The Three Eras of Real-World Data

What real-world data is

Kevin opens with a piece Sandeep wrote for Manifold on how real-world data has evolved. Sandeep starts with a definition: real-world data is patient-level healthcare data generated as patients interact with the health system, along with data patients produce themselves, such as readings from a wearable. Put together, it lets researchers follow a patient’s journey over time.

That data is messy by nature, he says, because it comes from so many sources. Its value is in what it shows biopharma: how patients behave in a real-world setting, how treatments affect them, and what the outcomes are. In a nutshell, he calls it any and all biomedical data available on a patient, and the raw material for real-world evidence that informs both research and patient care.

Era one: claims data

Because the US healthcare system runs on payers, the first large source of real-world data was claims. When a patient fills a prescription, a claim is lodged with the payer. Once the payer reimburses it and the claim is closed, it becomes a data point. Multiply that across millions of patients and thousands of therapies, Sandeep says, and you can see treatment patterns, reimbursement costs, and how care differs between academic centers and community practices.

He places the start of this era in the late 1990s, with heavy pharmaceutical use in the early 2000s, mostly on the commercial side for geographic patterns and physician and patient targeting. Payers aggregate and de-identify the data at a cohort level. Claims remain, in his words, a scalable, widely available, almost commoditized asset, and closed claims still support robust real-world evidence.

Era two: clinical data from electronic health records

The second era added clinical data. Sandeep points to Flatiron Health, which he dates to around 2012, and its oncology EHR, OncoEMR, as an early, standardized source of research-grade clinical data that he still counts among the best available. Data from general EHR systems is harder to use. Even different Epic installations produce different data feeds, so pulling from 30 systems means wading through a lot of mess.

The most valuable detail, he says, sits in unstructured physician notes, and converting those notes into structured data through abstraction and curation remains the gold standard. Clinical data moves the picture beyond cost and utilization to why a patient received a treatment and how they progressed. His example is a patient with metastatic non-small cell lung cancer who has had chemotherapy, then a first-line treatment, then docetaxel. Claims can’t show that journey, along with comorbidities and social determinants of health. Clinical data can.

For drug developers, that detail supports patient selection and stratification for trials. A company can watch how a defined population progresses at three, six, and twelve months, which tells it when patients may become eligible for the drug it is developing. Kevin adds that the same data can show which patient populations still need new drugs, guiding where research dollars go.

Why pharma wasn’t ready

Kevin asks whether biopharma knew how to use EHR data as it arrived. Not at all, Sandeep says, and the gap isn’t fully closed. Data has to be structured, standardized, analytics-ready, and now AI-ready, across millions of patient lives and many organizations, systems, and analytical tools.

That is why real-world data providers such as Flatiron, ConcertAI, and TriNetX have done well, in his view. They do the heavy upfront work of bringing data streams together and structuring them so a pharma company can apply its own processes downstream.

Era three: multimodal data

The third era layers in omics, imaging, and wearable data. Sandeep credits the falling cost of sequencing, which he says is approaching about $100 per genome, for making DNA and RNA sequencing common, first in oncology and increasingly in rare disease and immunology. Add pathology images and social determinants of health, and real-world data goes from kilobytes and megabytes to gigabytes and petabytes. He names molecular diagnostics companies such as Tempus, Caris, and Guardant Health as strong in this kind of data, which he calls the bedrock of pharmaceutical R&D.

More data types also mean tighter cohorts. With claims, he says, you start from everything, say 100 million patient lives, and filter. Combining sequencing, proteomics, clinical, and claims data defines a much narrower population. The practical answer to the volume is to avoid moving the data at all: bring compute, tools, and capabilities to where it lives. Behind the scenes, standardization and harmonization are still burdensome, even with AI.

Trusted research environments

Few people outside the field know what a TRE is, Sandeep says, though architecturally it is simple. The data sits in the owner’s cloud storage, such as an S3 bucket or Azure blob, and a platform is deployed into that environment so the owner’s compute and storage are used in place. The platform carries the burden of keeping the environment secure.

In practice it is a controlled computing environment. Only approved researchers come in, they analyze sensitive patient data, and the data never leaves its source. No copies are made, provenance is clear, and access is limited to the data each person is approved to see. Kevin sums it up as a single, scalable, secure source of truth with controlled access for each type of user.

Bring your own tools, and trusted execution

Sandeep says biopharma has resisted platforms that limit researchers to built-in tools, because those researchers have their own scripts and pipelines. Manifold’s answer is bring your own tools: RStudio, SAS if you hold a license, Python, and bioinformatics pipelines in WDL or Nextflow. Kevin notes that this eases change management, since researchers can keep running analyses they wrote and validated.

AI shifts the question from which tools a person can use to which tools an AI can call. Sandeep suggests thinking of a trusted execution environment: a sandbox that gives the model the right context, a harness, and approved tools so it can carry out a natural-language request. Kevin connects this to the difference between probabilistic exploration and deterministic, validated software.

Sandeep argues that limiting an AI to specific data sets and tools makes its output far more deterministic. He shares an example he heard about: a physician at a client used AI on the platform for a grant analysis, gave the AI’s approach to a graduate student, and the student reproduced the same result by hand before it went into the application.

Forward-deployed engineers and why pilots stall

Even a good platform is complex, so Kevin asks how organizations get on board. Sandeep’s answer is forward-deployed engineers. Pilots are often run on very clean data, so they succeed, but things fall apart when a company tries to scale across the organization. Where vendors once sent translational research and clinical development experts to work with pharma counterparts, he sees medical affairs, commercial, and clinical development teams now working with forward-deployed engineers to deploy AI in-house.

In his framing, the TRE solves the governance problem, while forward-deployed engineering and tooling for self-serve ingestion and harmonization solve the adoption problem. The upfront work remains, and he expects the data producers who make their data easy to analyze to win. While preparing for the episode, he says, he saw that Axtria aimed to have a thousand forward-deployed engineers working at biopharma sites, much like the on-site consultants of his own consulting days.

Kevin, whose company Galang AI presents the show, says the same question applies across industries: what is the goal, what are the inputs, and where is AI best used. Point it at the wrong step and you optimize something that doesn’t matter. Sandeep tells of a Claude Cowork workshop where the instructor described a coffee shop owner who runs nearly everything, from ordering supplies to bookkeeping, through Claude. Few people work that way, he says, which is why companies need help implementing AI methodically rather than spraying and praying.

Frontier AI labs, the wet lab, and wearables

Kevin asks about frontier AI labs moving into life sciences. Sandeep believes they see a real chance to change how drugs are discovered and developed, from target identification to lead optimization, and to shorten what he calls a 15-year development cycle. He also thinks they recognize that, unlike coding, life sciences can’t be taken on all at once, so they will move piece by piece. He says Anthropic has set up a wet lab to learn which processes AI can change.

There is also a revenue story, he says. He compares the revenue of the top two GLP-1 drugs this year with what Anthropic, or Anthropic and OpenAI together, made in revenue, and qualifies the comparison with “if I’m correct.” His point is that pharma is a very large market the labs want to take part in.

Kevin notes that software development collapsed as a barrier faster than anyone expected, but biology still needs physical experiments that play out over time. He mentions Ginkgo Bioworks’ automated labs as one way that gap may narrow. Sandeep adds that wearables such as the Oura ring are, in his view, less about the device than about the data they accumulate and what it can reveal about the human condition.

Trusting AI with sensitive data

Kevin closes by asking whether organizations trust frontier labs with their data and prompts. Sandeep won’t guess how the labs handle it. He notes that Anthropic and OpenAI are widely deployed in biopharma, and assumes the data isn’t going to them, though he says he could be wrong.

On Manifold’s side, protecting data providers’ patient-level data is paramount. But protection alone isn’t enough, he says. People still need access and insight, so the platform runs AI, with model switching, inside a sandbox. Controls limit what approved users can do, such as cutting cohorts and exporting outputs in permitted ways, so no one can quietly train a model on a provider’s data and walk away with it. Over time he sees federated querying and federated learning extending the same approach across data sets.