Healthcare AI: Is Your Patient Data Ready?

The following is a guest article by Bob Stanley director of special projects at Melissa

Excellent AI models can’t win if the patient is not represented accurately 

AI has arrived in healthcare, and adoption has scaled up quickly. As many as 81% of physicians now use some form of AI professionally, versus 38% in 2023 according to a recent AMA survey. No, 81% are not practicing AI-assisted medicine, but they are increasingly relying on it to support ordinary clinical work. Summarizing research and standards of care was found to be an AI function for 39% of respondents, while 30% rely on it for creating discharge instructions, care plans, or progress notes, and 28% use it in documenting billing codes, charts, and visit notes. AI’s technical value has become clear in administrative and workflow automations, but there is much more ahead. 

As adoption accelerates, hospital and health organizations face the challenge of ensuring reliability and effective governance of these new systems. More than just understanding what an AI model can do, these leaders must be 100% confident in their ability to successfully deploy, manage, and benefit from its advancements. 

Two key questions determine their readiness: How reliable is the data used to build and validate the model? And is the real-world patient data fueling the model verified to be accurate, complete, current, and connected to the right patient? ‘Mostly or ‘almost’ are inadequate responses, as trustworthy patient data must be at the core of any healthcare AI implementation. Over the next 12-18 months, healthcare organizations can anticipate close coordination between data quality programs and the approval processes for getting their AI platforms off the ground. Ultimately, data deemed unsafe or low‑quality is likely to (correctly) stand in the way, justifying a block or suspension of an AI deployment.

A checklist for great data feeding your AI model

Bad results can still be generated by an excellent healthcare AI model, if its reasoning is based on an inaccurate representation of the patient. The basics matter, with common issues like duplicate records, incorrect identities, outdated addresses, missing medications, or fragmented medical histories becoming significant in new ways.  The industry needs to acknowledge accurate patient identity and trustworthy records are key to developing safe AI infrastructure.

It’s not about completely reworking the data environment supporting AI but rather ensuring ongoing governance that prioritizes patient data accuracy. Optimized processes must cover end-to-end data quality, following a checklist that begins well upstream of AI and works continuously through ongoing data monitoring:

1. Starting at patient registration, validate and standardize identity attributes including name, address, and contact data aligned to Project US@, a standardization initiative of ONC (Office of the National Coordinator for Health Information Technology). 
2. Before records train or trigger an AI process, resolve and deduplicate across EHR, claims, ancillary, and affiliate feeds.
3. Pay attention to the completeness and recency of data fields consumed by your AI platform, profiling info such as medications, allergies, problems, coverage, and contact information.
4. Maintain continuous monitoring of identity breaks; these are not back-office cleanup tasks and need to be handled as AI incidents.
5. Establish requirements for evidence of data fitness as part of your AI implementation plan, equally as important as model-performance evidence.

Understanding that AI is a magnifying glass

AI amplifies whatever condition the patient record is in. Weaknesses that once caused an administrative headache (i.e., outdated address, misspelled name, or duplicate records) become more consequential when the same data is consumed at scale – and considered the foundation of truth for that patient. Medication data puts it in context. If the patient record includes an incorrect dose, inaccuracies in the route of administration, or an incomplete list of medications, an AI system may process all data correctly but arrive at the wrong conclusion. This applies to all kinds of automated workflows like risk scoring, patient alerts, and insurance eligibility decisions. Even as mistakes are being made, your AI may be acting confidently on information that was wrong before it ever got to the model.

 Any organization implementing AI must be confident that the records feeding the platform actually belong to that patient and are accurate across their medical history.  Patient records must be consolidated and reliable, supported by processes ensuring identity resolution, record matching, and deduplication. These should be considered ‘standards of data care,’ protecting safety and excellence in overall care, but also extending that value as the same data becomes fuel to power AI processes. For example, Melissa applies matching and deduplication technologies to reconcile records across systems, and create a more reliable, unified patient view.

Consider that duplicate patient records are already recognized as an issue in healthcare business operations as well as delivery-of-care problems, impacting both healthcare research and clinic interactions. (Melissa client data analysis consistently shows duplicate rates of 8–10%. That is 80,000 duplicates for every million patient records, creating a cascade of impact on cost, risk, and correctness of medical and administrative decision supported by AI at scale.) Fragmented records add to the potential for error and additional risk of poorly applied AI processes, for example if a patient record is missing potentially contraindicated supplements or medications. 

Trustworthy patient records are the hidden infrastructure behind reliable AI 

Optimum data quality begins when the information enters the healthcare system, and features processes for ensuring data is correct and validated well before the workflow of any AI algorithm. For address validation and standardization at patient registration, Melissa tools and solutions are compliant to the requirements of Project US@, a unified effort to standardize patient addresses across and between healthcare systems. Project US@’s technical specification is designed to eliminate fragmented or duplicate records downstream by preventing variations in names, addresses, and other identifying data.

Ultimately, AI is only as safe as the dynamic data quality management and governance that supports it. AI readiness is an ongoing exercise in keeping data clean; as new records enter the organization and change over time, healthcare organizations must be committed to mechanisms that profile, validate, cleanse, match, and monitor information from the beginning and throughout the data’s lifecycle. Melissa role in this effort continues to grow in step with today’s complex and diverse healthcare systems – keeping patient data ready for evolving AI and digital health tools.

About Bob Stanley

Bob Stanley, Melissa’s director of special projects, helps Melissa Informatics’ clients harness the entire data lifecycle for business, pharmaceutical and clinical data insight and discovery. Click here to access Melissa’s customer data health checklist.  Connect with Bob via LinkedIn.

Melissa is a proud sponsor of Healthcare Scene.