Following the launch of ChatGPT in November 2022, the early hype surrounding generative AI was that it would revolutionize healthcare by replacing physician and nurse decision-making due to its vast data sources and rapid analytical capabilities.
GenAI creates new content – such as text, images, and code – through algorithms that learn from existing datasets and then generate content similar to the training data. The technology’s potential made it seem as if healthcare was on the cusp of having an AI tool that could diagnose and treat patients faster and with greater accuracy.
But survey data shows healthcare professionals are conflicted about GenAI, their excitement mixed with apprehension and even fear. In different surveys from 2024, 85% of hospital and health system leaders report GenAI as “the most exciting technology to emerge in healthcare.”
Nonetheless, 83% report concerns about the role of AI in patient care, while 91% believe that AI in healthcare needs supervision. As a result, only 25% have implemented AI at their organizations, and 21% feel that they are missing out on significant tech-driven opportunities.
Uncertainty Reigns
Some of the reluctance surrounding AI adoption has been influenced by the hype. Healthcare leaders have been told and read so often that “AI will transform healthcare” and that starting on that journey may seem intimidating. Their concerns are understandable given that electronic health record (EHR) systems (another massively hyped event in healthcare) have in the past 15 years been attributed to hospital bankruptcies, clinician burnout, and lower patient experiences.
AI, however, does already have numerous, smaller, practical use cases in healthcare – that while perhaps not world-changing – can deliver positive clinical, operational, and financial benefits immediately. These AI entry points can help organizations build experience, knowledge, and the confidence to take on more ambitious, higher impact solutions in the years ahead.
AI’s Value Today
Media coverage of AI in healthcare today is dominated by clickbait, like the time when ChatGPT passed the U.S. medical licensing exams or when it co-authored a research paper published in a respected medical journal. Although certainly newsworthy, these anecdotes fail to demonstrate the way that AI is transforming healthcare today.
Specially trained large language models (LLMs), a subset of GenAI that applies machine learning to massive data sets for linguistic tasks such as text generation and answering questions, are finally helping health systems leverage data in much less time and with far fewer IT or financial resources. AI is now delivering many of the data management benefits that we hoped for back when the debate on IT-system interoperability began over a decade ago.
Since then, health systems have added dozens of new systems to their enterprises to manage imaging studies, lab results, cardiovascular data, clinical decision support, remote patient monitoring, and many other needs. Instead of seamlessly sharing data, however, each system was required to have their own IT staff design custom integration programs (ETLs) to clean and standardize the data to be entered into the organization’s enterprise data warehouse. If a new data set became available that the ETL did not recognize, the program would be unusable until the IT team was required to update the software.
The many variabilities and inconsistencies across healthcare data sets made updating these ETLs a routine – and costly — task within IT departments. Organizations that can afford it employ entire teams dedicated to maintaining such systems.
AI, in effect, eliminates the need for ETLs. A properly trained LLM can recognize and extract data from nearly any system, no matter how it was input, and enter it into an analytics tool, research or registry database, or any other system using the coding language and/or format specified.
A simplified hypothetical illustration of this capability is documenting patient height. In some systems, height is entered as a number followed by the words “feet” and “inches”, another may list it as only inches, and another may contain an apostrophe to indicate feet and double apostrophe for inches.
What if, for your use case, the height needs to be in centimeters? Instead of reprogramming, testing and troubleshooting an ETL program, which may take days or weeks, a clinician can query the AI to identify, extract, and enter the height of a specific patient population. In a few seconds or minutes, the algorithm can recognize all the possible variables for the value, pull the data, convert it to centimeters, and enter it into the database.
Searching for, identifying, and translating patient height values in different databases is tedious and time consuming. Yet, clinical data abstractors must perform this task hundreds of times a week to fulfill required submissions of this data to various clinical registries.
Clinical data abstractors, who are typically highly trained and experienced nurses, need to track down numerous data points beyond height, including patient demographics, vital signs, procedures, medications and outcomes, depending on the registry. Filling out just one form can take two to three hours. Updating just one clinical data registry can consume weeks of time because information is listed in structured and unstructured formats in numerous nomenclatures. That valuable time could instead be applied to care delivery that would free up hospital beds sooner, shorten emergency department wait times, generate revenue, and reduce feelings of nurse burnout associated with such data processing tasks.
Leading health systems have recognized that clinical data abstraction is an ideal entry point for adopting AI. It is the ideal task – repetitive, labor-intensive and time-consuming – for LLMs to tackle where health systems can experience immediate time, cost and patient-care quality benefits. Updating clinical data registries builds experience and sets the stage for additional care-quality initiatives inside organizations where large amounts of structured and unstructured data in different systems and servers need to be accurately and reliably cleansed and inputted.
Certainly, AI today can also be applied to back-office tasks that automate and accelerate steps in the revenue cycle, such as claims coding, prior authorization requests, and claims denials requests. Clinical data registry updates, however, are an ideal entry point for clinical-care use cases that enable providers and administrators to envision a future where AI continuously delivers quality and safety insights pulled from a much broader data set. The next thing they know, they might realize that their organization is part of the AI “healthcare revolution.”
Christopher Mazzanti
Christopher Mazzanti is Chief Operating Officer at Carta Healthcare.
