).

The Common Longitudinal ICU data format (CLIF) enables rapid federated multisite research

Founded in 2023, CLIF is a network of critical care investigators with expertise in EHR-based research. CLIF’s mission is to establish a comprehensive ecosystem of essential acute care data elements to enhance quality, reproducibility, and efficiency in critical illness research. Currently, CLIF includes 12 leading North American academic health systems, representing a demographically and geographically diverse cohort of over 850 000 critically ill patients from 46 hospitals (https://clif-icu.com/).¹¹ Our guiding values have been (1) consistent, understandable, and harmonizable data elements and structures; (2) reproducible, transparent, and scalable research; (3) upholding data privacy and security via federated analysis; and (4) open-source development.¹² Early experience highlights that CLIF’s effectiveness relies on 3 integrated components: a common data model, contributors with expertise in both critical care medicine and data science, and federated governance.

Successful federated critical care research requires a focused common data format

Multicenter ICU research relies on a shared conceptual model of critical illness and syntactically harmonized ICU data elements.¹³ Common data models standardize naming, typing, and relationships of variables such that data elements are synonymous across institutions. Currently, the Observational Medical Outcomes Partnership (OMOP) provides the most prominent EHR-relevant common data model harmonizing large-scale clinical data.¹⁴ However, OMOP’s applicability to critical care research is limited by inadequate granularity and inconsistent representation of ICU-specific concepts such as ventilator settings and medication dosages.¹⁵ Similar challenges exist for the Health Care Systems Research Network data model. Open-access datasets (eg, the Medical Information Mart for Intensive Care [MIMIC]) provide detailed, high-resolution ICU data,¹⁶ but are not themselves data models, limiting their broader use for collaborative research. Recently, the Society for Critical Care Medicine introduced the Critical Care Data Dictionary (C2D2) to address the unmet need for an ICU-specific common data model.¹⁷ This important resource is well-suited for manual extraction into case report forms but is not designed for high-resolution EHR data integration across systems.

CLIF addresses this challenge by developing common data elements dynamically through research project requirements

CLIF’s clinician-curated ICU data model is built around standardized minimum Common ICU Data Elements (mCIDEs) that explicitly define variable meanings, naming conventions, and encoding methods to accurately reflect clinical concepts. Analogous to oncology’s approach to standardized data elements,^5,18 CLIF mCIDEs complement OMOP by clarifying ambiguous clinical definitions, detailing ICU-specific variables (eg, ventilator modes and settings), and adding temporal resolution crucial for ICU research. Harmonization through mCIDEs accelerates cohort creation, increases analytic rigor, and reduces technical and governance burdens.

Collaborations vary in how they identify and resolve data element issues. Models serving broad communities often employ centralized quality control and multistep consensus processes to manage change, ensuring wide applicability but prolonging refinement timelines. For instance, OMOP version 6.0 remains unsupported by key analytical tools 7 years postintroduction.¹⁹ Quality assurance processes based on theoretical or vocabulary-level considerations often recognize practical implementation challenges only during analysis—triggering iterative cycles of issue identification and consensus-based resolution after completed versioning. In contrast, CLIF’s syntactic alignment for mCIDEs occurs through iterative collaboration during active research. Data elements mature through repeated practical application, becoming finalized only after successful use across multiple projects. This dynamic project-driven (as opposed to vocabulary-driven) approach recognizes that real-world problems in data structures typically emerge during analysis.²⁰ This rapid-cycle maturation is not necessarily unique to the federated paradigm, but is enabled by CLIF’s size and retention of local data control.

Reuse of mCIDEs and modular code further accelerates analysis and allows parallel project execution. Shared code repositories eliminate redundancy, while contributions from local data scientists enhance code applicability and quality. CLIF continues to strengthen its infrastructure through reusable analytic templates and standardized onboarding processes, promoting reproducibility, transparency, and efficiency in multicenter ICU research.

Successful federated critical care data science requires contributors with clinical and informatics fluency

CLIF’s mCIDEs alone are insufficient for federated success; active contributors with dual clinical and data expertise are also essential. While CLIF teams are multidisciplinary, clinicians skilled in data science are suited for detecting physiologically implausible patterns or documentation errors during data cleaning that could otherwise go unnoticed (eg, ventilator metrics incompatible with documented mode). Cross-site discussions lead to shared heuristics for cleaning these data, strengthening consistency across studies. Embedding clinical reasoning into the data science workflow thus improves both efficiency and validity of data preparation and analyses.²¹

This work also builds critical expertise in how EHR data are generated, structured, and interpreted. Shared repositories with standard project templates enable rapid project initiation and scaling while reducing the local analytic burden when updates are needed. Cross-site comparisons have surfaced subtle inconsistencies in EHR data that would have gone unnoticed in isolation. For example, one site identified system-wide documentation errors through these collaborative exercises, highlighting CLIF’s potential as a tool for audit and feedback. A “buddy-code” system pairs each lead site with a partner site to test project-specific code, identifying bugs early and reducing downstream rework.

Successful federated critical care research requires effective governance

Beyond technical architecture, federation depends on governance that balances institutional autonomy with collaborative standards. CLIF sites collectively agree on data models and analysis protocols while retaining discretion over which projects to support and what data elements to implement locally—a feature inherent to federated architectures where participation is voluntary rather than mandated. CLIF’s “minimum viable model” approach allows sites to prioritize core concepts—such as ventilator parameters, vital signs, and laboratory results—while deferring less commonly used features. Accommodating differences in local capacity and priorities facilitates meaningful participation even for resource-constrained institutions, which would be difficult in centralized models where comprehensive data transfer is typically an all-or-nothing proposition. This distributed decision-making creates challenges and opportunities. Reaching consensus without centralized authority requires regular discussions among members for model refinement and project prioritization. However, this process ensures that standards reflect real-world implementation constraints across diverse institutions.

Federated structures create direct incentives for sustained collaboration. Each site’s investment in maintaining local data in a common format promotes ongoing participation and scientific exchange, whereas in centralized models, sites may have limited engagement after initial data transfer. Furthermore, federated approaches enable direct deployment of collaboratively developed tools and models at participating sites—a pathway that centralized or enclave-based models cannot easily provide. This creates a virtuous cycle where research conducted through the network can immediately inform clinical practice at member institutions, forming an active learning network rather than static data infrastructure.

Challenges and future opportunities for federated ICU research

Federated research poses several practical challenges. First, mapping local EHR data to any common data model involves upfront effort from clinicians and data engineers, and distributed harmonization means this effort occurs at each site rather than once centrally. However, investigators must invest in data extraction regardless of the collaboration model—either pulling data themselves or providing detailed guidance to local brokers. On-site harmonization adds incrementally to this work but is enhanced by local knowledge developed during extraction. Moreover, federated collaborations mitigate costs by sharing harmonization scripts, validation frameworks, and collective troubleshooting, allowing sites to inform model development rather than adapting to external standards and potentially increasing data accuracy through local context. Coordinating this mapping once per site enables participation in multiple projects without repeating foundational data preparation, whereas project-specific collaborations might require re-extraction for each study to conform to unique data dictionaries and variable definitions.

While the distributed database is a strength of CLIF’s approach, it requires longitudinal access to institutional data warehouses that could potentially exclude sites with limited infrastructure. We have therefore prepared the MIMIC dataset in CLIF format (https://github.com/Common-Longitudinal-ICU-data-Format/CLIF-MIMIC) so investigators can develop projects against standardized data without committing resources to building local CLIF databases. Because initiating a CLIF project does not necessarily mean contributing data, the threshold for engagement is lower. Even with these resources, though, sites must cultivate close partnership between clinical and technical teams. Running CLIF projects requires familiarity with the high-level programming languages used within the consortium (ie, R and Python), version control workflows (eg, GitHub), and local analytic infrastructure.²²

Federated research also faces methodological challenges. Because patient-level data do not leave an institution, analytical flexibility may be reduced. Federated analyses presently work best for efficiently summarizable measures (eg, counts, means) and meta-analyzable effect estimates, but further development in distributed and privacy-preserving methods may be needed for pooled modeling of tasks like individualized treatment effect estimation. More broadly, the rise of large language models has prompted speculation about whether automated harmonization might supplant common data models altogether. While these tools may support efficient and accurate semantic alignment across heterogeneous EHRs in the near future, they cannot yet replace the transparency, reproducibility, or rigor of a well-structured shared data model. For federated science to be trusted and auditable, variable definitions and transformations must be explicit, version-controlled, and reproducible—goals that CLIF is designed to fulfill.

All observational research is at risk of bias. CLIF’s grassroots development may introduce selection bias, as participation has grown through professional networks rather than formal sampling. Academic tertiary and quaternary hospitals are highly represented, although many participating systems include community and rural hospitals. More broadly, however, CLIF-based studies may be less susceptible to certain biases than analyses using existing datasets. For example, centralized enclaves often depend on complete-case observations, introducing selection bias through informative missingness. Moreover, datasets limited to ICU-treated patients—by excluding patients who die before critical care, whose care goals preclude escalation, or who have “borderline” illness severity (ie, might have been in an ICU at a different hospital)—face intrinsic risk of collider bias for studies of many exposure/outcome permutations.

CLIF remains a work in progress, with several core tables and core concepts still under active development. CLIF’s modular architecture allows scaling across age groups and clinical contexts (eg, pediatric or transplant-specific cohorts). Moreover, CLIF’s underlying principles and infrastructure extend beyond critical care, with broad applicability for any hospital locations where timely, structured EHR data can enhance research and quality improvement. Future development will expand CLIF’s interoperability with OMOP and Health Level Seven standards for data exchange.²³ Despite these limitations, the federated approach has enabled us to complete projects that would remain stalled under centralized models, and at substantially larger scales. By shifting the primary bottleneck from data access to analysis—where, in our view, it rightfully belongs—federation accelerates, broadens the scope, and enhances the impact of ICU research.

Supplementary Material

Supplementary material is available at Annals of the American Thoracic Society online.