By Dr. Ron Razmi

A few weeks ago, I attended the World Economic Forum in Davos and the discussion of AI as the next great technology dominated every corner of this important meeting. From the main event to almost every side event, it seemed that Artificial Intelligence had to be part of every agenda. The breakthroughs that made modern AI possible happened over a decade ago and now we’re starting to actually learn how to use it in every industry. Healthcare is no exception. I have a hard time reading a newsletter or any medical journals these days without running into an article about the applications of AI in healthcare. As a former Cardiologist, who has spent the last 15 years on the business side of healthcare, a good part of those years building a health AI company and writing a book about this topic, I couldn’t be more excited to see all of this. Yet, I spend most of my time writing and speaking about how there are major barriers that will slow down the diffusion of this technology into healthcare.

In my new book, AI Doctor: The Rise of Artificial Intelligence in Healthcare, I devote an entire chapter to these barriers to AI adoption in healthcare and another one to discussing the business models for healthcare executives considering investments in AI solutions. Both of these topics merit some discussion in this article and until they’re addressed, we will not start seeing much of the promise for this technology in healthcare. As I write in the book, I’m very optimistic about what AI will eventually achieve to improve everyone’s health but see a tough road ahead to achieve many of the high-value use cases. Why is that?

If there’s anything that is not a luxury in AI, it’s data. And, when it comes to health, AI needs complete and accurate data. Amazon and Netflix can figure out something about your shopping and tv viewing habits with the limited time you spend on their sites and improve your overall experience. If they get something wrong or if their algorithms mess up, c'est la vie. There will be no real harm that will come your way from those issues. In healthcare, one piece of missing or old data, like your creatinine level, can change the entire picture of your health. An algorithm that makes recommendations without the missing data can do real harm to you. If those using AI for doing research about health do not have access to high-quality data in genomics, proteomics, phenotype data, labs, etc., their research efforts will be significantly hampered, and AI would provide limited benefit.

The current state of data in healthcare is chaotic, to put it nicely. A person’s data resides in different databases (Electronic Health Records,) is often outdated, and contains errors. Most people do not have genetic data and the state of the available data often makes it hard to use with AI. For example, much of the historic data used in drug discovery experiments is recorded in lab notebooks and has to be digitized before AI can learn from those experiments. The EHR data that currently sits in closed vaults within medical centers would be of great use to those researchers but getting that data is not easy for them. There are a myriad of political, legal, privacy, security, technical, and other challenges that slow down or prevent easy access to this valuable data for researchers.

Beyond the data issue, there are a number of other barriers. These include ambiguous regulatory guidelines; yet to be determined return on investments for such solutions; lack of clinical trial data that establish the efficacy and safety of using these solutions in the real world environment; concerns regarding privacy and safety; high complexity in choosing and implementing the best AI solutions; designs that leads to poor fit with existing workflows; lack of enough talent that knows how to use AI solutions in healthcare and life sciences; fear amongst medical practitioners about the medical-legal risks, job security, and risk to their income, and more. That’s a lot of barriers. This explains why, in spite of its enormous promise, we have not seen widespread adoption of AI-powered solutions in healthcare. None of these barriers are insurmountable and I fully expect that in the next couple of decades AI will be the technology that will take us to the promised land in health. But given these barriers, it is good to be realistic about the pace of adoption and what needs to be done to make this a reality.

For the healthcare and life sciences executives who are hearing so much about AI everyday and want to get into the game, the good news is that less than perfect will be the name of the game for years to come. There is no perfect use case or technology to start with. How you get started is really a matter of your business needs, your human and capital resources, and your ambition level. AI is not plug and play technology. Each use case has to be stood up to provide benefits to your business. That means a complex implementation process, ongoing monitoring, and maintenance, and plenty of change management. As such, following the crowd to launch a cool radiology algorithm to read your routine X-rays may not end up having the hoped-for impact. However, exploring some of the emerging use cases in using generative AI in simplifying clinical documentation or improving coding can lead to huge benefits with the clinical staff satisfaction.

AI is a foundational technology, so its use cases span everything from administration to patient care to supply chain management to population health, and more. Where you have a lot of data (almost everywhere in healthcare) there will be opportunities for intelligent automation using AI. Some of the initial use cases that have seen some adoption include assistance in reading radiology scans, administrative workflows such as prior authorization and coding, operational use cases in hospitals, intensive care clinical use cases, patient engagement, and more. Each one of these use cases can have significant economic benefits for health systems. For example, the radiology use cases are very strong as the technology will serve as an immediate screener for critical findings on radiology scans and can alert the medical team to immediately check the images to confirm the presence of abnormalities and intervene in a timely manner. This can potentially improve patient outcomes, improve the workflow of the clinicians, and be used to attract more patients with conditions like stroke, trauma, and the critically ill.

I personally believe that the initial use cases that can have significant business benefits are the administrative and operational use cases. Although some may not get as excited about these use cases, I think they are an excellent way to get started. They are lower risk and if the data is not as ready or if technology is not fully mature, they can still provide benefits and not risk any clinical harm to the patients. Also, given the serious shortage of resources and the burned-out staff, helping them with documentation, coding, prior authorization workflows, generating patient communication material and referral letters, etc would be of great value. Revenue cycle management is a critical function for health systems and improving it using AI is within reach. Given the improvement in natural language processing using the large language models that underpin generative AI, these administrative and operational use cases can start to provide benefits in a reasonable amount of time. For executives who want to get started with AI and are trying to decide how and with what use case, they represent an excellent and safe place to start.

It is important to establish clear criteria for evaluating the potential use cases and the specific companies under consideration. In AI Doctor, I lay out guidelines for executives as they evaluate AI solutions. These guidelines assess the potential economic impact of the use case, availability of the needed data, workflow implications, potential for insurance reimbursement, total cost of ownership, governance requirements, complexity of implementation, the product’s track record with similar customers, level of evidence for the product to date, and more. Although this type of assessment may require cross-functional teams from the organization, as well as potential help from consultants and contractors, it will pay huge dividends when the right use cases are prioritized, and the organization gains clear benefits from its investments.

About the author: Dr. Ronald Razmi, is Co-Founder and Managing Director at Zoi Capital. Ron began his career as a cardiologist and was one of the pioneers in application of MRI systems in managing cardiac patients. He built the software system, cardiovascularmri.com, to serve as a platform for education and training of cardiologists in using MRI in managing their patients. He is a co-author of the textbook “Handbook of Cardiovascular Magnetic Resonance Imaging.”

As a McKinsey consultant, Ron worked with the world’s top life sciences companies in strategy, M&A, and product development. In 2011, he founded Acupera, a population health management software company. This unique platform enables healthcare organizations to use data, care pathways, and AI to industrialize key aspects of care delivery and patient management.

Ron is the author of the upcoming book, “AI Doctor: The Rise of Artificial Intelligence in Healthcare.” Ron has multiple publications in peer-reviewed scientific journals and has been a prolific speaker and writer in the field of digital health.