The Healthcare AI Boom's Blind Spot: Patient Identity and Data Integrity

The excitement in healthcare AI seems to track its value. The numbers are big enough to warrant attention by themselves. In a recent Signal Room discussion, health information expert Lorraine Fernandes said estimates cite about $38 to $39 billion in AI spending last year, with about $50 billion in expected spending by the end of this year, and highlighted that healthcare adoption is accelerating at twice the rate of other industries. What does not get as much coverage is what determines if any of that spending pays off. In a word, Fernandes said it is trust. Trust is not a marketing gimmick; it is won or lost at the patient level, at the identity and data levels, long before any model outputs an answer.

With 50 years of health information management experience, including a tenure on the board of IFHIMA, the International Federation of Health Information Management Associations, where she still directs marketing and communications, Fernandes has a strong position. On the podcast, her argument was simple. The same systems that are capturing that extensive AI funding are also capturing data that is created in the real world by people and processes that are overlooked. If that data is not trustworthy, the benefits that the spending intended to create will not be realized.

The Boom Is Easy to Count; the Risk Is Hard to See

Fernandes defies the conventional view, which treats spending as an indicator of capacity. Reading these figures as capability is less insightful; Fernandes sees them as a problem of exposure. Every dollar invested integrates itself into systems that combine, analyze, and disseminate clinical and administrative data, and, more and more often, those systems, or the systems they use, operate on AI. If they are not doing so now, Fernandes contends, they will be in a matter of weeks or months. Despite the economic boom, a successful model depends on the data moving through the system being accurate and properly attributed to the right patient.

This is the oversimplified story of the boom in the spending. The model can be technically very good, and still wrong. The reason can be a fragmented or inconsistent identity issue, and the failure can remain hidden and manifest later in the form of a prediction that misfires or a decision that confidently advances on a flawed foundation. The integrity problem does not shout; the spending is noticeable, and the integrity problem is, in effect, invisible.

Trust Is an Engineering Property, Not a Sentiment

Fernandes was careful to say that conversations surrounding trust are not new. Healthcare has been discussing this for decades and long before the current trend of AI. The difference today is the ramifications of getting it wrong. Errors used to mean a clinician misinterpreting a chart, where a human could catch the mistake. If the system that is consuming the data is automated and reasoning at a massive scale, the error is propagated.

Trust must be built across the entire life cycle, from the creation and analysis to the data and systems that consume it. It is important to understand that relocating the work is the most important thing. A vendor's ethics statement or assurance will not create trust. If the identity and the record and the meaning of the data are protected at each handoff, then the trust is created. Otherwise, the purpose of the automation will be lost, and the digital health and AI initiatives and the value behind them will not be recovered.

Those Transitioning From Data Creators to Data Stewards

Fernandes anchors this in health information management for good reason, as this profession has experienced that entire journey. HIM's transition from paper to electronic records came after the significant investment in federal digitization in the U.S. between 2008 and 2012. HIM professionals moved from the creation of data, through the release of information, clinical coding, registration, and patient access, to becoming data-analyst and governance and clinical-documentation roles, and then stewards of data through governance, compliance, privacy, and advocacy. HIM professionals constantly reshape their roles to fit the needs of their profession at any given time, and this highlights the need to specify who owns data integrity in the age of AI.

Fernandes made that assertion clear. Who would be better to manage data than the professionals who created it, know its limitations, and now use it in entirely new ways? HIM professionals understand where records are thin, where coded data is imprecise, and where two systems disagree about the same patient. HIM professionals are able to articulate those limitations, and that is precisely the context that an AI program would require in order to trust the data, not after.

The Connectors Between Clinical, Technical, and Administrative

One of the more interesting observations offered by Fernandes was that the profession has evolved into a set of connectors, a very specific type of bridging, spanning the clinical and technical functions and the administrative ones, which each contain a segment of the patient record yet, at times, do not communicate with each other. Patient identity is where those segments either resolve or fragment. A laboratory system, a registration desk, and a coding workflow can each be, from an internal perspective, right and still describe the same person differently, and no model that is processing the output can notice.

That is why she considers the human factor to be foundational as opposed to incidental. The toolkit that IFHIMA launched, which she explained was the outcome of around ten contributors from their respective countries, which included Barbados, Iran, Kenya, Spain, and Australia, was created partially to demonstrate why HIM should be included in the design and deployment of AI, not only in the aftermath. The connector role only operates when the systems in question are being constructed.

Why the Identity Problem Differs Globally

Fernandes acknowledged the discipline tying all this together is unevenly resourced. Countries vary in the HIM role, she said, and many of them are hampered by outdated job and labor classification codes, the absence of HIM formal academic programs, or a lack of recognized credentials. In the US, where neither programs nor credentials are the constraint, she cited the live tension of ensuring compensation is equitable to the investment the work requires.

There is also a standards element to identity and integrity. She commented on ICD-11, SNOMED, and the mappings between them, along with WHO classifications including ICF and ICHI. She observed that the US will be a late adopter of ICD-11 while the rest of the world embraces it first. Mapping across these systems is where a diagnosis or a procedure can be preserved or distorted as data crosses borders and systems, another quiet place where integrity is won or lost. While this may not be the most exciting job, it is of the utmost importance in ensuring that the data the AI model works with is appropriate and correct.

How Hutchins Approaches Patient Identity and Data Integrity

Our work tends to start where the boom narrative ends — with whether the records underneath a planned AI use case can be trusted to represent the right patient, consistently, across the systems that need to join into one view. We treat identity and integrity as stewardship work, not a one-time cleanup: the same discipline of governance, accountability, and ongoing monitoring that keeps a foundation trustworthy as systems, codes, and feeds keep changing. That is why we approach this alongside healthcare data quality and data governance, and why we assess it against a specific use case rather than in the abstract. These themes run throughout The Signal Room podcast, where practitioners describe what trustworthy data actually takes before a model is allowed to depend on it.

Authoritative sources

• HL7 FHIR interoperability standard
• ONC — HealthIT.gov
• WHO — Ethics & governance of AI for health
• U.S. Department of Health & Human Services — HIPAA
• Agency for Healthcare Research and Quality