To Scale AI in Healthcare, Fix the Data Foundation First

As a healthcare data leader, I have come to the conclusion that hidden issues in our data infrastructure, rather than the technology itself, are the main reason why most AI pilots fail. I have worked in healthcare data leadership for a little over 20 years, both in the large nonprofit systems like Mass General Brigham and in the rapidly expanding for-profit systems. From my experience, I would say that many AI pilots have been met with enthusiastic launches and have come to a halt with a deafening absence of activity.

What is causing this? It is almost never the algorithm.

It is a concerning trend that the data maturity that is required to scale AI in healthcare is often overlooked. The focus tends to be on large language models, ambient listening, and clinical decision support. All of these are inarguably valuable. However, if your data is fragmented and inconsistent or there is high technical debt, all of those tools will be ineffective.

What Is Holding Us Back Is Unreliable Data

In most of the large healthcare systems I have worked with, I have seen the same issue: even though the systems have the same EHR, the ways that clinicians document care are different. One clinician uses structured documentation; one uses unstructured. When this is done hundreds of times, the data is not fit for AI at scale.

This is not a problem for the IT department; this is the responsibility of all company leadership. Unlabeled and siloed data will prevent large language models from being able to make clinical connections. AI is not able to find the answers to questions if we have not done the work to standardize the data for it.

What a Real Scalable Data Strategy Looks Like

Investing in the early building blocks of an initiative looks unglamorous. All the blocks must be accounted for if AI is to do more than produce show pieces and be dismissed.

Normalization means having a consistent data set for every input. This includes consistent application and spelling for data labels and timestamps. Data governance means instituting a standard for the responsible application of data and the validation of the data for clinical application. Effective clinical partnerships require frontline clinicians in the design stage of the workflow for the integration of AI.

While in charge of data consolidation at the enterprise level, I further aligned data governance and operational workflows. I opened the pipes for data integration, consolidation and alignment to be clinical and operational data for decision support. As an example, we built specific data integration pipes for the needs of Endocrinology as the general data integration pipes could not support the clinical requirements.

Four Field-Based Applications

Here are my recommendations if you want to take GenAI seriously in a health system:

Understand what your technology lacks. You can’t build AI on a foundation of disarray.

Create the right infrastructure for humans. Pair data scientists and clinical informatics to design the right data.

Prioritize low-risk and high-reward projects. This is the best way to earn points in your organization.

Incorporate the voice of the end user and clinicians. Frequent validation ensures product adoption.

Without taking these steps, you’re adding features to a system that no one will want to use.

Conclusion

We have no obligation to follow trends in new technologies. Our responsibility is to enable nurses to spend sufficient time with patients, allow doctors the opportunity to maintain focus with patients rather than being distracted by the need to document, and facilitate the ability of clinicians to return home to their families in a timely manner.

GenAI has the potential to be a revolutionary tool, but only when it is implemented with high quality and reliable data. If we are successful in providing this foundation, we will shift our focus from conducting pilots to having a meaningful and positive impact in the world.