Federated analytics has proven it can work: Now, pharma must make it repeatable

Digital
Analytics concept

While pilot programmes are proving the impact of federated analytics on research and the production of traceable, reusable evidence, building a repeatable system that goes beyond individual studies remains a challenge.

Decisions about which medicines to develop, which patients to study, and how to demonstrate value to regulators and payers depend on a detailed understanding of the people a treatment is intended to help.

For researchers, this implies patient-level data that combines clinical history, molecular characteristics, treatments, and outcomes for precisely defined populations. Finding enough patients with the required depth and diversity of information will need access to data sources in different organisations and countries. Yet, bringing those records together is not always practical or permitted.

Federated analytics lets pharma researchers investigate research questions across those sources without moving patient-level records into their own systems. Study-specific analyses run where the data is held, with approved aggregate results returned to the sponsor for combined analysis.

The appeal is clear: research can extend beyond datasets that can be brought together centrally. A pilot can demonstrate that this approach works for a defined research question. But that success does not establish a dependable, repeatable way to deliver subsequent studies across different sites, countries, and patient populations. For pharma, the question is what can carry forward, and what needs to be built from scratch.

Making it work more than once

Developing a bespoke software solution and building an enterprise platform are fundamentally different undertakings. A bespoke solution can be designed around a specific set of requirements, with tailored integrations and dedicated implementation effort. An enterprise platform must go further. It needs reusable components, standardised interfaces, and consistent security and governance controls, all built on foundations that can accommodate new requirements without being rebuilt each time.

The same distinction applies to federated analytics of real-world data (RWD). A single study can be delivered through close working relationships, individually negotiated approvals, and substantial manual coordination. A reusable federated network, however, requires more than the ability to execute a study across multiple sites. It needs the technical infrastructure, governance frameworks, and operational processes to support successive studies consistently, while accommodating different research questions, participating institutions, and sponsors.

Where implementation breaks down

In my experience, implementation encounters three recurring obstacles: differences in governance and approval processes, analytical interoperability, and uneven local delivery capacity.

For a pharmaceutical company sourcing data through a vendor, governance differences can affect which sources are available and when an analysis can begin. Hospitals, registries, and other data custodians do not necessarily approve data use in the same way, even within one country. The delivery plan needs to account for those approval responsibilities alongside differences in technical arrangements and local capacity. A shared technical connection does not resolve them.

The second issue is analytical interoperability. Connecting sites and running the same code does not establish that their data describes equivalent patients, treatments, or outcomes. If hospitals record heart failure differently or code social determinants of health with varying definitions, the analysis may execute successfully while selecting different patient populations at each site. A harmonised and curated set of common definitions and checks is needed before those results can be meaningfully compared.

The third issue is local delivery capacity. Some sites have well-resourced data teams and established, secure environments for analysis; others are still digitising records or depend on a single analyst to handle requests. A model built around the most advanced site leaves the rest without a realistic route to participate. The operating model needs to account for those differences in the responsibilities and support expected at each site.

What needs to be in place before the technology

Before technology is deployed, the sponsor needs to define the research question and intended use of the evidence and agree on the data requirements and analytical approach with its vendor. In a vendor-led engagement, the vendor coordinates delivery with data custodians, including shared definitions, local responsibilities, approval processes, and output checks. The sponsor needs visibility into those arrangements and their implications for the study, rather than responsibility for organising every site. Together, these decisions define the operating model that the technology must support.

Planning for data standards and federated analysis when writing the protocol makes it easier to build on the work in later studies. Each new research question will still need its own assessment, but teams can draw on definitions and methods they have already agreed, with a clear understanding of who reviews the outputs. They do not have to start those discussions from scratch every time.

Technology then has a clear job to do, and trusted research environments (TREs) are a good example. TREs give data custodians a controlled place to hold their data, where researchers can analyse it without it being transferred. Running standardised analytics across a federated network of TREs, and combining the results, produces high-quality, traceable evidence. However, TREs only deliver repeatable evidence when they sit within an agreed way of working.

Shared definitions should also be agreed, because this is where many programmes fail. Approved analytics models can only run consistently across sites if those sites describe and annotate their data in a standardised way. For example, the Observational Medical Outcomes Partnership (OMOP) Common Data Model gives organisations a shared structure and data vocabulary, so one analysis can run unchanged across many datasets without data needing to move.

A real-world example

A pharma company has a product treating a rare lung disorder in late-stage trials. As it is a novel therapy, they have an emergency use authorisation in the US. Their existing drug for the same indication and its competitor have had their patents expire and are competing with generic alternatives. This novel drug also has a competitor at the same stage in the pipeline. The pharma company has been running trials in the EU, but they need RWD to answer various current and future questions. The sponsor began with a single RWD study across multiple sites in the EU, for a specific research question, but when they expanded to APAC, and wanted to convert the current RWD study framework into one that could become a reusable federated evidence generation network, they encountered delays in redoing the governance and approvals. Importantly, the analytics and data assets were not suitable for reuse, as they were built specifically to answer a single research question. And not all of the sites were able to provide data in time or with the same longitudinal quality. Federated analytics worked well for the single study, but they had to almost begin afresh to build reusability into the federated network.

Making it repeatable

Federated analytics will not scale through increased awareness alone. Pharma and its partners need an operating model that makes differences in systems, governance, and local capability manageable across successive studies. The practical test is whether subsequent studies can build on agreed definitions, approval responsibilities, analytical methods, and output-review processes, with changes assessed for each new research question.

If each study requires those arrangements to be rebuilt, delivery still depends on exceptional effort, rather than a repeatable model. For pharma, the question is whether its partner can explain what is already in place, what needs to change for the next research question, and how those changes affect feasibility, timing, and confidence in the findings.

About the author

Narasimha Kumar is global head of technology and data services at BC Platforms. Kumar brings 25+ years of experience in technology consulting, AI product management, and compliance across life sciences and healthcare. Previously chief product and commercial officer at Datafoundry AI and senior leader at PAREXEL International, Kumar specialises in AI-driven platforms and cloud transformation.

Image
Narasimha Kumar
profile mask

Narasimha Kumar