Outcomes Articles

Your Archived Patient Data Isn't Automatically Ready for AI

October 26, 2026

About This Report

Featuring insights from Erik Johnson, Heather Grey, and Jeff Thomas hosted by our CEO, Saul Marquez.

An old patient record can be easy to retrieve and still be difficult to use. A research team might find the note it needs, yet remain unsure what a field means, which context survived migration or whether the proposed reuse is permitted. Keeping the record solves an important problem. It doesn't answer the next research question.

That distinction matters when an AI project starts with a large archive and a promise to put it to work. The project owner needs more than a count of available records. Clinical specialists need enough context to interpret them, governance teams need to confirm permitted use and technology owners need to deliver them under the conditions the application requires.

Start with one proposed use, then examine what preservation has made possible and what remains unresolved. Three separate Outcomes Rocket conversations connect historical access, clinical meaning and practical delivery. The payoff is a specific review of whether these records fit this purpose, not an AI-ready label attached to the entire archive.

Preserve Usable Access, Not Just a Stored Record

When a health system retires an older application, what happens to the information inside it? The records may still exist, but that doesn't tell you whether someone can retrieve them, understand their original context, or use them for a new purpose.

Erik Johnson, Vice President of Marketing at Harmony Healthcare IT, discussed this challenge in his Outcomes Rocket interview. Healthcare organizations accumulate decades of clinical, financial, and operational information across different systems. Mergers, acquisitions, and EHR transitions can leave that history scattered across multiple applications, each with its own way of organizing and presenting information.

The conversation describes archiving as an active process of preserving, managing, and making historical data accessible. That can involve extracting information from older applications, maintaining access after migration, and preparing records for potential integration with newer systems. It's an important distinction. An archive should preserve the ability to work with information when a legitimate need arises.

But accessibility has limits. Suppose a research team wants to examine historical treatment patterns using records from a retired system. Finding the relevant files is a useful first step. The team still needs to know which fields survived, what context accompanied them, and whether the extracted information supports the question being asked.

There's also a governance question. Technical access to historical patient information doesn't automatically establish permission to reuse it for research or AI development. The responsible team must confirm the proposed use with the appropriate data-governance or privacy authority and determine which access restrictions and permissions apply.

An archive owner can begin by identifying the sources needed for one proposed use, what context survives and who is permitted to retrieve the records. Any unanswered questions become part of the project's fitness review.

Erik's discussion establishes why keeping historical information accessible matters. It doesn't establish that preserved records are suitable for training an AI model or producing a clinically meaningful result.

Accessible information still needs to mean what the next analysis assumes it means.

Recover the Clinical Meaning Behind the Fields

The next review moves from finding the record to interpreting it. A familiar abbreviation can look like a clean data field while concealing a distinction the research question depends on. More records don't settle that ambiguity. Someone has to examine what the original documentation actually says.

Heather Grey, Senior Vice President and General Manager of Real World Data, Clinical Trials and Registry at Omega Healthcare, described this challenge in her Outcomes Rocket interview with Saul Marquez. As Heather explained, "the gap between I have a lot of data and I have a lot of usable data is pretty huge."

She shared an example of a health system with a substantial backlog of historical information buried in unstructured clinical notes. In her example, notes using PSA, prostate-specific antigen and prostate-cancer testing illustrated how different wording can obscure a shared clinical concept. Her point was the work required to normalize that language into usable data.

The editorial takeaway for a new research use is to check the surrounding clinical context rather than treat related terms as automatically interchangeable. Heather described Omega's approach as combining technology with physicians, nurses, and clinical specialists who curate information and review its quality. Their work illustrates why extracting text or populating a field doesn't automatically establish that the resulting data accurately represents the clinical concept being studied.

For the proposed research question, identify the clinical concepts that must be represented and give reviewers examples of ambiguous notes. Specify what context they need, who resolves uncertain cases and how those decisions are checked. This turns a general request for clean data into a review that the clinical-data team can actually perform.

These questions help establish whether historical information is suitable for further evaluation. They don't certify its accuracy, establish reuse rights, or validate an AI model.

Even a carefully interpreted dataset must reach its intended user under the conditions the use actually requires.

Deliver Data on the Terms of the Intended Use

A well-interpreted dataset can still arrive in the wrong form or at the wrong time. Delivery must fit the work: who retrieves it, how recent it must be and what the request asks of the systems providing it. Those are operating requirements, not properties guaranteed by the archive.

Jeff Thomas, Senior Vice President and Chief Technology Officer at Sentara Health, discussed these operational constraints in his Outcomes Rocket interview. He described how valuable healthcare information can remain locked inside existing systems, making it difficult to support newer AI workloads. Accessing that information also creates demands on infrastructure originally designed for different purposes.

Jeff discussed approaches to making data available for AI inference without placing the same demands directly on production systems. His experience highlights a practical consideration: the way information is retrieved matters alongside what the information contains.

Freshness depends on the intended use. A research team analyzing treatment patterns over several years may need a consistent historical dataset with reliable context. A clinical workflow responding to a patient's current condition may require much more recent information. Applying the same freshness requirement to both would miss what each application actually needs.

The operational review should therefore ask who is authorized to access the information, how current it must be, what workload the existing systems can support, and how the result will reach its intended users. Security and privacy requirements also need review by the responsible teams.

The Sentara conversation also emphasizes helping people through organizational change. An accessible data layer doesn't teach an intended user which result to trust or where the new information belongs in daily work. Include those people when defining the delivery requirements, rather than handing them a completed technical design.

Readiness is now a test of one use against its actual requirements, not a badge attached to the whole archive.

Choose One Use and Identify the Missing Evidence

Before declaring an archive ready for AI, bring the archive owner, clinical-data specialist, operational technology owner, and intended user together around one proposed application. Choose a specific research question or operational decision, then create a short fitness brief documenting what that use actually requires.

Keep the fitness brief specific to that application:

Access: The historical sources, retrieval method and responsible archive owner.
Meaning: The clinical concepts, ambiguous cases and specialist responsible for interpretation.
Quality: The checks still needed and the person who reviews their results.
Permission: The proposed reuse and the governance authority responsible for confirming it.
Delivery: The freshness the intended user needs and the capacity and workflow the technology owner must assess.

Name an owner for every unanswered question. If a critical requirement cannot be supported, narrow the proposed use or defer it until the missing evidence is available. If those requirements can be met, proceed to the appropriate technical, predictive, or clinical validation. Completing the brief doesn't establish that an AI system will perform safely or effectively.

For the historical note in the opening, that means knowing not only where it is stored, but what it supports, who may use it and what needs checking. Until those questions have owners and answers, record volume is a poor substitute for readiness.

Keep the history. Define the use before declaring it ready.

About the Conversations

These are three separate Outcomes Rocket conversations, not a shared case study. During the time of recording, the guest names and titles were as follows: Erik Johnson, Vice President of Marketing at Harmony Healthcare IT. Heather Grey, Senior Vice President and General Manager of Real World Data, Clinical Trials and Registry at Omega Healthcare. Jeff Thomas, Senior Vice President and Chief Technology Officer at Sentara Health.