By using website you agree to our use of cookies as described in our cookie policy. Learn More

TDWI Upside - Where Data Means Business

Text-Based AI in Healthcare: The Challenges and Possibilities

Leveraging AI to continuously summarize the combined healthcare literature makes it possible to give every patient the best possible treatment informed by the most current research.

With medical knowledge doubling approximately every 73 days, it is simply impossible for any healthcare practitioner or researcher to stay on top of the latest advances. If there is one field that needs AI's help to deal with an information deluge, it is healthcare.

For Further Reading:

Why Emotion AI Is the Key to Mental Health Treatment

Q&A: Fragmented Patient Data Impedes Analytics in Healthcare

Healthcare Analytics in the Face of Heavy Volume

There has been significant progress in the use of artificial intelligence to analyze text, arguably the most famous of which are the GPT-2 and now GPT-3 tools from Open AI. These advances are exciting, but as the authors point out, the models are still "brittle" and come with a fair amount of error. In the world of healthcare, there is very little tolerance for error.

These models are also trained on a more general corpus and, therefore, are not appropriate for healthcare without first being refined on a medical corpus. As a result, practical applications of AI in the analysis of healthcare text are limited. As the use of AI for text analysis in healthcare expands, here are some key considerations:

Have a clear objective. Being able to predict the next word in a sentence is a cool parlor trick, but generalized tools will require a much more customized approach with a specific goal in mind. Arguably, the most important goal in healthcare is to create an AI tool to automatically summarize all the medical literature and decide the next-best treatment for a particular patient. To achieve this goal, we must be able to identify whether an article relates to a treatment advance, automatically extract the critical information, and automatically add those results to all the previous related research.

Recognize the need for medical experts. Medical experts must review the results of any text-based AI to make sure the output is of the highest quality. That means that any AI efforts will take time and be costly. In nearly every healthcare use of AI for text extraction, medical experts create a gold-standard database on which any model is trained. For example, with the goal of creating ongoing meta-analyses, medical experts are first called to review a representative set of articles and extract the salient information, such as the treatment used and the number of patients treated, from the research. With this database in hand, AI models can be trained to automatically summarize research and extract the key information.

Take into account scientific progress. Not every piece of research is equally valuable. With the goal of automatically summarizing cumulative research results, there must be a way to weigh each piece of research based on the incremental value it provides. Luckily, in healthcare there are a couple of elements that signal quality. The first is the type of study; randomized controlled trials produce the most reliable results when compared to single-arm studies or retrospective analyses. The second is the size of the study. A study of 10 patients is much less reliable than a study of 1,000 patients. By extracting these critical elements from each article, results are weighed appropriately when developing cumulative results.

Understand each disease is unique. There are approximately 10,000 identified diseases in humans, many with unique genetic markers and specialized tests and treatments. Although the full set of biomedical literature is vast, the amount of specific literature available for many of these diseases is too sparse to train a text-based AI model. To tackle this, it's critical to identify how each disease relates to others so we can leverage the full biomedical literature while providing accurate information for each disease -- no small task. This is an active area of exploration that takes into account genotypic and phenotypic similarities to identify how closely related diseases are.

A Final Word

Today, doctors deploy a variety of tactics to stay updated about their field, including focusing on a narrow disease area, only reading publications from the top journals, and attending conferences and continuing medical education events. Unfortunately, although these practices are effective, they are incomplete. As a result, and despite their best efforts, it's nearly impossible for healthcare professionals to stay current on all the advances in their field. By leveraging AI to continuously summarize the combined healthcare literature, it will be possible for every patient to have the best possible treatment informed by the most current research.

About the Author

Patrick Howie is the CEO and founder of MediFind, which he started after watching his brother struggle to navigate the healthcare system to treat his rare cancer. Howie is the former head of Global Analytics at Merck and has held senior leadership positions in multiple healthcare startups. He is also the author of "The Evolution of Revolutions." Connect with him on LinkedIn.

TDWI Membership

Accelerate Your Projects,
and Your Career

TDWI Members have access to exclusive research reports, publications, communities and training.

Individual, Student, and Team memberships available.