An AI Fix? That May Mean Major Data Remodeling

When healthcare organizations call Brian Laberge for help deploying artificial intelligence (AI) applications, they usually have some awareness that their data has problems.

“It’s not always about inaccurate AI outputs,” says Laberge, a solution engineer at the consultancy Wolters Kluwer Health. “Sometimes the first signs are subtle: inconsistent reports, unexplained abnormalities or friction in workflow.”

Brian Laberge

Often, those seemingly minor glitches are signs of a much bigger problem. “It’s only when teams begin integrating data across systems or preparing it for AI that they encounter what I call the ‘subfloor moment,’ ” Laberge says. “You think you’re just replacing the carpet, but then discover the subfloor is rotted in multiple places.”

As more healthcare organizations begin to leverage the power of AI, Laberge has become accustomed to “subfloor moments,” in which healthcare organizations quickly realize that the ways they have cataloged and stored data in the past are not well suited to the AI era. That means those organizations must find ways to prepare and repair their data infrastructures so they can make good use of the efficiencies and insights provided by AI.

Shrikanth Shetty, the chief growth officer and global head of life sciences and healthcare industries at the IT services provider HCLTech, says one of the first hurdles to overcome is data “silos.” Many systems that house healthcare data were built to operate independently, making it difficult to obtain data across systems. AI algorithms, however, can only generate solid insights if they have access to complete and accurate information. In other words, AI needs information that is free-flowing and not bottled up in silos.

Streamlining everything

Yet, Shetty says the problem is not simply about barriers. Separate data sets often have their own protocols and terminology. Even if the “silos” could communicate, they would not speak the same language. “Think of simple things: the patient record,” he explains. “If you see it across multiple systems, you may not be able to reconcile it across those systems because the nomenclature is not there, the standardized system is not there.”

Shrikanth Shetty

Shetty says the problem stems from how information technology (IT) evolved. Fifteen or 20 years ago, the goal of data storage was not to power AI. “At that point in time, the objective was to digitize,” he says. Nowadays, the goal is quite different. Instead of thinking of information technology as a way to streamline recordkeeping, healthcare executives see it as a way to streamline just about everything.

Although the healthcare industry has a reputation for being a late adopter of technological advances, partly because it is such a highly regulated part of the economy, Shetty says healthcare leaders are moving quickly to embrace AI because they see straightforward financial benefits. “There is a clear ROI [return on investment] here,” he says. “It’s not experimentation.” Yet, as Laberge’s “subfloor moments” illustrate, eagerness to embrace AI differs greatly from readiness to utilize it. A recent survey from the health and technology consultancy Nordic is instructive. When the company surveyed 127 healthcare industry executives earlier this year, they found that 25% of respondents reported being “very” confident in the governance structures in place to manage data at their healthcare organizations and 45% were “somewhat confident.” However, when asked whether their data governance was “well established,” only a quarter of respondents responded affirmatively. Kevin Erdal, Nordic’s senior vice president for transformation and innovation services, said the survey results show a disconnect between perception and reality.

“While many executives believe they are taking the right steps now to adopt AI, there are many components required for long-term success,” he said in a May news release.

Laberge says the consequences of a lack of data preparedness can be significant. He recalls the problems health plans can face when they have incomplete or improperly coded lab data. “For instance, in one pilot process, we discovered that in just over 30% of the data, the code provided didn’t match the expected description of the lab test,” he says. “This discrepancy can lead to significant problems in ensuring the accuracy of health metrics and measures like HEDIS [Health Effectiveness Data and Information Set].”

Quick ROI

Fixing the readiness gap is not a simple task, but neither is it insurmountable. Laberge says an important first step is to catalog or audit all the organization’s existing data sources to get a clear understanding of the current state of the organization’s data.

“Understanding the accuracy and completeness of these data sources is crucial,” he says.

The eventual goal is to get to what is known as a “golden record” — a unified record that is accurate and can work across all the entity’s various IT systems. Getting to that end goal starts with agreeing upon a standardized format for data storage and a standardized nomenclature. Rather than transforming data all at once, Shetty advises clients to find areas where they believe they already have “clean” data and then find AI use cases with which to experiment. He says that approach is much more manageable than trying to do large-scale transformation up front. “Chunk it out, see where those early low-hanging fruits are, where you can get some early benefits, and where you can get clear ROI,” he says.

A few areas will yield fairly quick ROIs, according to Shetty. The first is using AI to optimize internal IT processes, also known as “AI for business transformation.” Another easy opportunity is streamlining internal processes like revenue cycle management or providing services to employees. A third low-hanging fruit is improving patient or member experience by making it easier to do things like schedule appointments or consult a nurse.

Laberge says companies can also be proactive by embedding a terminology server, a piece of software that can help capture pieces of information even when synonyms or alternate terminology are used. He says the server can ensure that inbound data are clinically validated, properly mapped and categorized. “This helps jump-start data quality improvement as additional data [are] received,” he says.

Some caution

One of the main reasons healthcare organizations have often been wary of IT innovations is that mistakes in healthcare technology implementation can have life-and-death consequences. That is why Shetty says most healthcare organizations remain skittish about AI use cases that directly impact patient care. “If something bad happens, it’s a huge reputational
loss,” he says.

Shetty adds that patient-care use cases are also delicate due to the data security and privacy implications. “Remember, healthcare records have far more value on the dark web than simple credit card numbers,” he notes.

Still, Shetty says AI is making its way into the clinic, primarily by offering insights that can guide human practitioners. For instance, he says AI can be used to identify patients at high risk of readmission. “You’re not doing actual clinical work [with AI], but you are utilizing the data that [are] there to advise the doctors,” he says.

Shetty says one advantage of the AI era is that it is not necessary to invest a gigantic sum in hardware to get started. Most of what is needed can be purchased as a service, and firms like HCLTech offer accelerator products that expedite AI integration. “They can be bought as services, and you can scale them depending on how you want to scale them,” he says.

Ultimately, Shetty says the task of cleaning up and aligning data systems is necessary, even if a healthcare organization is not ready to embrace AI just yet. “Everyone realizes that their internal data [are] the most crucial thing that any organization has,” he says.

Any investment made to repair an organization’s data foundation will lead to long-term benefits, he says. “These are things which have to be done,” Shetty says. “These are foundational, and these will bear results — if not now, down the line as well.”